Datasets:

kpriyanshu256

/

the_verge-linustechtips-two_min

like 0

Dear Fellow Scholars, this is two-minute papers with Karo Zona Ifeher. Through over 200 episodes of this series, we talked about many learning-based algorithms that are able to solve problems that previously seemed completely impossible. They can look at an image and describe what they depict in a sentence, or even turn video game graphics into reality and back. Amazing new results keep appearing every single week. However, an important thing that we need to solve is that if we deploy these neural networks in a production environment, we would want to know if we are relying on a good or bad AI's decision. The narrative is very simple. If we don't trust a classifier, we won't use it. And perhaps the best way of earning the trust of a human would be if the AI could explain how it came to a given decision. Strictly speaking, a neural network can explain it to us, but it will show us hundreds of thousands of neural activations that are completely unusable for any sort of intuitive reasoning. So, what is even more difficult to solve is that this explanation happens in a way that we can interpret. An earlier approach used decision trees that described what the learner looks at and how it uses this information to arrive to a conclusion. This new work is quite different. For instance, imagine that a neural network would look at all the information we know about the patient and tell us that this patient likely has the flu. And in the meantime, it could tell us that the fact that the patient has a headache and sneezes a lot contributed to the conclusion that he has the flu, but the lack of fatigue is notable evidence against it. Our doctor could take this information and instead of blindly relying on the output, could make a more informed decision. A fine example of a case where AI does not replace but augment human labor. An elegant tool for a more civilized age. Here, we see an example image where the classifier explains which region contributes to the decision that this image depicts a cat and which region seems to be counter evidence. We can use this not only for tabulated patient data and images, but text as well. In this other example, we try to find out whether a piece of written text is about Christianity or atheism. Note that the decision itself is not as simple as looking for a few keywords. Even a mid-tier classifier is much more sophisticated than that. But it can tell us about the main contributing factors. A big additional selling point is that this technique is model agnostic, which means that it can be applied to other learning algorithms that are able to perform classification. It is also a possibility that an AI is only right by chance and if this is the case, we should definitely know about that. And here, in this example, with the additional explanation, it is rather easy to find that we have a bad model that looks at the background of the image and thinks that it is the fur of a wolf. The tests indicate that humans make significantly better decisions when they lean on explanations that are extracted by this technique. The source code of this project is also available. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Kato Yjona Yfahir. Due to popular demand, here is the new Deep Mind Paper on WaveNet. WaveNet is a text to speech algorithm that takes a sentence as an input and gives us audio footage of these words being uttered by a person of our choice. Let's listen to some results from the original algorithm. Note that these are all synthesized by the AI. The Blue Lagoon is a 1980 American romance and adventure film directed by Randall Cliser. The Blue Lagoon is a 1980 American romance and adventure film directed by Randall Cliser. Aspects of the Sublime in English Poetry and Painting, 1770-1850. Aspects of the Sublime in English Poetry and Painting, 1770-1850. All this requires is some training data from this person's voice, typically 10-30 hours, and a ton of computational power. The computational power part is especially of interest because we have to produce over 16-24,000 samples for each second of continuous audio footage. And unfortunately, as you can see here, these new samples are generated one by one. And since today's graphics cards are highly parallel, this means that it is a waste to get them to have one compute unit that does all the work while the others are sitting there twiddling their thumbs. We need to make this more parallel somehow. So, the solution is simple. Instead of one, we can just simply make more samples in parallel. No, no, no, no, no. It doesn't work like that. And the reason for this is that speech is not like Randall's noise. It is highly coherent where the new samples are highly dependent on the previous ones. We can only create one new sample at a time. So, how can we create the new waveform in one go using these many compute units in parallel? This new wave-nut variant starts out from white noise and applies changes to it over time to morph it into the output speech waveform. The changes take place in parallel over the entirety of the signal, so that's a good sign. It works by creating a reference network that is slow, but correct. Let's call this the Teacher Network. And the new algorithm arises as a student network which tries to mimic what the teacher does, but the student tries to be more efficient at that. This has a similar vibe to generative adversarial networks where we have two networks. One is actively trying to fool the other one, while this other one tries to better distinguish fake inputs from real ones. However, it is fundamentally different because of the fact that the student does not try to fool the teacher, but mimic it while being more efficient. And this yields a blistering fast version of wave-nut that is over a thousand times faster than its predecessor. It is not real time, it is 20 times faster than real time. And you know what the best part is? Usually, there are heavy trade-offs for this. But this time, the validation section of the paper reveals that there is no perceived difference in the outputs from the original algorithm. Hell yeah! So, where can we try it? Well, it is already deployed online in Google Assistant, in multiple English and Japanese voices. So, as you see, I was wrong. I said that a few papers down the line, it will definitely be done in real time. Apparently, with this new work, it is not a few papers down the line, it is one, and it is not a bit faster, but a thousand times faster. Things are getting out of hand real quick, and I mean this in the best possible way. What a time to be alive! This is one incredible and highly inspiring work. Make sure to have a look at the paper, perfect training for the mind. As always, it is available in the video description. Thanks for watching, and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Zsolnai-Fehir. This paper is about simulating on a computer what happens when bubbles collide. Prepare for lots of beautiful footage. This is typically done by simulating the Navier Stokes equations that describe the evolution of the velocity within a piece of fluid over time. However, because the world around us is a continuum, we cannot compute these quantities in an infinite number of points. So, we have to subdivide the 3D space into a grid and compute them only in these grid points. The finer the grid, the more details appear in our simulations. If we try to simulate what happens when these bubbles collide, we would need to create a grid that can capture these details. This is an issue because the thickness of a bubble film is in the order of 10 to 800 nanometers and this would require a hopelessly fine high resolution grid. By the way, measuring the thickness of bubbles is a science of its own, there is a fantastic reddit discussion on it, I put a link to it in the video description, make sure to check it out. So, these overly fine grids take too long to compute, so what do we do? Well, first we need to focus on how to directly compute how the shape of soap bubbles evolves over time. Fortunately, from Belgian physicist Joseph Plato, we know that they seek to reduce their surface area, but retain their volume over time. One of the many beautiful phenomena in nature. So, this shall be the first step. We simulate forces that create the appropriate shape changes and proceed into an intermediate state. However, by pushing the film inwards, its volume has decreased. Therefore, this intermediate state is not how it should look in nature. This is to be remedied now where we apply a volume correction step. In the validation section, it is shown that the results follow Plato's laws quite closely. Also, you know well that my favorite kind of validation is when we let reality be our judge, and in this work, the results have been compared to a real life experimental setup and proved to be very close to it. Take a little time to absorb this. We can write a computer program that reproduces what would happen in reality and result in lots of beautiful video footage. Loving it. And the best part is that the first surface evolution step is done through an effective implementation of the hyperbolic mean curvature flow, which means that the entirety of the process is typically 3 to 20 times faster than the state of the art while being more robust in handling splitting and merging scenarios. The computation times are now in the order of milliseconds instead of seconds. The earlier work in this comparison was also showcased in two minute papers if I see it correctly, it was in episode number 18. Holy matter of papers, how far we have come since. I've put a link to it in the video description. The paper is beautifully written and there are plenty of goodies there in, for instance, an issue with non-manifold junctions is addressed, so make sure to have a look. The source code of this project is also available. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Jolnai-Fehir. This work is about building an AI vocoder that is able to synthesize believable singing from MIDI and lyrics as inputs. But first, what is a vocoder? It works kind of like this. A vocoder is an audio processor that is almost unmistakable in its character. It is used to combine the tonal qualities of one sound source, call a carrier, with the frequency-specific movements of a second signal called a modulator. This allows you to create a computer voice. Fellow Scholars, who are fans of Jean-Michel-Jarge music, are likely very familiar with this effect. I've put a link to an example song in the video description. Make sure to leave a comment with your favorite songs with vocoders, so I and other fellow Scholars can also nerd out on them. And now, about the MIDI and lyrics terms. The lyrics part is a simple text file containing the words that this synthesized voice should sing, and the MIDI is data that describes the pitch, length, and velocity of each sound. With a little simplification, we could say that the score is given as an input and the algorithm has to output the singing footage. We will talk about the algorithm in a moment, but for now, let's listen to it. Wow! So this is a vocoder. This means that it separates the pitch and thumb-bar components of the voice, therefore, the waveforms are not generated directly, which is a key difference from Google DeepMind's wave net. This leads to two big advantages. One, the generation times are quite favorable, and by favorable, I guess you're hoping for real time. Well, hold on to your papers because it is not real time, it is 10 to 15 times real time. And two, this way, the algorithm will only need a modest amount of training data to function well. Here you can see the input phonemes that make up the syllables of the lyrics, each typically corresponding to one note. This is then connected to a modified wave net architecture that uses two by one dilated convolutions. This means that the dilation factor is doubled in each layer, thereby introducing an exponential growth in the receptive field of the model. This helps us keep the parameter count down, which enables training on small datasets. As validation, the mean opinion scores have been recorded in a previous episode, we discussed that this is a number that describes how a sound sample would pass as genuine human speech or singing. The test showed that this new method is well ahead of the competition, approximately, midway between the previous works and the reference singing footage. There are plenty of other tests in the paper. This is just one of many, so make sure to have a look. This is one important stepping stone towards synthesizing singing that is highly usable in digital media and where generation is faster than real time. Creating a MIDI input is a piece of cake with a MIDI master keyboard or we can even draw the notes by hand in many digital audio workstation programs. After that, writing the lyrics is as simple as it gets and doesn't need any additional software. Tools like this are going to make this process accessible to everyone, loving it. If you would like to help us create more elaborate videos, please consider supporting us on Patreon. We also support one-time payments through cryptos like Bitcoin, Ethereum and Litecoin. Everything is available in the video description. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two minute papers with Kato Zsolnai-Fehir. When we are talking about deep learning, we are talking about neural networks that have tens, sometimes hundreds of layers and hundreds of neurons within these layers. This is an enormous number of parameters to train, and clearly there should be some redundancy some duplication in the information within. This paper is trying to throw out many of these neurons of the network without affecting its accuracy too much. This process we shall call pruning, and it helps creating neural networks that are faster and smaller. The accuracy term I used typically means a score on a classification task, in other words, how good this learning algorithm is in telling what an image or video depicts. This particular technique is specialized for pruning convolutional neural networks, where the neurons are in doubt with a small receptive field and are better suited for images. These neurons are also commonly referred to as filters, so here we have to provide a good mathematical definition of a proper pruning. The authors proposed a definition where we can specify a maximum accuracy drop that we deemed to be acceptable, which will be denoted with the letter B in a moment, and the goal is to prune as many filters as we can without going over the specified accuracy loss budget. The pruning process is controlled by an accuracy and efficiency term, and the goal is to have some sort of balance between the two. To get a more visual understanding of what is happening, here the filters you see outlined with the red border are kept by the algorithm, and the rest are discarded. As you can see, the algorithm is not as trivial as many previous approaches that just prune away filters with weaker responses. Here you see the table with the B numbers. Several tests reveal that around a quarter of the filters can be pruned with an accuracy loss of 0.3% and with a higher B we can prune more than 75% of the filters with a loss of around 3%. This is incredible. Image segmentation tasks are about finding the regions that different objects inhabit. Interestingly, when trying the pruning for this task, it not only introduces a minimal loss of accuracy, in some cases the pruned version of the neural network performs even better. How cool is that? And of course, the best part is that we can choose a trade-off that is appropriate for our application. For instance, if we are looking for a light cleanup, we can use the first option at a minimal penalty, or if we wish to have a tiny, tiny neural network that can run on a mobile device, we can look for the more heavy-handed approach by sacrificing just a tiny bit more accuracy. And we have everything in between. There is plenty more validation for the method in the paper, make sure to have a look. It is really great to see that new research works make neural networks not only more powerful over time, but there are efforts in making them smaller and more efficient at the same time. Great news indeed. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is 2 Minute Papers with Károly Zsolnai-Fehér. Earlier, we talked about Google's WaveNet, a learning-based text-to-speech engine. This means that we give it a piece of written text, and after a training step using someone's voice, it has to read it aloud using this person's voice as convincingly as possible. And this follow-up work is about making it even more convincing. Before we go into it, let's marvel at these new results together. Generative adversarial network or variational autoencoder. He has read the whole thing. He reads books. This is really awesome. This is your personal assistant, Google Home. This is your personal assistant, Google Home. The buses aren't the problem, they actually provide a solution. The buses aren't the problem, they actually provide a solution. She sells seashells on the seashore. The shells she sells are seashells, I'm sure. As you can hear, it is great at prosody, stress, and intonation, which leads to really believable human speech. The magic component in the original WaveNet paper was introducing dilated convolutions for this problem. This makes large skips in the input data, so we have a better global view of it. It is a bit like increasing the receptive field of the eye so we can see the entire landscape and not only a tree on a photograph. The magic component in this new work is using mouse spectrograms as an input to WaveNet. This is an intermediate representation that is based on the human perception that records not only how different words should be pronounced, but the expected volumes and intonations as well. The new model was trained on about 24 hours of speech data. And of course, no research work should come without some sort of validation. The first is recording the mean opinion scores for previous algorithms, this one, and real professional voice recordings. The mean opinion score is a number that describes how a sound sample would pass as genuine human speech. The new algorithm passed with flying colors. And even more practical evaluation was also done in the form of a user study where people were listening to the synthesized samples and professional voice narrators and had to guess which one is which. And this is truly incredible because most of the time people had no idea which was which. If you don't believe it, we'll try this ourselves in a moment. A very small but statistically significant tendency towards favoring the real footage was recorded likely because some words like Merlot are mispronounced. Only voiced audiobooks, automatic voice narration for video games. Bring it on! What a time to be alive! Note that producing these waveforms is not real time and still takes quite a while. To progress along that direction, scientists at DeepMind wrote a hack of a paper where they spat a wave nut up a thousand times. Leave a comment if you would like to hear more about it in a future episode. And of course, new inventions like this will also raise new challenges down the line. It may be that voice recordings will become much easier to forge and be less useful as evidence unless we find new measures to verify their authenticity, for instance to sign them like we do with software. In closing, a few audio sample pairs. One of them is real, one of them is synthesized. What do you think? Which is which? Leave a comment below. That girl did a video about Star Wars lipstick. That girl did a video about Star Wars lipstick. She earned a doctorate in sociology at Columbia University. She earned a doctorate in sociology at Columbia University. George Washington was the first president of the United States. George Washington was the first president of the United States. I'm too busy for romance. I'm too busy for romance. I'll just leave a quick hint here that I found on the webpage. Up, there you go. If you have enjoyed this episode, please make sure to support us on Patreon. This is how we can keep the show running, and you know the drill. One dollar is almost nothing, but it keeps the papers coming. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Kato Ejolna Ifehir. This project is about the data set created through a joint effort between MIT and Facebook. As it turns out, this data set is way more useful than I initially thought I'll tell you in a moment why. Data sets are used to train and test the quality of learning algorithms. This particular data set contains short video clips. These clips are passed to a neural network which is asked to classify the kind of activity that is taking place in the video. In this data set, there are many cases where everything is given to come to a logical answer that is wrong. We may be in a room with the climbing wall, but exercising is not necessarily happening. We could be around the swimming pool, but swimming is not necessarily happening. I'm pretty sure this has happened to you too. This is a brilliant idea because it is super easy for a neural network to assume that if there is a swimming pool, swimming is probably happening, but it takes a great deal of understanding to actually know what constitutes the swimming part. A few episodes ago, we discussed that this could potentially be a stepping stone towards creating machines that think like humans. Without looking into it, it would be easy to think that creating a data set is basically throwing a bunch of training samples together and calling it a day. I can assure you that this is not the case and that creating a data set like this was a Herculean effort as it contains more than half a million videos and almost 2 million annotations for 200 different activities. And there are plenty of pre-processing steps that one has to perform to make it usable. The collection procedure contains a video crawling step where a large number of videos are obtained from YouTube which are to be de-duplicated, which means removing videos that are too similar to one already contained in the database. A classical case is many different kinds of commentary on the same footage. This amounted to the removal of more than 150,000 videos. Then all of these videos undergo a shot and person detection step where relevant subclips are extracted that contain some kind of human activity. These are then looked at by two different classifiers and depending on whether there was a consensus between the two, a decision is made whether the clip is to be discarded or not. This step helps balancing the ratio of videos where there is some sort of relevant action compared to the clips where there is no relevant action happening. This also makes the negative samples much harder because the context may be correct but the expected activity may not be there. This is the classical hard case with the swimming pool and people in swimming suits twiddling their thumbs instead of swimming. And here comes the more interesting part. When trying to train a neural network for other, loosely related tasks using this dataset for pre-training improves the scores significantly. I'll try to give a little context for the numbers because these numbers are absolutely incredible. There are cases where the success rate is improved by over 30% which speaks for itself. However, there are other cases where the difference is about 10 to 15% that is also remarkable when we are talking about high numbers because the closer the classifier gets to 100% the more difficult the remaining corner cases are that improve the accuracy. In these cases even a 3% improvement is remarkable. And before we go, greetings and best regards to Lucas, the little scholar who seems to be absorbing the papers along with the mother's milk. Excellent, you can start early enough in the pursuit of knowledge. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Zsolnai-Fehir. This footage that you see here came freshly from Google DeepMind's lab and is about benchmarking reinforcement learning algorithms. Here, you see the classical cardboard swing-up task from this package. As the algorithm starts to play, a score is recorded that indicates how well it is doing and the learner has to choose the appropriate actions depending on the state of the environment to maximize this score. Reinforcement learning is an established research subfield within machine learning with hundreds of papers appearing every year. However, we see that most of them cherry pick a few problems and test against previous works on this very particular selection of tasks. This paper describes a package that is not about the algorithm itself but about helping future research projects to be able to test their results against previous works on an equal footing. This is a great idea which has been addressed earlier by OpenAI with their learning environment by the name Jim. So the first question is, why do we need a new one? The DeepMind Control Suite provides a few differentiating features. One, Jim contains both discrete and continuous tasks where this one is concentrated on continuous problems only. This means that state, time, and action are all continuous which is usually the hallmark of more challenging and lifelike problems. For an algorithm to do well, it has to be able to learn the concept of velocity, acceleration, and other meaningful physical concepts and understand their evolution over time. Two, there are domains where the new control suite is a superset of Jim meaning that it offers equivalent tasks and then some more. And three, the action and reward structures are standardized. This means that the results and learning curves are much more informative and easier to read. This is crucial because research scientists read hundreds of papers every year and this means that they don't necessarily have to look at videos. They immediately have an intuition of how an algorithm works and how it relates to previous techniques just by looking at the learning curve plots. Many tasks also include a much more challenging variant with more sparse rewards. We discussed these sparse rewards in a bit more detail in the previous episode. If you are interested, make sure to click the card on the lower right at the end of this video. The paper also contains an exciting roadmap for future development including quadruped locomotion, multithreaded dynamics, and more. Of course, the whole suite is available free of charge for everyone. The link is available in the description. Super excited to see a deluge of upcoming AI papers and see how they beat the living hell out of each other in 2018. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Kato Ejona-Ifahir. This work is about improving reinforcement learning. Reinforcement learning is a learning algorithm that we can use to choose a set of actions in an environment to maximize a score. Our classical example applications are helicopter control where the score to be maximized would be proportional to the distance that we traveled safely or any computer game of your choice where a score can describe how well we are doing. For instance, in Frostbite, our score describes how many jumps we have survived without dying and this score is subject to maximization. Earlier, scientists at DeepMind combined reinforcement learner with a deep neural network so the algorithm could look at the screen and play the game much like a human player would. This problem is especially difficult when the rewards are sparse. This is similar to what a confused student would experience after a written exam where only one grade is given but the results for the individual problems are not shown. It is quite hard to know where we did well and where we missed the mark and it is much more challenging to choose the appropriate topics to study to do better next time. When starting out, the learner starts exploring the parameter space and performs crazy, seemingly non-sensical actions until it finds a few scenarios where it is able to do well. This can be thought of as adding noise to the actions of the agent. Scientists at OpenAI propose an approach where they add noise not directly to the actions but to the parameters of the agent which results in perturbations that depend on the information that the agent senses. This leads to less flailing and the more systematic exploration that substantially decreases the time taken to learn tasks with sparse rewards. For instance, it makes a profound difference if we use it in the Walker game. As you can see here, the algorithm with the parameter space noise is able to learn the concept of galloping while the traditional method does, well, I am not sure what it is doing to be honest but it is significantly less efficient. The solution does not come without challenges. For instance, different layers respond differently to this added noise and the effect of the noise on the outputs grows over time which requires changing the amount of noise to be added depending on its expected effect on the output. This technique is called Adaptive Noise Scaling. There are plenty of comparisons and other cool details in the paper make sure to have a look it is available in the video description. DeepMind's deep reinforcement learning was published in 2015 with some breathtaking results and superhuman plays on a number of different games and it has already been improved leaves and bounds beyond its initial version. And we are talking about OpenAI so of course the source code of this project is available under the permissive MIT license. In the meantime we have recently been able to upgrade the entire TFR sound recording pipeline through your support on Patreon. I have been yearning for this for a long, long time now and not only that but we could also extend our software pipeline with sound processing units that use AI and work like magic. Quite fitting for the series right? Next up is a recording room or recording corner with acoustic treatment depending on our budget. And thank you for your support it makes a huge difference. A more detailed write up on this is available in the video description. Have a look. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Karo Zsolnai-Fehir. This work is about creating an AI that can perform audio-visual correspondence. This means two really cool tasks. One, when given a piece of video and audio, it can guess whether they match each other. And two, it can localize the source of the sounds heard in the video. And wait, because this gets even better, as opposed to previous works, here the entire network is trained from scratch and is able to perform cross-model retrieval. Cross-model retrieval means that we are able to give it an input sound and it will be able to find pictures that would produce similar sounds. Or vice versa. For instance, here the input is the sound of a guitar, note the loudspeaker icon in the corner, and it shows us a bunch of either images or sounds that are similar. Marvelous. The training is unsupervised, which means that the algorithm is given a bunch of data and learns without additional labels or instructions. The architecture and results are compared to a previous work by the name Look, Listen, and Learn that we covered earlier in the series, the link is available in the video description. As you can see, both of them run a convolution on your own network. This is one of my favorite parts about deep learning. The very same algorithm is able to process and understand signals of very different kinds, video and audio. The old work concatenates this information and produces a binary yes-no decision whether it thinks the two streams match. This new work tries to produce a number that encodes the distance between the video and the audio. Kind of like the distance between two countries on a map, but both video and audio signals are embedded in the same map. And the output decision always depends on how small or big this distance is. This distance metric is quite useful. If we have an input video or audio signal, choosing other video and audio snippets that have a low distance is one of the important steps that opens up the door to this magical cross-model retrieval. What a time to be alive. Some results are easy to verify, others may spark some debate. For instance, it is quite interesting to see that the algorithm highlights the entirety of the guitar string as a sound source. If you are curious about this mysterious blue image here, make sure to have a look at the paper for an explanation. Now this is a story that we would like to tell to as many people as possible. Everyone needs to hear about this. If you would like to help us with our quest, please consider supporting us on Patreon. You can pick up some cool perks like getting early access to these videos or deciding the order of upcoming episodes. Details are available in the video description. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karojona Ifei. This paper discusses possible roadmaps towards building machines that are endowed with human like thinking. And before we go into that, the first question would be, is there value in building machines that think like people? Do they really need to think like people? Isn't it a bit agotistical to say, if they are to become any good at this and this task, they have to think like us? And the answer is, well, in some cases, yes. If you remember DeepMind's DeepQ learning algorithm, it was able to play on a superhuman level on 29 out of 49 different Atari games. For instance, it did quite well in breakout, but less so in frostbite. And by frostbite, I mean not the game engine, but the Atari game from 1983 where we need to hop from ice flow to ice flow and construct an igloo. However, we are not meant to jump around arbitrarily, we can gather these pieces by jumping on the active ice flows only, and these are shown with white color. Have a look at this plot. It shows the score it was able to produce as a function of game experience in hours. As you can see, the original DQN is doing quite poorly, while the extended versions of the technique can reach a relatively high score over time. This looks really good. Until we look at the x-axis, because then we see that this takes around 462 hours and the scores plateau afterwards. Well, compared that to humans that can do at least as well, or a bit better, after a mere 2 hours of training. So clearly, there are cases where there is an argument to be made for the usefulness of human like AI. The paper describes several possible directions that may help us achieve this. Two of them is understanding intuitive physics and intuitive psychology. Even young infants understand that objects follow smooth paths and expect liquids to go around barriers. We can try to endow an AI with similar knowledge by feeding it with physics simulations and their evolution over time to get an understanding of similar phenomena. This could be used to augment already existing neural networks and give them a better understanding of the world around us. Active psychology is also present in young infants. They can tell people from objects or distinguish other social and anti-social agents. They can also learn goal-based reasoning quite early. This means that a human who looks at an experienced player play frostbite can easily derive the rules of the game in a matter of minutes. Kind of what we are doing now. Neural networks also have a limited understanding of compositionality and causality, and often perform poorly when describing the content of images that contain previously known objects interacting in novel and scene ways. There are several ways of achieving each of these elements described in the paper. If we manage to build an AI that is endowed with these properties, it may be able to think like humans and, through self-improvement, may achieve the kind of intelligence that we see in all these science fiction movies. There is lots more in the paper, learning to learn approximate models for thinking faster, model-free reinforcement learning, and a nice Q&A section with responses to common questions and criticisms. It is a great read, and it is easy to understand for everyone. I encourage you to have a look at the video description for the link to it. Scientists at Google DeepMind have also written a commentary article where they largely agree with the premises described in this paper, and add some thoughts about the importance of autonomy in building human-like intelligence. Both papers are available in the video description, and both are great reads, so make sure to have a look at them. It is really cool that we have plenty of discussions on potential ways to create a more general intelligence that is at least as potent as humans in a variety of different tasks. What a time to be alive! Thanks for watching and for your generous support, and I'll see you next time!

Dear Fellow Scholars, this is 2 Minute Papers with Károly Zsolnai-Fehér. The goal of this work is to have a robot that automatically creates a 3D model of an indoor space, including path planning and controlling attention. Now this immediately sounds like quite a challenging task. This robot uses an RGB-D camera, so beyond the colors it also gets some depth information that describes how far the viewed objects are from the observer. From this information, it tries to create a 3D digital model of these interiors. It not only does that, but it also constantly replants its trajectory based on newly found areas. These paths have to be smooth and walkable without bumping into objects and of course adapt to the topology of the building. You'll see in a moment why even this mundane sounding smooth part is really challenging to accomplish. Spoiler alert, it's about singularities. With this proposed technique, the robot takes a lap in the building and builds a rough representation of it, kind of like a mini-map in your favorite computer games. Previous techniques worked with potential and gradient fields to guide the navigation. The issue with these is that the green dots that you see here represent singularities. These are the generate points that introduce ambiguity to the path planning process and reduce the efficiency of the navigation. The red dots are sinks, which are even worse because they can trap the robot. This new proposed tensor field representation contains a lot fewer singularities and its favorable mathematical properties make it sink free. This leads to much better path planning, which is crucial for maintaining high reconstruction quality. If you have a look at the paper, you'll see several occurrences of the word advection. This is particularly cool because these robot paths are planned in these gradient or tensor fields represent the vertices and flow directions similarly to how fluid flows are computed in simulations, many of which you have seen in this series. Beautiful. Love it. However, as advection can't guarantee that we'll have a full coverage of the building, this proposed technique borrows classic structures from graph theory. Graph theory is used to model the connections between railway stations or to represent people and their relationships in a social network. And here, a method to construct a minimum spanning tree was borrowed to help deciding which direction to take at intersections for optimal coverage with minimal effort. Robots, fluids, graph theory, some of my favorite topics of all time, so you probably know how happy I was to see all these theories come together to create something really practical. This is an amazing paper. Make sure to have a look at it. It is available in the video description. The source code of this project is also available. If you have enjoyed this episode, make sure to subscribe to the series and click the bell icon to never miss an episode. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Ejona Efehir. Deep learning means that we are working with neural networks that contain many inner layers. As neurons in each layer combine information from the layer before, the deeper we go in these networks, the more elaborate details we're going to see. Let's have a look at an example. For instance, if we train a neural network to recognize images of human faces, first we'll see an edge detector and as a combination of edges, object parts will emerge in the next layer. And in the later layers, a combination of object parts create object models. Neural texture synthesis is about creating lots of new images based on an input texture and these new images have to resemble but not copy the input. Previous works on neural texture synthesis focused on how different features in a given layer relate to the ones before and after it. The issue is that because neurons in convolutional neural networks are in doubt with a small receptive field, they can only look at an input texture at one scale. So for instance, if you look here, you see that with previous techniques, trying to create small scale details in a synthesized texture is going to lead to rather poor results. This new method is about changing the inputs and the outputs of the network to be able to process these images at different scales. These scales range from coarser to finer versions of the same images. Sounds simple enough, right? This simple idea makes all the difference. Here, you can see the input texture and here's the output. As you can see, it has different patterns but has very similar properties to the input. And if we zoom into both of these images, we can see that this one is able to create beautiful, high frequency details as well. Wow, this is some really, really crisp output. Now, it has to be emphasized that this means that the statistical properties of the original image are being mimicked really well. What it doesn't mean is that it takes into consideration the meaning of these images. Just have a look at the synthesized bubbles or the flowers here. The statistical properties of the synthesized textures may be correct, but the semantic meaning of the input is not captured well. In a future work, it would be super useful to extend this algorithm to have a greater understanding of the structure and the symmetries of the input images into consideration. The source code is available under the permissive MIT license, so don't hold back those crazy experiments. If you have enjoyed this episode and you think the series provides you value or entertainment, please consider supporting us on Patreon. One-time payments and cryptocurrencies like Bitcoin or Ethereum are also supported and have been massively successful. I am really out of words. Thank you so much. The details are available in the video description. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Kato Ejona Ifehir. It has been a while now since we've talked about fluid simulations and now it is time for us to have a look at an amazing technique that creates simulations with viscoelastic fluids plus rigid and deformable bodies. This is a possibility with previous techniques that takes forever to compute and typically involves computational errors that add up over time and lead to a perceptible loss of viscoelasticity. I'll try to explain these terms and what's going on in the moment. There will be lots of eye candy to feast your eyes on throughout the video but I can assure you that this scene takes the cake. The simulation phase here took place at one frame per second which is still not yet fast enough for the much coveted real-time applications, however, is super competitive compared to previous offline techniques that could do this. So what does viscosity mean? Viscosity is the resistance of a fluid against deformation. Water has a low viscosity and rapidly takes the form of the cup we pour it into where honey, ketchup and peanut butter have a higher viscosity and are much more resistant to these external forces. Elasticity is a little less elusive concept that describes to what degree a piece of fluid behaves like elastic solids. However, viscous and elastic are not binary yes or no concepts. There is a continuum between the two and if we have the proper machinery we can create viscoelastic fluid simulations. The tau parameter that you see here controls how viscous or elastic the fluid should be and is referred to in the paper as relaxation time. As tau is increased towards infinity, the friction dominates the internal elastic forces and the polymer won't be able to recover its structure well. The opposite cases where we reduce the tau parameter to 0, then the internal elastic forces will dominate and our polymer will be more rigid. The alpha parameter stands for compliance, which describes the fluidity of the model. Using a lower alpha leads to more solid behavior and higher alpha leads to more fluid behavior. The cool thing is that as a combination of these two parameters we can produce a lot of really cool materials ranging from viscous to elastic to inviscid fluid simulations. Have a look at this honey pouring scene. Hmmm. This simulation uses more than 100,000 particles and 6 of these frames can be simulated in just one second. Wow. If we reduce the number of particles to a few tens of thousands real time human interaction with these simulations also becomes a possibility. A limitation of this technique is that most of our decisions involving the physical properties of the fluid are collapsed into the tau and alpha parameters if we are looking for more esoteric fluid models we should look elsewhere. I am hoping that since part of the algorithm runs on the graphics card the speed of this technique can be further improved in the near future. That would be awesome. Admittedly we've only been scratching the surface so make sure to have a look at the paper for more details. Thanks for watching and for your generous support. See you next time.

Dear Fellow Scholars, this is two-minute papers with Kato Ejeona Ifehir. This work is about performing useful image restoration tasks with a convolution on your own network with an additional twist. Its main use cases are as follows. One, in the case of JPEG artifact removal, the input is this image with many blocky artifacts that materialize during compression and the output is a restored version of this image. Two, image inpainting, where some regions of the input image are missing and are to be filled with useful and hopefully plausible information. Three, super resolution, where the input image is intact but is very coarse and has low resolution, and the output should be a more detailed, higher resolution version of the same image. This is the classic enhanced scenario from the CSI-TV series. It is typically hard to do because there is a stupendously large number of possible high-resolution image solutions that we could come up with as an output. Four, image denoising is also a possibility. The standard way of doing these is that we train such a network on a large database of images so that they can learn the concept of many object classes such as humans, animals and more and also the typical features and motives that are used to construct such images. These networks have some sort of understanding of these images and hence can perform these operations better than most handcrafted algorithms. So let's have a look at some comparisons. Do you see these bold lettered labels that classify these algorithms as trained or untrained? The bi-cubic interpolation is a classic untrained algorithm that almost naively tries to guess the pixel colors by averaging its neighbors. This is clearly untrained because it does not take a database of images to learn on. Understandably, the fact that these results are leg-luster is to show that non-learning based algorithms are not great at this. The SR-ResNet is a state-of-the-art learning-based technique for super-resolution that was trained on a large database of input images. It is clearly doing way better than bi-cubic interpolation. And look, we have this deep prior algorithm that performs comparably well but is labeled to be untrained. So what is going on here? And here comes the twist. This convolutional neural network is actually untrained. This means that the neural weights are randomly initialized which generally leads to completely useless results on most problems. So no aspect of this network works through the data it has learned on, all the required information is contained within the structure of the network itself. We all know that the structure of these neural networks matter a great deal, but in this case it is shown that it is at least as important as the training data itself. A very interesting and esoteric idea indeed, please make sure to have a look at the paper for details as there are many details to be understood to get a more complete view of this conclusion. In the comparisons, beyond the images, researchers often publish this PSNR number that you see for each image. This is the peak signal to noise ratio, which means how close the output image is to the ground truth and this number is of course always subject to maximization. Remarkably this untrained network performs well on both images with natural patterns and man-made objects. Reconstruction from a pair of flash and no flash photography images is also a possibility and the algorithm does not contain the light leaks produced by a highly competitive handcrafted algorithm, a joint bilateral filter. Quite remarkable indeed. The supplementary materials and the project website contain a ton of comparisons against competing techniques, so make sure to have a look at that if you would like to know more. The source code of this project is available under the permissive Apache 2.0 license. If you have enjoyed this episode and you feel that 8 of these videos a month is worth a dollar, please consider supporting us on Patreon. One dollar is almost nothing, but it keeps the papers coming. Recently we have also added the possibility of one-time payments through PayPal and Cryptocurrencies. I was stunned to see how generous our crypto-loving fellow scholars are. Since most of these crypto donations are anonymous and it is not possible to say thank you to everyone individually, I would like to say a huge thanks to everyone who supports the series and this applies to everyone regardless of contribution, just watching the series and spreading the word is already a great deal of help for us. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Károly Zsolnai-Fehér. Since the latest explosion in AI research, virtually no field of science remains untouched by neural networks. These are amazing tools that help us solve problems where the solutions are easy to identify but difficult to explain. For instance, we all know a backflip when we see one, but mathematically defining all the required forces, rotations and torque is much more challenging. Neural networks excel at these kinds of tasks provided that we can supply them a large number of training samples. If we pick inside these neural networks, we see more and more layers and more and more neurons within these layers as years go by. The final decision depends on what neurons are activated by our inputs. They are highly efficient, however, trying to understand how a decision is being made by these networks is going to be a fruitless endeavor. This is especially troublesome when the network gives us a wrong answer that we, without having access to any sort of explanation, may erroneously accept without proper consideration. This piece of work is about distillation, which means that we take a neural network and try to express its inner workings in the form of a decision tree. Decision trees take into consideration a series of variables and provide a clear roadmap towards the decision based on them. For instance, they are useful in using the age and the amount of free time of people to try to guess whether they are likely to play video games or deciding who should get a loan from the bank based on their age, occupation and income. Yeah, this sounds great. However, the main issue is that decision trees are not good substitutes for neural networks. The theory says that we have a generalization versus interpretability trade off situation, which means that trees that provide us good decisions overfit the training data and generalize poorly and the ones that are easy to interpret are inaccurate. So in order to break out of this trade off situation, a key idea of this piece of work is to ask the neural network to build a decision tree by taking an input dataset for training, trying to generate more training data that follows the same properties and feed all this to the decision tree. Here are some results for the classical problem of identifying digits in the MNIST dataset. As each decision is meant to cut the number of output options in half, it shows really well that we can very effectively perform the classification in only four decisions from a given input. And not only that, but it also shows what it is looking for. For instance, here we can see that the final decision before we conclude whether the input number is a 3 or 8, it looks for the presence of a tiny area that joins the ends of the 3 to make an 8. A different visualization of the Connect 4 game dataset reveals that the neural network quickly tries to distinguish two types of strategies. One, where the players start playing on the inner and one with the outer region of the board. It is shown that these trees perform better than traditional decision trees. Once more, they are only slightly worse than the corresponding neural networks, but can explain their decisions much more clearly and are also faster. In summary, the rate of progress in machine learning research is truly insane these days, and I am all for papers that try to provide us a greater understanding of what is happening under the hood. I am loving this idea in particular. We had an earlier episode on how to supercharge these decision trees via tree boosting. If you are interested in learning more about it, the link is available in the video description. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Ejonei Fahir. This technique is about creating high-resolution images from semantic maps. A semantic map is a colorful image where each of the colors denote an object class, such as pedestrians, cars, traffic signs and lights, buildings and so on. Normally, we use light simulation programs or restoration to render such an image, but AI researchers ask the question, why do we even need a renderer if we can code up a learning algorithm that synthesizes the images by itself? Whoa! This generative adversarial network takes this input semantic map and synthesizes a high resolution photorealistic image from it. Previous techniques were mostly capable of creating coarser, lower resolution images and also they were rarely photorealistic. And get this, this one produces 2K by 1K pixel outputs, which is close to full HD in terms of pixel count. If we wish to change something in a photorealistic image, we likely need a graphic designer and lots of expertise in Photoshop and similar tools. In the end, even simpler edits are very laborious to make because the human eye is very difficult to fool. An advantage of working with these semantic maps is that they are super easy to edit without any expertise. For instance, we can exert control on the outputs by choosing from a number of different possible options to fill the labels. These are often not just risky versions of the same car or road, but can represent a vastly different solution by changing the material of the road from concrete to dirt. Or, it is super easy to replace trees with buildings, all we have to do is rename the labels in the input image. These results are not restricted to outdoor traffic images, individual parts of human faces are also editable. For instance, adding a mustache has never been easier. The results are compared to a previous technique by the name Pix2Pix and against Cascaded Refinement Networks. You can see that the quality of the outputs vastly outperforms both of them and the images are also a visibly higher resolution. It is quite interesting to say that these are previous work because both of these papers came out this year. For instance, our episode on Pix2Pix came nine months ago and it has already been improved by a significant margin. The joys of machine learning research. Part of the trick is that the semantic map is not only used by itself, but a boundary map is also created to encourage the algorithm to create outputs with better segmentation. This boundary information turned out to be just as useful as the labels themselves. Part trick is to create multiple discriminator networks and run them on a variety of course to find scale images. There is much, much more in the paper, make sure to have a look for more details. Since it is difficult to mathematically evaluate the quality of these images, a user study was carried out in the paper. In the end, if we take a practical mindset, these tools are to be used by artists and it is reasonable to say that whichever one is favored by humans should be accepted as a superior method for now. This tool is going to be a complete powerhouse for artists in the industry. And by this, I mean right now, because the source code of this project is available to everyone, free of charge. In the meantime, we have an opening at our institute at the Vienna University of Technology for one PhD student and one postdoc. The link is available in the video description, read it carefully to make sure you qualify and if you apply through the email address of Professor Mikhail Vima. Make sure to mention two minute papers in your message. This is an excellent opportunity to turn your life around, live in an amazing city, learn a lot and write amazing papers. It doesn't get any better than that. That line is end of January. Thanks for watching and for your generous support and I'll see you next year.

Dear Fellow Scholars, this is two-minute papers with Karojona Ifehr. After defeating pretty much every highly ranked professional player in the game of Go, Google DeepMind now ventured into the realm of chess. They recently challenged not the best humans. No, no, no, that was long ago. They challenged Stockfish, the best computer chess engine in existence in quite possibly the most exciting chess-related event since Casper of Matches against Deep Blue. I will note that I was told by DeepMind that this is the preliminary version of the paper, so now we shall have an initial look and perhaps make a part 2 video with the newer results on the final paper drops. Alpha Zero is based on a neural network and reinforcement learning and is trained entirely through self-play after being given the rules of the game. It is not to be confused with Alpha Go Zero that played Go. It is also noted that this is not simply Alpha Go Zero applied to chess. This is a new variant of the algorithm. The differences include one, the rules of chess are asymmetric. For instance, pawns only move forward, castling is different on kingside and queenside, and this means that neural network-based techniques are less effective at it. Two, the algorithm not only has to predict a binary win or loss probability when given a move, but draws are also a possibility and that is to be taken into consideration. Sometimes a draw is the best we can do, actually. There are many more changes to the previous incarnation of the algorithm, please make sure to have a look at the paper for details. Before we start with the results and more details, a word on ILO ratings for perspective. The ILO rating is a number that measures the relative skill level of a player. Currently, the human player with the highest ILO rating, Magnus Carson, is hovering around 2800. This man played chess blindfolded against 10 opponents simultaneously in Vienna a couple years ago and won most of these games. That's how good he is. And Stockfish is one of the best current chess engines with an ILO rating over 3300. The difference of 500 ILO points means that if it were to play against Magnus Carson, it would be expected to win at least 95 games out of 100, though it is noted that there is a rule suggesting a hard cutoff at around 400 points difference. The two algorithms then played each other, Alpha Zero vs Stockfish. They were both given 60 seconds of thinking time per move, which is considered to be plenty, and that both of the algorithms take around 10 seconds at most per move. And here are the results. Alpha Zero was able to outperform Stockfish in about 4 hours of learning from scratch. They played 100 games. Alpha Zero won 28 times, drew 72 times, and never lost to Stockfish. Holy matter of papers, do you hear that? Stockfish is already unfathomably powerful compared to even the best human prodigies and Alpha Zero basically crushed it after 4 hours of self-play. And it was run with a similar hardware as Alpha Go Zero, one machine with 4 tensor processing units. This is hardly commodity hardware, but given the trajectory of the improvements we've seen lately, it might very well be in a couple of years. Note that Stockfish does not use machine learning and is a handcrafted algorithm. People like to refer to computer opponents in computer games as AI, but it is not doing any sort of learning. So, you know what the best part is? Alpha Zero is a much more general algorithm that can also play Shogi on an extremely high level, which is also referred to as Japanese chess, and this is one of the most interesting points. Alpha Zero would be highly useful even if it were slightly weaker than Stockfish, because it is built on more general learning algorithms that can be reused for other tasks without investing significant human effort. But in fact, it is more general and it also crashes Stockfish. With every paper from DeepMind, the algorithm becomes better and more and more general. I can tell you this is very, very rarely the case. Totally insanity. Two more interesting tidbits about the paper. One, all the domain knowledge the algorithm is given is stated precisely for clarity. Two, one might think that as computers and processing power increases over time, all we have to do is add more brute force to the algorithm and just evaluate more positions. If you think this is the case, have a look at this. It is noted that Alpha Zero was able to reliably defeat Stockfish while evaluating 10 times fewer positions per second. Maybe we could call this the AI equivalent of intuition. In other words, being able to identify a small number of promising moves and focusing on them. Let us run down my spine as I read this paper. Being a researcher is the best job in the world and we are even being paid for this. Unreal. This is a hot paper. There's a lot of discussions out there on this, lots of chess experts analyze and try to make sense of the games. I had a ton of fun reading and watching through some of these as always, two minute papers encourages you to explore and read more and the video description is ample in useful materials. You will find videos with some really cool analysis from Grandmaster Daniel King, International chess master Daniel Orange and the YouTube channel chess network, all quality materials. And if you have enjoyed this episode and you think that eight of these videos a month is worth a few dollars, please throw a coin our way on Patreon. Or if you favor cryptocurrencies instead, you can throw Bitcoin or Ethereum our way. Our support has been amazing as always, thanks so much for keeping with us through thick and thin, even in times when weird Patreon decisions happen. Luckily, this last one has been reverted. I'm honored to have supporters like you fellow scholars. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Kato Jolna Ife here. This is another one of those amazing papers that I am really excited about. And the reason for that is that this is in the intersection of computer graphics and machine learning, which, as you know, is already enough to make me happy, but when I first seen the quality of the results, I was delighted to see that it delivered exactly what I was hoping for. Light simulation programs are an important subfield of computer graphics where we try to create a photorealistic image of a 3D digital scene by simulating the path of millions and millions of light rays. First, we start out with a noisy image, and as we compute more paths, it slowly clears up. However, it takes a very long time to get a perfectly clear image, and depending on the scene and the algorithm, it can take from minutes to hours. In an earlier work, we had a beautiful but pathological scene that took weeks to render on several machines. If you would like to hear more about that, the link is available in the video description. So in order to alleviate this problem, many noise filtering algorithms surfaced over the years. The goal of these algorithms is that instead of computing more and more paths until the image clears up, we stop at a noisy image and try to guess what the final image would look like. This often happens in the presence of some additional depth and geometry information, additional images that are often referred to as feature buffers or auxiliary buffers. This information helps the noise filter to get a better understanding of the scene and produce higher quality outputs. Recently, a few learning-based algorithms emerged with excellent results. Well, excellent would be an understatement, since these can take an extremely noisy image that we rendered with one ray per pixel. This is as noisy as it gets I'm afraid, and it is absolutely stunning that we can still get usable images out of this. However, these algorithms are not capable of dealing with sequences of data and are condemned to deal with each of these images in isolation. They have no understanding of the fact that we are dealing with an animation. What does this mean exactly? What this means is that the network has no memory of how it dealt with the previous image and if we combine it with the fact that a trace amount of noise still remains in the images, we get a disturbing flickering effect. This is because the remainder of the noise is different from image to image. This technique uses a recurrent neural network which is able to deal with sequences of data, for instance, in our case, video. It remembers how it dealt with the previous images a few moments ago, and as a result, it can adjust and produce outputs that are temporarily stable. Computer graphics researchers like to call this spatio-temporal filtering. You can see in this camera panning experiment how much smoother this new technique is. Let's try the same footage, slow down, and see if we get a better view of the flickering. Yep, all good. Recurrent neural networks are by no means easy to train and need quite a few implementation details to get it right, so make sure to have a look at the paper for details. Temporal coherent light simulation reconstruction of noisy images from one sample per pixel. And for video. This is insanity. I would go out on a limp and say that in the very near future, we'll run learning based noise filters that take images that are so noisy, they don't even have one ray sample per pixel. Maybe one every other pixel or so. This is going to be the new milestone. If someone told me that this would be possible, when I started doing light transport as an undergrad student, I wouldn't have believed a word of it. Computer games, VR, and all kinds of real-time applications will be able to get photorealistic light simulation graphics in real-time, and temporarily stable. I need to take some time to digest this. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karojona Ifehir. In this work, a 121-layer convolution on your own network is trained to recognize pneumonia and 13 different diseases. Numonia is an inflammatory lung condition that is responsible for a million hospitalizations and 50,000 deaths per year in the U.S. alone. Such an algorithm requires a training set of formidable size to work properly. This means a bunch of input output pairs. In this case, one training sample is an input frontal x-ray image of the chest and the outputs are annotations by experts who mark which of the 14 sought diseases are present in this sample. So they say like, this image contains pneumonia here and this doesn't. This is not just a binary yes or no answer, but a more detailed heat map of possible regions that fit the diagnosis. The training set used for this algorithm contained over 100,000 images of 30,000 patients. This is then given to the neural network and its task is to learn the properties of these diseases by itself. Then after the learning process took place, previously unseen images are given to the algorithm and a set of radiologists. This is called a test set and of course it is crucial that both the training and the test sets are reliable. If the training and test set is created by one expert radiologist and then we again benchmark a neural network against a different randomly picked radiologist, that's not a very reliable process because each of the humans may be wrong in more than a few cases. Instead, the training and test set annotation data is created by asking multiple radiologists and taking a majority vote on their decisions. So now that the training and test data is reliable, we can properly benchmark a human versus a neural network. And here's the result. This learning algorithm outperforms the average human radiologist. The performance was measured in a 2D space where sensitivity and specificity were the two interesting metrics. Sensitivity means the proportion of positive samples that were classified as positive and specificity means the proportion of negative samples that were classified as negative. The crosses mean the human doctors and as you can see whichever radiologist we look at, even though they have very different false positive and negative ratios, they are all located below the blue curve which denotes the results of the learning algorithm. This is a simple diagram but if you think about what it actually means, this is an incredible application of machine intelligence. And now a word on limitations. It is noted that this was an isolated test. For instance, the radiologists were only given one image and usually when diagnosing someone, they know more about the history of the patient that may further help their decisions. For instance, a history of a strong cough and high fever is highly useful supplementary information for humans when diagnosing someone who may have pneumonia. Beyond only the frontal view of the chest, it is also standard practice to use the lateral view as well if the results are inconclusive. These views are not available in this data set and it is conjectured that it may sway the comparison towards humans. However, I'll note that this information may also benefit the AI just as much as the radiologists and this seems like a suitable direction for future work. Finally, this is not the only algorithm for pneumonia detection and it has been compared to the state of the art for all 14 diseases and this new technique came out on top on all of them. Also have a look at the paper for details because training a 121 layer neural network requires some clever shenanigans as this was the case here too. It is really delightful to see that these learning algorithms can help diagnosing serious illnesses and provide higher quality healthcare to more and more people around the world, especially in places where access to expert radiologists is limited. Everyone needs to hear about this. If you wish to help us spreading the word and telling these incredible stories to even more people, please consider supporting us on Patreon. We also know that many of you are crazy for Bitcoin so we also set up a Bitcoin address as well. Details are available in the video description. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karojona Ifehir. Let's have a look at some recent results on mural style transfer. You know the drill, we take a photo with some content, and for example, a painting with the desired style, and the output is an image where this style is applied to our content. If this is done well and with good taste, it really looks like magic. However, for pretty much all the previous techniques, there are always some mysterious styles that result in failure cases. And the reason for this is the fact that these techniques are trained on a set of style images, and if they face a style that is wildly different from these training images, the results won't be very usable. This new algorithm is also based on neural networks, and it doesn't need to be trained on these style images, but it can perform high-quality style transfer, and it works on arbitrary styles. This sounds a bit like black magic. So how does this happen exactly? First, an auto-ank order is trained for image reconstruction. An auto-ank order is a neural network where the input and output image is supposed to be the same thing. So far, this doesn't make any sense, because all the neural network does is copy and pasting the inputs to the outputs. Not very useful. However, if we reduce the number of neurons in one of the middle layers to very, very few neurons compared to the others, we get a bottleneck. This bottleneck essentially hamstrings the neural network and forces it to first come up with a highly compressed representation of an image. This is the encoder network, and then reconstruct the full image from this compressed representation. This is called the decoder network. So encoding is compression, decoding is decompression, or more intuitively reconstruction. This compressed representation can be thought of as the essence of the image, which is a very concise representation, but carefully crafted such that a full reconstruction of the image can take place based on it. Auto-ank orders are previous work, and if you would like to hear more about them, check the video description as we have dedicated an earlier episode to it. And now, the value proposition of this work comes from the fact that we don't just use the auto-ank order as is, but rip this network in half and use the encoder part on both the input style and content images. This way the concept of style transfer is much, much simpler in this compressed representation. In the end, we are not stuck with this compressed result because if you remember, we also have a decoder, which is the second part of the neural network that performs a reconstruction of an image from this compressed essence. As a result, we don't have to train this neural network on the style images, and it will work with any chosen style. Tell you, with most style transfer techniques, we are given an output image and we either have to take it or leave it because we can't apply any meaningful edits to it. A cool corollary of this design decision is that we can also get closer to our artistic vision by fiddling with parameters. For instance, the scale and weight of the style transfer can be changed on the fly to our liking. As always, the new technique is compared to a bunch of other competing algorithms. Due to the general and lightweight nature of this method, it seems to perform more consistently across a set of widely varying input styles. We can also create some mats for our target image and apply different artistic styles to different parts of it. Local parts of a style can also be transferred. Remember, the first style transfer technique was amazing, but very limited and took an hour on a state of the art graphics card in a desktop computer. This one takes less than a second and works for any style. Now as more new phones contain chips for performing deep learning, we can likely look forward to a totally amazing future where style transfer can be done in our pockets and in real time. What a time it is to be alive. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Zsolnai-Fehair. As we know from the series, neural network-based techniques are extraordinarily successful in defeating problems that were considered to be absolutely impossible as little as 10 years ago. When we'd like to use them for something, choosing the right kind of neural network is one part of the task, but usually the even bigger problem is choosing the right architecture. Typically, at a bare minimum, means the type and number of layers in the network and the number of neurons to be used in each layer. Bigger networks can learn solutions for more complex problems, so it seems that the answer is quite easy. Just throw the biggest possible neural network we can at the problem and hope for the best, but if you think that it is that easy or trivial, you need to think again. Here's why. Bigger networks come at a cost. They take longer to train and even worse, if we have networks that are too big, we bump into the problem of overfitting. Overfitting is a phenomenon when a learning algorithm starts essentially memorizing the training data without actually doing the learning. As a result, its knowledge is not going to generalize for unseen data at all. Imagine a student in a school who has a tremendous aptitude in memorizing everything from the textbook. If the exam questions happen to be the same, this student will do extremely well, but in the case of even the slightest deviations, well, too bad. Even though people like to call this road learning, there is nothing about the whole process that resembles any kind of learning at all. A smaller neural network, a less knowledgeable student who has done their homework properly would do way way better. So this is overfitting, the bane of so many modern learning algorithms. It can be kind of defeated by using techniques like L1 and L2 regularization or dropout. These often help, but none of them are silver bullets. If you would like to hear more about these, we've covered them in an earlier episode, actually two episodes. As always, the links are in the video description for the more curious fellow scholars out there. So the algorithm itself is learning, but for some reason we have to design their architecture by hand. As we discussed, some architectures, like some students, of course, significantly outperform other ones, and we are left to perform a lengthy trial and error to find the best ones by hand. So speaking about learning algorithms, why don't we make them learn their own architectures? And this new algorithm is about architecture search that does exactly that. I note that this is by far not the first crack at this problem, but it definitely is a remarkable improvement over the state of the art. It represents the neural network architecture as an organism and makes it evolve via genetic programming. This is just as cool as you would think it is and not half as complex as you may imagine at first. We have an earlier episode on genetic algorithms. I wrote some source code as well, which is available free of charge for everyone. Make sure to have a look at the video description for more on that. You'll love it. In this chart, you can see the number of evolution steps on the horizontal x-axis and the performance of these evolved architectures over time on the vertical y-axis. Finally, after taking about one and a half days to perform these few thousand evolutionary steps, the best architectures found by this algorithm are only slightly inferior to the best existing neural networks for many classical datasets, which is bloody amazing. Please refer to the paper for details and comparisons against the state of the art neural networks and other architecture search approaches, there are lots of very easily readable results reported there. Note that this is still a preliminary work and uses hundreds of graphics cards in the process. However, if you remember how it went with AlphaGo, the computational costs were cut down by a factor of 10 within a little more than a year. And until that happens, we have learning algorithms that learn to optimize themselves. This sounds like science fiction. How cool is that? Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Zonai Fahir. This one is going to be a treat. As you know, all too well after watching at least a few episodes of this series, Neural Networks offer us amazingly powerful tools to defeat problems that we didn't stand a chance against for a long, long time. We are now in the Golden Age of AI and no business or field of science is going to remain unaffected by this revolution. However, this approach comes with its own disadvantage compared to previous handcrafted algorithms. It is harder to know what is really happening under the hood. That's also kind of the advantage of Neural Networks because they can deal with complexities that we humans are not built to comprehend. But still, it is always nice to peek within a Neural Network and see if it is trying to learn the correct concepts that are relevant to our application. Maybe later we'll be able to look into a Neural Network, learn what it is trying to do, simplify it, and create a more reliable handcrafted algorithm that mimics it. What's even more, maybe they will be able to write this piece of code by themselves. So clearly, there's lots of value to be had from the visualizations, however, this topic is way more complex than one would think at first. Earlier, we talked about a technique that we called activation maximization, which was about trying to find an input that makes a given neuron as excited as possible. Here you can see what several individual neurons have learned when I trained them to recognize wooden patterns. In this first layer, it is looking for colors. Then, in the second layer, some basic patterns emerge. As we look into the third layer, we see that it starts to recognize horizontal, vertical, and diagonal patterns, and in the fourth and fifth layers, it uses combinations of the previously seen features, and as you can see, beautiful, somewhat symmetric figures emerge. If you would like to see more on this, I put a link to a previous episode in the video description. Then, a follow-up work came for multifaceted neuron visualizations that unveiled even more beautiful and relevant visualizations, a good example was showing which neuron is responsible for recognizing groceries. A new distale article on this topic has recently appeared by Christopher Ola and his colleagues at Google. Distale is a journal that is about publishing clear explanations to common interesting phenomena in machine learning research. All their articles so far are beyond amazing, so make sure to have a look at this new journal as a whole, as always, the link is available in the video description. They usually include some web demos that you can also play with. I'll show you one in a moment. This article gives a nice rundown of recent works in optimization-based feature visualization. The optimization part can take place in a number of different ways, but it generally means that we start out with a noisy image and look to change this image to maximize the activation of a particular neuron. This means that we slowly morph this piece of noise into an image that provides us information on what the network has learned. It is indeed a powerful way to perform visualization, often more informative than just choosing the most exciting images for a neuron from the training database. It unveils exactly the information the neuron is looking for, not something that only correlates with that information. There is more about not only visualizing the neurons in isolation, but getting a more detailed understanding of the interactions between these neurons. After all, a neuron network produces an output as a combination of these neuron activations, so we might as well try to get a detailed look at how they interact. Different regularization techniques to guide the visualization process towards more informative results are also discussed. You can also play with some of these web demos, for instance, this one shows the neuron activations with respect to the learning rates. There is so much more in the article I urge you to read the whole thing. It doesn't take that long and it is a wondrous adventure into the imagination of neural networks. How cool is that? If you have enjoyed this episode, you can pick up some really cool perks on Patreon, like Early Access, voting on the order of the next few episodes, or getting your name in the video description as a key contributor. This also helps us make better videos in the future, and we also use part of these funds to empower research projects and conferences. Details are available in the video description. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Károly Zsolnai-Fehér. Reinforcement learning is a technique where we have a virtual creature that tries to learn an optimal set of actions to maximize a reward in a changing environment. Playing video games, helicopter control, and even optimizing light transport simulations are among the more awesome example use cases for it. But if we train a reinforcement learner from scratch, we'll see that it typically starts out with a brute force search in the space of the simplest, lowest-level actions. This not only leads to crazy behavior early on, but is also highly ineffective, requires way more experience than humans do, and the obtained knowledge cannot be reused for similar tasks. It can learn the game it was trained on, often even on a superhuman level, but if we needed to function in a new environment, all this previous knowledge has to be thrown away. This algorithm is very much like how humans learn. It breaks down a big and complex task into sequences of smaller actions. These are called sub-polices and can be shared between tasks. Learning to walk and crawl are excellent examples of that and will likely be reused for a variety of different problems and will lead to rapid learning on new unseen tasks even if they differ significantly from the previously seen problems. Not only that, but the search space over sub-polices can easily be 100 or more times smaller than the original search space of all possible actions, therefore this kind of search is way more efficient than previous techniques. Of course, creating a good selection of sub-polices is challenging, because they have to be robust enough to be helpful on many possible tasks, but not too specific to one problem, otherwise they lose their utility. A few episodes ago, we mentioned a related technique by the name Neural Task Programming, and it seems that this one is capable of generalization not only over different variations of the same task, but across different tasks as well. These ends were trained to traverse several different mazes one after another and quickly realized that the basic movement directions should be retained. Creating more general learning algorithms is one of the holy grail problems of AI research and this one seems to be a proper, proper step towards defeating it. We are not there yet, but it's hard not to be optimistic with this incredible rate of progress each year. Really excited to see how this area improves over the next few months. The source code of this project is also available. Oh, and before we go, make sure to check out the channel of Robert Miles, who makes excellent videos about AI, and I'd recommend starting with one of his videos that you are objectively guaranteed to enjoy. If you wish to find out why, you'll see the link in the video description or just click the cat picture appearing here on the screen in a moment. If you indeed enjoyed it, make sure to subscribe to his channel. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karajola Ifehir. Image Matting is the process of taking an input image and separating its foreground from the background. It is an important preliminary step for creating visual effects where we cut an actor out from green screen footage and change the background to something else. And Image Matting is also an important part of these new, awesome portrait mode selfies where the background looks blurry and out of focus for a neat, artistic effect. To perform this properly, we need to know how to separate the foreground from the background. Matting human hair and telling accurately which hair strand is the foreground and which is the background is one of the more difficult parts of this problem. This is also the reason for many of the failure cases of the portrait mode photos made with the new iPhone and Pixel cameras. The input of this problem formulation is a colored image or video and the output is an alpha-matte where white and lighter colors encode the foreground and darker colors are assigned to the background. After this step, it is easy to separate and cut out the different layers and selectively replace some of them. Traditional techniques rely on useful heuristics like assuming that the foreground and the background are dominated by different colors. This is useful, but of course it's not always true. And clearly, we would get the best results if we had a human artist creating these alpha mats. Of course, this is usually prohibitively expensive for real world use and costs a ton of time and money. The main reason why humans are successful at this is that they have an understanding of the objects in the scene. So perhaps we could come up with a neural network-based learning solution that could replicate this ideal case. The first part of this algorithm is a deep neural network that takes images as an input and outputs an alpha-matte which was trained on close to 50,000 input output pairs. So here comes the second refinement stage where we take the output mat from the first step and use a more shallow neural network to further refine the edges and sharper details. There are a ton of comparisons in the paper and we are going to have a look at some of them and as you can see, it works remarkably well for difficult situations where many tiny hair strands are to be matted properly. If you look closely here, you can also see the minute differences between the results of the raw and refined steps. And it is shown that the refined version is more similar to the ground truth solution and is abbreviated with GT here. By the way, creating a dataset with tons of ground truth data is also a huge endeavor in and of itself. So thank you very much for the folks at alpha-matting.com for creating this dataset and you can see how important this kind of work is to make it easier to compare state-of-the-art research works more easily. Adobe was part of this research project so if everything goes well, we can soon expect such a feature to appear in their products. Also, if you are interested, we also have some nice two-minute paper shirts for your enjoyment. If you are located in the US, check two-minutepapers.com and for worldwide shipping, check the video description for the links. All photos of you wearing them are appreciated. Plus, scholarly points if it depicts you reading a paper. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Károly Zsolnai-Fehér. We have recently witnessed the emergence of neural network-based techniques that are able to synthesize all sorts of images. Our previous episode was about Nvidia's algorithm that created high-resolution images of imaginary celebrities that was a really cool application of generative adversarial networks. This architecture means that we have a pair of neural networks, one that learns to generate new images, and the other learns to tell a fake image from a real one. As they compete against each other, they get better and better without any human interaction. So we can clearly use them to create 2D images, but why stop there? Why not use this technique, for instance, to create assets for digital media? So instead of 2D images, let's try to adapt these networks to generate high-resolution 3D models of terrains that we can use to populate a virtual world. Both computer games and the motion picture industry could benefit greatly from such a tool. This process is typically done via procedural generation, which is basically a sort of guided random terrain generation. Here, we can have a more direct effect on the output without putting in tens of hours of work to get the job done. In this first training step, this technique learns how an image of a terrain corresponds to input drawings. Then we will be able to sketch a draft of a landscape with rivers, ridges, valleys, and the algorithm will output a high-quality model of the terrain itself. During this process, we can have a look at the current output and refine our drawings in the meantime, leading to a super-efficient process where we can go from a thought to high-quality final results within a few seconds without being bogged down with the technical details. Once more, it can also not only deal with erased subregions, but it can also automatically fill them with sensible information to save time for us. What an outstanding convenience feature. And the algorithm can also perform physical manipulations like erosion to the final results. After the training for the erosion step is done, the computational cost is practically zero. For instance, running an erosion simulator on this piece of data would take around 40 seconds, where the neural network can do it in 25 milliseconds. The full simulation would almost be a minute where the network can mimic its results practically instantaneously. A limitation of this technique is that if the input is too sparse, unpleasant grid artifacts may appear. There are tons of more cool features in the paper, make sure to have a look as always it is available in the video description. This is a really well thought out and well-presented work that I expect to be a true powerhouse, for terrain authoring in the future. And in the meantime, we have reached 100,000 subscribers. A hundred thousand fellow scholars. Wow! This is absolutely amazing and honestly, I never thought that this would ever happen. So, happy paperversary. Thank you very much for coming along on this journey of science and I am very happy to see that the series brings joy and learning to more people than ever. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Károly Zsolnai-Fehér. Hold on to your papers because these results are completely out of this world. You'll soon see why. In this work, high-resolution images of imaginary celebrities are generated via a generative adversarial network. This is an architecture where two neural networks battle each other. The generator network is the artist who tries to create convincing, real-looking images and the discriminator network, the critic, tries to tell a fake image from a real one. The artist learns from the feedback of the critic and will improve itself to come up with better quality images, and in the meantime, the critic also develops a sharp eye for fake images. These two adversaries push each other until they are both adept at their tasks. A classical drawback of this architecture is that it is typically extremely slow to train, and these networks are often quite shallow, which means that we get low-resolution images that are devoid of sharp details. However, as you can see here, these are high-resolution images with tons of details. So, how is that possible? So here comes the solution from scientists at NVIDIA. Initially, they start out with tiny shallow neural networks for both the artist and the critic, and as time goes by, both of these neural networks are progressively grown. They get deeper and deeper over time. This way, the training process is more stable than using deeper neural networks from scratch. It not only generates pictures, but it can also compute high-resolution intermediate images via latent space interpolation. It can also learn object categories from a bunch of training data and generate new samples. And if you take a look at the roster of scientists on this project, you will see that they are computer graphics researchers who recently set foot in the world of machine learning. And man, do they know their stuff and how to present a piece of work? And now comes something that is the absolute most important part of devaluation that should be a must for every single paper in this area. These neural networks were trained on a bunch of images of celebrities and are now generating new ones. However, if all we are shown is a new image, we don't know how close it is to the closest image in the training set. If the network is severely overfitting, it would essentially copy-paste samples from there, like a student in class who hasn't learned a single thing, just memorize the textbook. Actually, what is even worse is that this would mean that the worst learning algorithm that hasn't learned anything, but memorize the whole database would look the best. That's not useful knowledge. And here, you see the nearest neighbors, the images that are the closest in this database to the newly synthesized images. It shows really well that the AI has learned the concept of a human face extremely well and can synthesize convincing looking new images that are not just copy-pasted from the training set. The source code, pre-trained network, and one hour of imaginary celebrities are also available in the description, check them out. Premium quality service right there. And if you feel that eight of these videos a month is worth a dollar, please consider supporting us on Patreon. You can also get really cool additional perks like early access, and it helps us make better videos, grow, and tell these incredible stories to a larger audience. Details are available in the description. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Zsolnai-Fehir. One of the holy grair problems of machine learning research is to achieve artificial general intelligence, AGI in short. Deep Blue was able to defeat the genius Kasperov in chess, but it was unable to tell us what the time was. Algorithms of this type we often refer to as weak AI, or narrow AI, a technique that excels, or is maybe even on a superhuman level at a task, but has zero or no knowledge about anything else. A key to extend these algorithms would be to design them in a way that their knowledge generalizes well to other problems. This is what we call transfer learning, and this collaboration between the Stanford AI lab and Kaltek goes by the name Neural Task Programming and tries to tackle this problem. A solution to practically any problem we are trying to solve can be written as a series of tasks. These are typically complex actions, like cleaning a table or performing a backflip that are difficult to transfer to a different problem. This technique is a bit like divide and conquer type algorithms that aggressively try to decompose big difficult tasks into smaller, more manageable pieces. The smaller and easier to understand the pieces are, the more reusable they are, and the better they generalize. Let's have a look at an example. For instance, in a problem where we need to pick and place objects, these series of tasks can be decomposed into picking and placing. These can be further diced into a series of even smaller tasks such as gripping, moving, and releasing actions. However, if the learning takes place like this, we can now specify different variations of these tasks, and the algorithm will quickly understand how to adapt the structure of these small tasks to efficiently solve new problems. The new algorithm generalizes really well for tasks with different lengths, topologies, and changing objectives. If you take a look at the paper, you'll also find some more information on adversarial dynamics, which lists some problem variants where a really unpleasant adversary pushes things around on the table from time to time to mess with the program, and there are some results that show that the algorithm is able to recover from these failure states quite well. Really cool. Now, please don't take this as a complete solution for AGI, because it is a fantastic piece of work, but it's definitely not that. However, it may be a valuable puzzle piece to build towards the final solution. This is Research. We advance one step at a time. Man, what an amazing time to be alive. Thanks for watching, and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two minute papers with Karojona Ifeher. Earlier, we had an episode about OpenAI's absolutely amazing algorithm that mastered Dota 2, a competitive online battle arena game and managed to beat some of the best players in the world in a relatively limited one versus one game mode. While the full 5 versus 5 version of this learning algorithm is still in the works, scientists at OpenAI experimented with some self-play in other games and have found some remarkable results. You can see here that most of these amusing experiments take place in a made up 3D game with simulated physics. For instance, performing well with these humanoid creatures means controlling 17 actuated joints properly. These agents use a reinforcement learning algorithm to maximize a reward, for instance, a sumo warrior gets a thousand points for pushing their opponent out of the ring. The first interesting thing is that a learning curriculum was used, which means that the algorithm was allowed to explore on their own by relaxing the strict scores that are given only one winning. This is combined with the fact that these agents play against themselves led to some remarkable emergent behaviors. Here you can see with the score how much of a difference this curriculum makes. And you also see that whenever a plot is symmetric, that means that they are zero sum games. So if one agent wins a given number of points, the other loses the same amount. The self-play part is also particularly interesting, as many agents are being trained in parallel at the same time. And if we are talking about one versus one games, we have to create some useful logic to decide who to pair with whom. It seems that training against an older version of a previously challenged opponent was the best strategy. This makes sense because they are running a similar algorithm. And for self-play, this means that the algorithm is asked to defeat an older version of itself. If it can reliably do that, it will lead to a smooth and predictable learning process. It is kind of incredible to think about the fact that we have a virtual world with a bunch of simulated learning creatures, and we are omnipotent beings trying to craft the optimal learning experience for them. The perks of being a researcher in machine learning. And we are even being paid for this. Isn't this incredible? Ssh, don't tell anyone about this. There are so many interesting results here and so much to talk about. For instance, we haven't even talked about transfer learning, where these creatures learn to generalize their knowledge learned from previous tasks to tackle new challenges more efficiently. Make sure to have a look at the paper and the source code is available for everyone free of charge. If you're one of those fellow thinkers, you'll be more than happy to look into the video description. If you wish to hear more about transfer learning, subscribe and turn on notifications because the next episode is going to be about some really cool results in this area. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Zsolnai-Fehir. This is a fully-in-house Disney paper on how to teach a neural network to capture the appearance of clouds. This topic is one of my absolute favorites because it is in the intersection of the two topics I love most, computer graphics and machine learning. Hell yeah! Generally, we use light simulation programs to render these clouds and the difficult part of this is that we have to perform something that is called volumetric path tracing. This is a technique where we have to simulate rays of light that do not necessarily bounce off of the surface of objects that may penetrate their surfaces and undergo many scattering events. Understandably, in the case of clouds, capturing volumetric scattering properly is a key element in modeling their physical appearance. However, we have to simulate millions and millions of light pass with potentially hundreds of scattering events which is a computationally demanding task even in the age of rapidly improving hardware. As you can see here, the more we bump up the number of possible simulated scattering events, the closer we get to reality, but the longer it takes to render an image. In the case of bright clouds here, rendering an image like this can take up to 30 hours. In this work, a nice hybrid approach is proposed where a neural network learns the concept of in-scattered radiance and predicts it rapidly so this part we don't have to compute ourselves. It is a hybrid because some parts of the renderer are still using the traditional algorithms. The dataset used for training the neural network contains 75 different clouds, some of which are procedurally generated by a computer and some are drawn by artists to expose the learning algorithm to a large variety of cases. As a result, these images can be rendered in a matter of seconds to minutes. Normally, this would take many, many hours on a powerful computer. Here is another result with traditional path tracing. And now the same with deep scattering. Yep, that's how long it takes. The scattering parameters can also be interactively edited without us having to wait for hours to see if the new settings are better than the previous ones. Dialing in the perfect results typically takes an extremely lengthy trial and error phase which now can be done almost instantaneously. The technique also supports a variety of different scattering models. As with all results, they have to be compared to the ground truth renderings and as you can see here, they seem mostly indistinguishable from reality. It is also temporarily stable so animation rendering can take place flicker free as is demonstrated here in the video. I think this work is also a great testament to show how these incredible learning algorithms can accelerate progress in practically all fields of science. And given that this work was done by Disney, I am pretty sure we can expect tons of photorealistic clouds in their upcoming movies in the near future. There are tons of more details discussed in the paper which is remarkably well produced. Make sure to have a look, the link is in the video description. This is a proper, proper paper you don't want to miss out on this one. Also, if you enjoy this episode and you feel that the series provides you value in the form of enjoyment or learning, please consider supporting us on Patreon. You can pick up cool perks there like deciding the order of the next few episodes and you also help us make better videos in the future. Those are available in the description. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Zsolnai-Fehair. Researchers at MVIDIA have been hitting it out of the park lately, and this work is no different. This technique performs image translation, which means that the input is an image of our choice, and the output is the same image with somewhat different semantics. For instance, an image of a city can be translated to a map of this city, or a daytime photo or video can be converted to a pier as if it were shot during the night. And throughout the video, you'll see so much more of these exciting applications. A typical way of accomplishing this is done by using generative adversarial networks. This is a pair of neural networks where the generator network creates new synthetic images trying to fool a discriminator network which learns to tell a fake synthesized image from a real one. These two neural networks learn together where one tries to come up with better solutions to fool the discriminator, where the discriminator seeks to get better at telling forgeries from the real photographs. In the end, this rivalry makes both of them get better and better, and the final result is an excellent technique to create convincing image translations. In this work, not two, but six of these networks are being used, so make sure to have a look at the paper for details. There was an earlier work that was able to perform image translation by leaning on a novel cycle consistency constraint. This means that we assume that the source image can be translated to the target image, and then this target image can be translated back to look exactly like the source. This kind of means that these translations are not arbitrary and are mathematically meaningful operations. Here, the new technique builds on a novel assumption that there exists a latent space in which the input and the output images can both coexist. This latent space is basically an intuitive and concise representation of some more complicated data. For instance, earlier, we experimented a bit with fonts and had seen that even though the theory of font design is not easy, we can create a two-dimensional latent space that encodes simple properties like curvature that can describe many, many fonts in an intuitive manner. Remarkably, with this new work, converting dogs and cats into different breeds is also a possibility. Interestingly, it can also perform real to synthetic image translation and vice versa. So that means that it can create video game footage from our real world videos and even more remarkably, convert video game footage to real world video. This is insanity, one of the craziest ideas I've seen in a while. Bravo! And now, hold on to your papers because it can also perform attribute-based image translation. This means that for instance, we can grab an image of a human face and transform the model's hair to blonde, add sunglasses or smiles to it at will. A limitation of this technique is that training is still non-trivial as it still relies on generative adversarial networks, and it is not yet clear whether there is a point to which the training converges or not. The source code of this project is also available. Make sure to take a good look at the license before doing anything because it is under the Creative Commons non-commercial and no-derivatives license. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karojona Ifehir. In this series, we talk a lot about different AI algorithms that solve a variety of super-difficult tasks. These are typically tested within a software environment in the form of a simulation program. However, this often leaves the question open whether these algorithms would really work in real-world environments. So what about that? This work from OpenAI goes by the name Domain Randomization and is about training an AI on relatively crude computer simulations in a way that can be transferred to the real world. The problem used to demonstrate this was localizing and grasping objects. Note that this algorithm has never seen any real images and was trained using simulated data. It only played a computer game, if you will. Now, the question we immediately think about is what the term Domain Randomization has to do with transferring simulation knowledge into reality. The key observation is that using simulated training data is okay, but we have to make sure that the AI is exposed to a diverse enough set of circumstances to obtain knowledge that generalizes properly, hence the term Domain Randomization. In these experiments, the following parameters were heavily randomized. Number of shapes and distractor objects on the table, positions and textures on the objects, table and the environment, number of lights, material properties, and the algorithm was even exposed to some random noise as well in the images. And it turns out that if we do this properly, leaning on the knowledge of only a few thousand images, when the algorithm is uploaded to a real robot arm, it becomes capable of grasping the correct prescribed objects. In this case, the objective was spam detection. Very amusing. I think the very interesting part is that it is not even using photorealistic rendering and light simulations. These programs are able to create high quality images that resemble the real world around us, and it is mostly clear that those would be useful to train such an algorithm. However, this only uses extremely crude data and the knowledge of the AI still generalizes to the real world. How about that? Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Ejolnai-Fehir. Hold on to your papers because this work on AlphaGo is absolute insanity. In the game of Go, the players put stones on the table where the objective is to surround more territory than the opponent. This is a beautiful game that is particularly interesting for AI research because the space of possible moves is vastly larger than in chess, which means that using any sort of exhaustive search is out of question, and we have to resort to smart algorithms that are able to identify a small number of strong moves within this dependently large search space. The first incarnation of DeepMind's Go AI AlphaGo uses a combination of a policy network that is responsible for predicting the moves and the value network that predicts the winner of the game after it plays it to the end against itself. These are both deep neural networks and they are then combined with a technique called Monte Carlo 3 Search to be able to narrow down the search in this large search space. This algorithm started out with a bootstrapping process where it was shown thousands of games that were used to learn the basics of Go. Based on this, it is clear that such an algorithm can learn to be as good as formidable human players. But the big question was, how could it possibly become even better than the professionals that it has observed? How could the disciple become better than its master? The solution is that after it has learned what it can from these games, it plays against itself many many times to improve its skills. This second phase is the main part of the training that takes the most time. Let's call this base algorithm AlphaGoFan, which was used to play a guessfant way, a two-dan European Go champion who was defeated 5-0. This was a historic moment and the first time an AI beat a professional Go player without a handicap. Funway described his experience as playing a guest a very strong and stable player and he also mentioned that the algorithm felt very human like. Some voiced their doubts within the Go community and noted that the algorithm would never be able to beat Lisa Dahl and 9-dan World Champion and winner of 18 international titles. Also give you an intuition of the difference. Based on their illow points, Lisa Dahl is expected to beat Funway 97 times out of 100 games. So a few months later, DeepMind organized a huge media event where they would challenge him to play against AlphaGo. This was a slightly modified version of the base algorithm that used a deeper neural network with more layers and was trained using more resources than the previous version. There was also an algorithmic change to the policy networks, the details on this are available in the paper in the description. It is a great read. Make sure to have a look. Let's call this algorithm AlphaGo Lee. This event was watched all around the world and can perhaps be compared to the cuspere of public chess games against DeepBlue. I have the fondest memories of waking up super early in the morning jumping out of the bed in excitement to watch all these Go matches. And in a long and nail biting series, Lisa Dahl was defeated 4-1 by the AI. With significantly less media attention, the next phase came, bearing the name AlphaGo Master, which used around 10 times less tensor processing units than the AlphaGo Lee and became an even stronger player. This algorithm played against human professionals online in January 2017 and won all 60 matches it had played. This is insanity. But if you think that's it, well, hold on to your papers now. In their newest work, AlphaGo has reached its next form AlphaGo Zero. This variant does not have access to any human played games in the first phase and learns completely through self play. It starts out from absolutely nothing with just the knowledge of the rules of the game. It was trained for 40 days and by day 3, it reached the level of AlphaGo Lee this is above world champion level. Around day 29, it hits the level of AlphaGo Master, which is practically unbeatable to all human beings. And get this, at 40 days, this version surpasses all previous AlphaGo versions and defeats the previously published WorldBedar version 100-0. This has kept me up for several nights now and I am completely out of words. In this version, the two neural networks are fused into one which can be trained more efficiently. It is beautiful to see these curves as they show this neural network starting from a random initialization. It knows the rules but beyond that it is completely clueless about the game itself and this rapidly becomes practically unbeatable. And I left the best part for last. It uses only one single machine. I think it is fair to say that this is history unfolding before our eyes. What a time to be alive. Congratulations to the DeepMine team for this remarkable achievement. And for me, I love talking about research to a wider audience and it is a true privilege to be able to tell these stories to you. Thank you very much for your generous support on Patreon and making me able to spend more and more time with what I love most. Absolutely amazing. And now I know it's a bit redundant but from muscle memory, I'll sign out to usual way.

Dear Fellow Scholars, this is two-minute papers with Károly Zsolnai-Fehér. This is our 200th episode, so I know you're expecting something great. See how you like this one. One of the most sought-after effects in light transport simulations is capturing indirect illumination. This is a beautiful effect where the color of multiple diffuse, matte surfaces bleed onto each other. And of course, computing such an effect is as costly as it is beautiful, because it requires following the path of millions and millions of light rays. This usually means several hours of waiting time. There have been countless research papers written on how to do this in real time, but the limitations were often much too crippling for practical use. But this time around, you will see soon that these results are just outstanding. And we will have a word on limitations at the end of this video. The key contribution of this work is that instead of computing the light transport between all possible point pairs in the scene, it uses radiance probes that measure the nearby illumination and tries to reconstruct the missing information from this sparse set of radiance probes. After that, we place a bunch of receiver points around the scene to places where we would like to know how the indirect illumination looks. There are several things to be taken care of in the implementation of this idea. For instance, in previous works, the hierarchy of the sender and receiver points was typically fixed. In this new work, it is shown that the much sparser set of carefully placed radiance probes is sufficient to create high-quality reconstructions. This seemingly small difference also gives rise to a lot of ambiguous cases that the researchers needed to work out how to deal with. For instance, possible occlusions between the probes and receiver points need special care. The entire algorithm is explained in a remarkably intuitive way in the paper, make sure to have a look at that. And, given that we can create images by performing much less computation with this technique, we can perform real-time light simulations. As you can see, 3.9 milliseconds is a typical value for computing an entire image, which means that this can be done with over 250 frames per second. That's not only real-time, that's several times real-time, if you will. Outstanding. And of course, now that we know that this technique is fast, the next question is how accurate is it. As expected, the outputs are always compared to the reference footage, so we can see how accurate the proposed technique is. Clearly, there are differences. However, probably many of us would fail to notice that we are not looking at the reference footage, especially if we don't have access to it, which is the case in most applications. And note that normally, we would have to wait for hours for results like this. Isn't this incredible? There are also tons of more comparisons in the paper. For instance, it is also shown how the density of radiance probes relates to the output quality, and where the possible sweet spots are for industry practitioners. It is also tested against many competing solutions. Not only the results, but the number and quality of comparisons is also top tier in this paper. However, like with all research works, no new idea comes without limitations. This method works extremely well for static scenes where not a lot of objects move around. Some movement is still fine as it is shown in the video here, but drastic changes to the structure of the scene, like a large opening door that remains unaccounted for by the probes will lead to dips in the quality of the reconstruction. I think this is an excellent direction for future research works. If you enjoyed this episode, make sure to subscribe and click the bell icon. We have some more amazing papers coming up. You don't want to miss that. Thanks for watching and for your generous support. I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Zsolnai-Fehir. This work doesn't have a ton of viewable footage, but I think it is an absolutely amazing piece of craftsmanship. So in the first half of this video, we'll roll some footage from earlier episodes, and in the second half, you'll see the new stuff. In this series, we often talk about reinforcement learning, which is a learning technique where an agent chooses an optimal series of actions in an environment to maximize the score. Learning computer games is a good example of a clearly defined score that is to be maximized. As long as we can say that the higher the score, the better the learning. The concept will work for helicopter control choosing the best spot for Wi-Fi connectivity or a large variety of different tasks. However, what about environments where multiple agents or players are present? Not all games are single-player focused, and not all helicopters have to fly alone. So what about that? To deal with cases like this, scientists at OpenAI and the University of Oxford came up with a work by the name Learning with Opponent Learning Awareness, L-O-L-A in short, or Lola. I have to say that the naming game at OpenAI has been quite strong lately. This is about multiplayer reinforcement learning, if you will. This new agent does not only care about maximizing its own score, but also inserts a new term into the equation, which is about anticipating the actions of other players in the environment. It is not only possible to do this, but they also show that it can be done in an effective way, and the best part is that it also gives rise to classical strategies that game theory practitioners will immediately recognize. For instance, it can learn Tit4Tet, which is a strategy that mirrors the other player's actions. This means that if the other player is cooperative, it will remain cooperative. And if it gets screwed over, it will also try to screw others over. You'll see in a moment why this is a big deal. The Prisoner's Delama is a game where two criminals are caught and are independently interrogated and have to choose whether they snitch on the other one or not. If anyone snitches out, there will be hell to pay for the other one. If they both affect, they both serve a fair amount of time in prison. The score to be minimized is therefore this time spent in prison. And this strategy is something that we call the Nash equilibrium. In other words, this is the best set of actions if we consider the options of the other actor as well, and expect that they do the same for us. The optimal solution of this game is when both criminals remain silent. And now, the first cool result is that if we run the Prisoner's Delama with two of these new Lola agents, they quickly find the Nash equilibrium. This is great. But wait, we have talked about this Tit4Tet thing. So what's the big deal with that? There is an iterated version of the Prisoner's Delama game where the snitching or cooperating game is replayed many, many times. It is an ideal benchmark because an advanced agent would know that we cooperated the last time, so it is likely that we can park her up this time around too. And now comes the even cooler thing. This is where the Tit4Tet strategy emerges. These Lola agents know that if the previous time they cooperated, they will immediately give each other another chance and again get away with the least amount of prison time. As you can see here, the results vastly outperform other naive agents and from the scores, it seems that previous techniques enter a snitching revenge war against each other and both will serve plenty of time in prison. Other games are also benchmarked against naive and cooperative agents vastly outperforming them. This is a fantastic paper. Make sure to check it out in the video description for more details. I found it to be very readable, so do not despair if your math kung fu is not that strong. Just dive into it. Videos like this tend to get less views because they have less visual fireworks than most other works we are discussing in a series. Fortunately we are super lucky because we have your support on Patreon and can tell these important stories without worrying about going viral. And if you have enjoyed this episode and you feel that eight of these videos a month is worth a dollar, please consider supporting us on Patreon. One back is almost nothing but it keeps the papers coming. Details are available in the video description. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Ejona Efehir. Now that facial recognition is becoming more and more of a hot topic, let's talk a bit about 3D face reconstruction. This is a problem where we have a 2D input photograph or a video of a person and the goal is to create a piece of 3D geometry from it. To accomplish this, previous works often required a combination of proper alignment of the face, multiple photographs and dense correspondences, which is a fancy name for additional data that identifies the same regions across these photographs. But this new formulation is the holy grail of all possible versions of this problem because it requires nothing else but one 2D photograph. The weapon of choice for this work was a convolutional neural network and the dataset the algorithm was trained on couldn't be simpler. It was given a large database of 2D input image and 3D output geometry pairs. This means that the neural network can look at a lot of these pairs and learn how these input photographs are mapped to 3D geometry. And as you can see, the results are absolutely insane, especially given the fact that it works for arbitrary face positions and many different expressions and even with occlusions. However, this is not your classical convolutional neural network because as we mentioned, the input is 2D and the output is 3D. So the question immediately arises, what kind of data structure should be used for the output? The authors went for a 3D voxel array, which is essentially a cube in which we build up the face from small, identical, Lego pieces. This representation is similar to the terrain in the game Minecraft. Only the resolution of these blocks is finer. The process of guessing how these voxel arrays should look based on the input photograph is referred to in the research community as volumetric regression. This is what this work is about. And now comes the best part. An online demo is also available where we can either try some prepared images or we can also upload our own. So while I run my own experiments, don't leave me out of the good stuff and make sure you post your results in the comment section. The source code is also available for you fellow tinkerers out there. The limitations of this technique includes the inability of detecting expressions that are very far away from the ones seen in the training set. And as you can see in the videos, temporal coherence could also use some help. This means that if we have video input, the reconstruction has some tiny differences in each frame. Maybe a recurrent neural network, like some variant of long short term memory, could address this in the near future. However, those are trickier and more resource intensive to train properly. They are excited to see how these solutions evolve and of course, two minute papers is going to be here for you to talk about some amazing upcoming works. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two minute papers with Karo Zsolnai-Fehir. With today's graphics technology, we can enjoy many really smooth videos that were created using 60 frames per second. We love it too, and we hope that you noticed that our last 100 or maybe even more episodes have been available in 60 hertz. However, it often times happens that we are given videos that have anything from 20 to 30 frames per second. This means that if we play them on a 60 FPS timeline, half or even more of these frames will not provide any new information. As we try to slow down the videos for some nice slow motion action, this ratio is even worse, creating an extremely choppy output video. Fortunately, there are techniques that are able to guess what happens in these intermediate frames and give them to us. This is what we call frame interpolation. We have had some previous experiments in this area where we tried to create an amazing slow motion version of a video with some bubbles merging. A simple and standard way of doing frame interpolation is called frame blending, which is a simple averaging of the closest two known frames. The more advanced techniques are optical flow-based, which is a method to determine what motions happened between these two frames and create new images based on that knowledge leading to higher quality results in most cases. This technique uses a convolutional neural network to accomplish something similar, but in the end, it doesn't give us an image but a set of convolution kernels. This is a transformation that is applied to the previous and the next frame to produce an intermediate image. It is not the image itself, but the recipe of how to produce it, if you will. We've had a ton of fun with convolutions earlier where we use them to create beautiful subsurface scattering effects for translucent materials in real time and are more loyal fellow-scalers remembered that at some point I also pulled out my guitar and showed what it would sound like inside a church using a convolution-based reverberation technique. The links are available in the video description, make sure to check them out. Since we have a neural network over here, it goes without saying that the training takes place on a large number of before, after image pairs so that the network is able to produce these convolution kernels. Of course, to validate this algorithm, we also need to have access to a ground truth reference to compare against. We can accomplish this by withholding some information about a few intermediate frames, so we have the true images which the algorithm would have to reproduce without seeing it. Kind of like giving a test to a student when we already know the answers. You can see such a comparison here. And now, let's have a look at these results. As you can see, they are extremely smooth and the technique retains a lot of high frequency details in these images. The videos also seem temporalic adherent, which means that it's devoid of the annoying flickering effect where the reconstruction takes place in a way that's a bit different in each subsequent frame. Kind of that happens here, which is an excellent property of this technique. The Python source code for this technique is available and is free for non-commercial uses. I've put a link in the description if you have given it a try and have some results of your own, make sure to post them in the comments section or our subreddit discussion. The link is available in the description. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Károly Zsolnai-Fehér. In this new age of AI, there is no shortage of articles and discussion about AI safety and, of course, rightfully so. These new learning algorithms started solving problems that were previously thought to be impossible in quick succession. Only 10 years ago, if we told someone about half of the things that have been covered in the last few Two Minute Papers episodes would have been declared insane. And of course, having such powerful algorithms, we have to make sure that they are used for good. This work is a collaboration between open AI and deep-mind security team and is about introducing more human control in reinforcement learning problems. The goal was to learn to perform a backflip through reinforcement learning. This is an algorithm that tries to perform a series of actions to maximize the score, kind of like playing computer games. For instance, in Atari Breakout, if we break a lot of breaks, we get a high score, so we know we did something well. If we see that happening, we keep doing what led to this result. If not, we go back to the drawing board and try something new. But this work is about no ordinary reinforcement learning algorithm because the score to be maximized comes from a human supervisor and we are trying to teach a digital creature to perform a backflip. I particularly like the choice of the backflip here because we can tell when we see one, but the mathematical specification of this in terms of movement actions is rather challenging. This is a problem formulation in which humans can overlook and control the learning process, which is going to be an increasingly important aspect of learning algorithms in the future. The feedback option is very simple, we just specify whether the sequence of motions achieved our prescribed goal or not. Did it fall or did it perform the backflip successfully? After around 700 human feedbacks, the algorithm was able to learn the concept of a backflip, which is quite remarkable given that these binary yes-no scores are extremely difficult to use for any sort of learning. In an earlier episode, we illustrated a similar case with a careless teacher who refuses to give out points for each problem on a written exam and only announces whether we have failed or passed. This clearly makes a dreadful learning experience and it is incredible that the algorithm is still able to learn using these. We provide feedback on less than 1% of the actions the algorithm makes and it can still learn difficult concepts off of these extremely sparse and vague rewards. Low quality teaching leads to high quality learning. How about that? This is significantly more complex than what other techniques were able to learn with human feedback and it works with other games too. A word about the collaboration itself. When a company hires a bunch of super smart scientists and spends a ton of money on research, it is understandable that they want to get an edge through these projects which often means keeping the results for themselves. This leads to excessive secrecy and the lack of collaboration with other groups as everyone wants to keep their cards close to their chest. The fact that such collaborations can happen between these two AR research giants is a testament to how devoted they are to working together and sharing their findings with everyone free of charge for the greater good. Awesome. As the media is all up in arms about the demise of the human race, I feel that it is important to show the other side of the coin as well. We have top people working on AI safety right now. If you wish to help us tell these stories to more and more people, please consider supporting us on Patreon. Details are available in the video description or just click the letter P that appears on the screen in a moment. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is 2 Minute Papers with Karo Ejola Ifehir. In most video games that we've seen for at least a few moments, we'll learn to anticipate what is going to happen in the next second and even more, if given the patience and skills we could attempt to recreate parts of the game itself. And what you see here in this work is actually even better because it requires neither the patience nor the skills to do that. So here's the million dollar idea. Let's have a learning algorithm look at some video game footage and then ask it to recreate that so we can indulge in playing it. The concept is demonstrated on the Super Mario game and later you will also see some results with the millennial childhood favorite, Mega Man. There are many previous works that hook into the source code of these games and try to read and predict what happens next by reading the code level instructions. But not in this case because this technique looks at the video output and the learning takes place on the level of pixels. Therefore no access to the inner workings of the game is necessary. The algorithm is given two things for the learning process. One, a sprite palette that contains all the possible elements that can appear in the game, including landscape tiles, enemies, coins and so on. And two, we also provide an input video sequence with one playthrough of the game to demonstrate the mechanics and possible interactions between these game elements. The video is a series of frames from which the technique learns how a frame can be advanced to the next one. After it has been exposed to enough training samples, it will be able to do this prediction by itself on unknown frames that it hasn't seen before. This pretty much means that we can start playing the game that it tries to mimic. And there are similarities across many games that could be exploited and how in the learning algorithms with knowledge reused from other games, making them able to recreate even higher quality computer games, even in cases where a given scenario hasn't played out in the training footage. It used to be the privilege of computer graphics researchers to play video games during work hours, but apparently scientists in machine learning also caught up in this regard. Way to go. A word about limitations, as the predictions are not very speedy and are based off of a set of facts learned from the video sequences, it is a question as to how well this technique would generalize to more complex 3D video games. As almost all research works, this is a stepping stone, but a very important one at that, as this is a proof of concept for a really cool idea. You know the drill, a couple papers down the line and will see the idea significantly improved. The results are clearly not perfect, but it is a nice demonstration of a new concept, knowing the rate of progress in machine learning research you will very soon see some absolutely unreal results. What's even more, I expect that new levels, enemy types and mechanics will soon be synthesized to already existing games via generative adversarial networks or generative latent optimization. If you would like to hear more about these, as always, the links are available in the video description. Also, if you enjoyed this episode, please make sure to help us tell these incredible stories to more and more people by supporting us on Patreon. Your support has always been absolutely amazing. Details are available in the video description. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Ejona Efeira. This work is doing something truly remarkable. If we have a piece of audio of a real person speaking and a target video footage, it will retime and change the video so that the target person appears to be uttering these words. This is different from what we've seen a few episodes ago where scientists at Nvidia worked on synthesizing lip-sync geometry for digital characters solely relying on audio footage. The results were quite amazing. Have a look. Listen up. The train your well houses is our main objective. Well, like say, you and your men will do that. You have to go in and out very quick. This was great for animating digital characters when all we have is sound. But this time around, we are interested in reanimating the footage of real existing people. The prerequisite to do this with a learning algorithm is to have a ton of data to train on, which we have in our possession as there are many hours of footage of the former president speaking during his weekly address. This is done using a recurrent neural network. Recurrent neural networks are learning algorithms where the inputs and outputs can be sequences of data. So here, in the first part, the input can be a piece of audio with the person saying something, and it is able to synthesize the appropriate mouth shapes and their evolution over time to match the audio. The next step is creating an actual mouth texture from this rough shape that comes out from the learning algorithm, which is then used as an input to the synthesizer. Furthermore, the algorithm is also endowed with an additional pose matching module to make sure that the synthesized mouth texture aligns with the posture of the head properly. The final re-timing step makes sure that the head motions follow the speech correctly. If you have any doubts whether this is required, here are some results with and without the re-timing step. I grew up without my father around. Without re-timing, he moves randomly and appears a natural. Well, I wonder what my life would have been like if he had been a greater presence. I've also tried extra hard to be a good dad for my own daughters. Like all dads, I worry about my girls' safety all the time. You can see that this indeed substantially enhances the realism of the final footage. Even better, when combined with Google DeepMind's wave net, given enough training data, we could skip the audio footage altogether, and just write a piece of text making Obama or someone else say what we have written. The Blue Lagoon is a 1980 American romance and adventure film directed by Randall Cliser. Aspects of the Sublime in English poetry and painting, 1770-1850. There are also a ton of other details to be worked out. For instance, there are cases where the mouth moves before the person starts to speak, which is to be taken into consideration. The dreaded arms and eyes are classical examples of that. Small confines of the legal community. I think it's real important to keep the focus on the broader world out there and see that for a lot of kids, the doors that have been opened to me aren't open to them. There's also an important jaw correction step and more. This is a brilliant piece of work with many non-trivial decisions that are described in the paper. Make sure to have a look at it for details, as always, the link is available in the video description. The results are also compared to the face-to-face paper from last year that we also covered in the series. It is absolutely insane to see this rate of progress over the laps of only one year. If you have enjoyed this episode and you feel that eight of these videos a month is worth a dollar, please consider supporting us on Patreon. You can pick up some really cool perks there and it is also a great deal of help for us to make better videos for you in the future. Earlier, I also wrote a few words about the changes we were able to make because of your amazing support. Details are available in the description. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karoizona Ifeher. Using light simulation programs, we are able to populate a virtual scene with objects, assign material models to them, and create a beautiful, photorealistic image of this digital scene. The theory of light transport follows the physics of light. Therefore, these images should be indistinguishable from reality, and fortunately, they often are. However, there are some cases when we can tell them apart from real images. And the reason for this is not the inaccuracies of the light transport algorithm, but the oversimplifications used in the geometry and material models. The main issue is that in our mathematical models, these materials are often defined to be too perfect. But in reality, metals are rarely perfectly polished, and the classical material models in light transport can rarely capture these microstructures that make surfaces imperfect. This algorithm is about rendering new material models that can represent the imperfect materials like scratched coating and metallic flakes on car paint, a leather sofa, wooden floor, or a teapot. Just look at these phenomenal images. Previous techniques exist to solve this problem, but they take extremely long and are typically limited to flat surfaces. One of the main difficulties of the problem is that these tiny flakes and scratches are typically orders of magnitude smaller than a pixel, and therefore, they require a lot of care and additional computations to render. This work provides an exact, closed-form solution to this that is highly efficient to render. It is over 100 times faster than previous techniques, has less limitations as it works on curved surfaces, and it only takes 40% longer to render it compared to the standard perfect material models. Only 40% more time for this? Sign me up! It is truly incredible that we can create images of this sophistication using science. It is also highly practical as it can be plugged in as a standard material model without any crazy modifications to the simulation program. Looking forward to seeing some amazing animations using more and more realistic material models in the near future. If you would like to learn more about light simulations, I have been holding a full master-level course on it at the Technical University of Vienna for a few years now. After a while, I get a strong feeling that the teachings shouldn't only be available for the lucky 30 people in the classroom who can afford a college education. The teachings should be available for everyone. And now, the entirety of this course is available free of charge for everyone where we learn the theory of light from scratch and implement a really cool light simulation program together. If you want to solve a few infinite dimensional integrals with me, give it a go. As always, details are available in the video description. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Zsolnai-Fehir. Reinforcement learning is an awesome algorithm that is able to play computer games, navigate helicopters, hit a baseball, or even defeat gold champions when combined together with a neural network and Monte Carlo research. It is a quite general algorithm that is able to take on a variety of difficult problems that involve observing an environment and coming up with a series of actions to maximize the score. In a previous episode, we had a look at deep minds algorithm where a set of movement actions had to be chosen to navigate in a difficult 3D environment efficiently. The score to be maximized was the distance measured from the starting point, the further our character went, the higher score it was given, and it has successfully learned the concept of locomotion. Really cool. A prerequisite for a reinforcement learner to work properly is that it has to be given informative reward signals. For instance, if we go to a written exam as an output, we would like to get a detailed breakdown of the number of points we got for each problem. This way we know where we did well and which kinds of problems need some more work. However, imagine having a really careless teacher who never tells us the points but would only tell us whether we have failed or passed. No explanation, no points for individual tasks, no telling whether we failed by a lot or just a tiny bit. Nothing. Just attempt, we failed. Next time, we failed again, and again, and again, and again. Now this would be a dreadful learning experience because we would have absolutely no idea what to improve. Clearly, this teacher would have to be fired. However, when formulating a reinforcement learning problem, instead of using more informative scores, it is much easier to just tell whether the algorithm was successful or not. It is very convenient for us to be this careless teacher. Otherwise, what score would make sense for a helicopter control problem when we almost crash into a tree? This part is called reward engineering, and the main issue is that we have to adapt the problem to the algorithm, where the best would be if the algorithm would adapt to the problem. This has been a long-standing problem in reinforcement learning research and the potential solution would open up the possibility of solving even harder and more interesting problems with learning algorithms. And this is exactly what researchers at OpenAI tried to solve by introducing hindsight experience replay, HER or HER in short. Very apt. This algorithm takes on problems where the scores are binary, which means that it either passed or failed the prescribed task. A classic careless teacher scenario. And these rewards are not only binary, but very sparse as well, which further exacerbates the difficulty of the problem. In the video, you can see a comparison with a previous algorithm with and without the HER extension. The higher the number of epochs you see above, the longer the algorithm was able to train. The incredible thing here is that it is able to achieve a goal even if it had never been able to reach it during training. The key idea is that we can learn just as much from undesirable outcomes as from desirable ones. Let me quote the authors. The authors are not only learning how to play hockey and are trying to shoot a puck into a net. You hit the puck, but it misses the net on the right side. The conclusion drawn by a standard reinforcement learning algorithm in such a situation would be that the performed sequence of actions does not lead to a successful shot and little if anything would be learned. It is however possible to draw another conclusion, namely that this sequence of actions would be successful if the net had been placed further to the right. They have achieved this by storing and replaying previous experiences with different potential goals. As always, the details are available in the paper, make sure to have a look. Now it is always good to test things out whether the whole system works well in software. However, its usefulness has been demonstrated by deploying it on a real robot arm. You can see the goal written on the screen alongside with the results. A really cool piece of work that can potentially open up new ways of thinking about reinforcement learning. After all, it's great to have learning algorithms that are so good they can solve problems that we formulate in such a lazy way that we'd have to be fired. And here's a quick question. Do you think 8 of these episodes a month is worth a dollar? If you have enjoyed this episode and your answer is yes, please consider supporting us on Patreon. Details are available in the video description. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karoizhou naifahir. In many previous episodes, we talked about generative adversarial networks, a recent new line in machine learning research with some absolutely fantastic results in a variety of areas. They can synthesize new images of animals, create 3D models from photos, or dream up new products based on our edits of an image. A generative adversarial network means that we have two neural networks battling each other in an arms race. The generator network tries to create more and more realistic images, and these are passed to the discriminator network, which tries to learn the difference between real photographs and fake forged images. During this process, the two neural networks learn and improve together until they become experts at their own craft. And as you can see, the results are fantastic. However, training these networks against each other is anything that roses and sunshine. We don't know if the process converges, or if we reach Nashak-Willibrium. Nashak-Willibrium is a state where both actors believe they have found an optimal strategy while taking into account the other actors' possible decisions and neither of them have any interest in changing their strategy. This is a classical scenario in game theory, where two convicted criminals are pondering whether they should sneak on each other without knowing how the other decided to act. If you wish to hear more about the Nashak-Willibrium, I've put a link to Khan Academy's video in the description, make sure to check it out, you'll love it. I find it highly exciting that there are parallels in AI and game theory. However, the even cooler thing is that here, we try to build a system where we don't have to deal with such a situation. This is called generative latent optimization, GLO in short, and it is about introducing tricks to do this by only using a generator network. If you have ever read up on font design, you know that it is a highly complex field. However, if we'd like to create a new font type, what we are typically interested in is only a few features, like how curvy they are, or whether we are dealing with a serif kind of font and simple descriptions like that. The same principle can be applied to human faces, animals, and most topics you can imagine. This means that there are many complex concepts that contain a ton of information, most of which can be captured by a simple description with only a few features. This is done by projecting this high-dimensional data onto a low-dimensional latent space. This latent space helps eliminating adversarial optimization, which makes this system much easier to train. And the main selling point is that it still retains the attractive properties of generative adversarial networks. This means that it can synthesize new samples from the Learned data set. If it had learned the concept of birds, it will be able to synthesize new bird species. It can perform continuous interpolation between data points. This means that, for instance, we can produce intermediate states between two chosen furniture types or light fixtures. It is also able to perform simple arithmetic operations between any number of data points. For instance, if A is males with sunglasses, B are males without sunglasses, and C are females, then A minus B plus C is going to generate females in sunglasses. It can also do super-resolution and much, much more. Make sure to have a look at the paper in the video description. Now, before we go, we shall address the elephant in the room. These images are tiny. Our season follow scholars know that for generative adversarial networks, there are plenty of works on how to synthesize high-resolution images with more details. This means that this is a piece of work that opens up exciting new horizons, but it is not to be measured against the tenth follow-up work on top of a more established line of research. Two-minute papers will be here for you to keep you updated on the progress, which is, as we know, staggeringly quick in machine learning research. Don't forget to subscribe and click the bell icon to never miss an episode. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Károly Zsolnai-Fehér. We have talked about some awesome previous works where we use learning algorithms to teach digital creatures to navigate in complex environments. The input is a terrain and a set of joints, feet and movement types, and the output has to be a series of motions that maximizes some kind of reward. This previous technique borrowed smaller snippets of movements, from a previously existing database of motions and learned to stitch them together in a way that looks natural. And as you can see, these results are phenomenal. And a selling point of this new one, which you might say looks less elaborate. However, it synthesizes them from scratch. The problem is typically solved via reinforcement learning, which is a technique that comes up with a series of decisions to maximize a prescribed score. This score typically needs to be something reasonably complex. Otherwise, the algorithm is given too much freedom to maximize it. For instance, we may want to teach a digital character to run or jump hurdles, but it may start crawling instead, which is still completely fine if our objective is too simple. For instance, just maximizing the distance from the starting point. To alleviate this, we typically resort to reward engineering, which means that we add additional terms to this reward function to regularize the behavior of these creatures. For instance, we can specify that throughout these motions, the body has to remain upright, which likely favors locomotion type solutions. However, one of the main advantages of machine learning is that we can reuse our solutions for a large set of problems. If we have to specialize our algorithm for all terrain and motion types and different kinds of games, we lose out on the biggest advantage of learning techniques. So researchers at DeepMind decided that they are going to solve this problem with a reward function, which is nothing else but forward progress. That's it. The further we get, the higher score we obtain. This is amazing because it doesn't require any specialized reward function, but at the same time, there are a ton of different solutions that get us far in these terrains. And as you see here, beyond bipeds, a bunch of different agent types are supported. The key factors to make this happen is to apply two modifications to the original reinforcement learning algorithm. One makes the learning process more robust and less dependent on what parameters we choose and the other one makes it more scalable, which means that it is able to efficiently deal with larger problems. Furthermore, the training process itself happens on a rich, carefully selected of challenging levels. Make sure to have a look at the paper for details. A byproduct of this kind of problem formulation is, as you can see, that even though this humanoid does its job with a slower body well, but in the meantime, it is flailing its arms like a madman. The reason is likely because there is not much of a difference in the reward between different arm motions. This means that we most likely get through a maze or a height field even when flailing. Therefore, the algorithm doesn't have any reason to favor more natural looking movements for the upper body. It will probably choose a random one, which is highly unlikely to be a natural motion. This creates high quality, albeit amusing results that I am sure some residents of the internet will honor with a sped up remix video with some Benihil music. In summary, no pre-computed motion database, no handcrafting of rewards, and no additional wizardry needed. Everything is learned from scratch with a few small modifications to the reinforcement learning algorithm. Highly remarkable work. If you've enjoyed this episode and would like to help us and support the series, have a look at our Patreon page. Details and cool perks are available in the video description, or just click the letter P at the end of this video. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Ejone Fahir. As you know from this series, fluid simulation techniques that are able to create high-fidelity video footage are in abundance in computer graphics research. These techniques all have their own trade-offs, and when we evaluate them, we often use terms like first or second-order accuracy, which are mathematical terms. We often have to evaluate these techniques against each other by means of mathematics, because this way we can set up consistent and unbiased comparisons that everyone understands and agrees upon. However, ultimately, in the show business, what matters is how the viewers perceive the end result, whether they think it looks fake, or if it keeps their suspension of this belief. We have the choice of not only simulation techniques, but all of them also have their own set of parameters. For instance, the higher the resolution of our simulations, the more high-frequency details appear in the footage. However, after a point, increasing the resolution further is extremely costly, and while we know what is to be gained in terms of mathematics, it is still unknown how well it would do with the users. So the ultimate question is this, what do I get for my money and time? This paper provides an exhaustive user study to answer this question, where the users are asked to look at two different simulations and as a binary choice, tell us which is the one they perceived to be closer to the reference. The reference footage is a real-world video of a water sloshing in a tank and the other footage that is to be judged is created via a fluid simulation algorithm. Turns out that the reference footage can be almost anything, as long as there are some splashing and sloshing going on in it. It also turns out that after a relatively favorable breaking point, which is denoted by 2x, further increasing the resolution does not make a significant difference in the user's course. But boy, does it change the computation times? So this is why such studies are super useful, and it's great to see that the accuracy of these techniques are measured both mathematically and also how convincing they actually look for users. Another curious finding is that if we deny access to the reference footage, we see a large change in different responses and a similar jump in ambiguity. This means that we are reasonably bad at predicting the fine details. Therefore if the simulation pushes the right buttons, the users will easily believe it to be correct even if it is far away from the ground truth solution. Here's a matrix with a ton of rendered footage. Horizontally you see the same thing with different simulation techniques and vertically we slowly go from transparent above to opaque below. To keep things fair and really reveal which choices are the best bank for the buck, there are also comparisons between techniques that have a similar computation time. In these cases the Floyd implicit particle, flip in short and the affine particle in cell are almost unanimously favorable. These are advanced techniques that combine particle and grid-based simulations. I think this is highly useful information for more time critical applications, so make sure to have a look at the paper for details. There are similar user studies with glossy and translucent material models and much more in the paper. The source code of this project is also available. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karojona Ifehir. This work is about using AI to create super detailed smoke simulations. Typically, creating a crude simulation doesn't take very long, but as we increase the resolution, the execution time and memory consumption skyrockets. In the age of AI, it only sounds logical to try to include some learning algorithms in this process. So what if we had an AI-based technique that would have some sort of understanding of smoke simulations, take our crude data and add the fine details to it. This way, we could obtain a high resolution smoke simulation without waiting several days or weeks for the computation. Now if you're a seasoned Fellow scholar, you may remember an earlier work by the name Wavelet Turbulence, which is one of my favorite papers of all time. So much so that it got the distinction of being showcased in the very first two-minute papers episode. I was a sophomore college student back then when I first seen it and was absolutely shocked by the quality of the results. That was an experience I'll never forget. It also won a technical Oscar award and it is not an overstatement to say that this was one of the most influential works that made me realize that research is my true calling. The link to the first episode is available in the video description and if you want to see how embarrassing it is, make sure to check it out. It did something similar, but instead of using AI, it used some heuristics that describe what is the ratio and distribution of smaller and bigger vertices in a piece of fluid or smoke. Using this information, it could create a somewhat similar effect, but ultimately that technique had an understanding of smoke simulations in general. And it didn't know anything about the scene that we have at hand right now. Another work that is related to this is showing a bunch of smoke simulation videos to an AI and teach it how to continue these simulations by itself. I would place this work as a middle ground solution because this work says that we should take a step back and not try to synthesize everything from scratch. Let's create a database of simulations, dice them up into tiny, tiny patches, look at the same footage in low and high resolutions and learn how they relate to each other. This way we can hand the neural network some low resolution footage and it will be able to make an educated guess as to which high resolution patch should be the best match for it. When we found the right patch, we just switched the core simulation to the most fitting high resolution patch in the database. You might say that in theory, creating such a Frankenstein smoke simulation sounds like a dreadful idea. But have a look at the results as they are absolutely brilliant. And as you can see, it takes a really crude base simulation and adds so many details to it, it's truly an incredible achievement. One neural network is trained to capture similarities in densities and one for vorticity. Using the two neural networks in tandem, we can take a low resolution fluid flow and synthesize the fine details on top of it in a way that is hardly believable. It also handles boundary conditions, which means that these details are correctly added even if our smoke puff hits an object. This was an issue with wavelet turbulence which had to be addressed with several follow-up works. There are also comparisons against this legendary algorithm and as you can see, the new techniques smokes it. However, it took 9 years to do this. This is exactly 9 eternities in the world of research, which is a huge testament to how powerful the original algorithm was. It is also really cool to get more and more messages where I get to know more about you fellow scholars. I was informed that the series is used in school classes in Brazil. It is also used to augment college education and it is a great topic for fun family conversations over dinner. That's just absolutely fantastic. Loving the fact that the series is an inspiration for many of you. Thanks for watching and for your generous support. See you next Thursday atríol slowly intro.

Dear Fellow Scholars, this is two-minute papers with Karo Ejona Efeir. This work is about changing already existing images and videos by adding new objects to them or editing their physical attributes. These editable attributes include gravity, mass, stiffness, or we can even add new physical forces to the scene. For instance, we can change the stiffness of the watches in this dali painting and create an animation from it. Physically accurate animations from paintings. How cool is that? This is approaching science fiction levels of craziness. Animating a stationary clothesline by adding a virtual wind effect to the scene or bending a bridge by changing its mass is also a possibility. The first reaction I had when I've looked at this work was, are you kidding me? You can't edit a photograph, especially that I've seen plenty of earlier works that try to do something similar, but each time the limitations were just too crippling for real-world usage. And the ultimate question is always, how much user interaction does this need? Is this trivial to use or is it a laborious process? What we need to do is roughly highlight the outline of the object that we'd like to manipulate. The algorithm uses a previously published technique to make sure that the outlines are accurately captured and then tries to create a 3D digital model from the selected area. We need one more step where we align the 3D model to the image or video input. Finally, the attribute changes and edits take place not on the video footage but on this 3D model through a physics simulation technique. A truly refreshing combination of old and new techniques with some killer applications, loving it. The biggest challenge is to make sure that the geometry and the visual consistency of the scene is preserved through these changes. There are plenty of details discussed in the paper, make sure to have a look at that, the link to it is available in the video description. As these 2D photo to 3D model generator algorithms improve, so will the quality of these editing techniques in the near future. Our previous episode was on this topic, make sure to have a look at that. Also, if you would like to get more updates on the newest and coolest works in this rapidly improving field, make sure to subscribe and click the bell icon to be notified when new 2 minute papers videos come up. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is 2 Minute Papers with Károly Zsolnai-Fehér. Today, we are going to talk about a task that humans are remarkably good at, but learning algorithms mostly flounder, and that is creating 3D geometry by looking at a 2D color image. In video games and animation films, this is a scenario that comes up very often. If we need a new weapon model in the game, we typically give the artist a photo, who will sit down with a 3D modeler program and spend a few hours sculpting a similar 3D geometry. And I will quickly note that our binocular vision is not entirely necessary to make this happen. We can look at 2D images all day long and still have a good idea about the shape of an airplane, even with one eye closed. We had previous episodes on this problem, and the verdict was that the results with previous techniques are great, but not very detailed. Some mathematicians like to say that this algorithm has a cubic complexity, or cubic scaling, which means that if we wish to increase the resolution of the 3D model, just a tiny bit. We have to wait not a tiny bit longer, but significantly longer. And the cubic part means that this trade-off becomes unbearable, even for moderately high resolutions. This paper offers a technique to break through this limitation. This new technique still uses a learning algorithm to predict the geometry, but it creates these 3D models hierarchically. This means that it starts out approximating the coarse geometry of the output and restarts the process by adding more and more fine details to it. The geometry becomes more and more refined over several steps. Now this refinement doesn't just work unless we have a carefully designed algorithm around it. The refinement happens by using additional information in each step from the created model. Namely, we imagine are predicted 3D geometry as a collection of small blocks, and each block is classified as either free space, occupied space, or as a surface. After this classification happened, we have the possibility to focus our efforts on refining the surface of the model, leading to a significant improvement in the execution time of the algorithm. As a result, we get 3D models that are of higher quality than the ones offered by previous techniques. The outputs are still not super high resolution, but they capture a fair number of surface detail. And you know the drill, research is a process, and every paper is a stepping stone. And this is one of those stepping stones that can potentially save many hours of work for 3D artists in the industry. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars,這兩位是Papers with Karo and Jona and Fahir你們真討厭,吃飽了和足了,就做哪看我說話這件事是在做了一場電影的時間這就說了一場電影的電影我們在做了一場電影的電影這就說了一場電影的電影的電影這就說了一場電影的電影這就是一場電影的電影這就是一場電影的電影這就是一場電影的電影的電影的電影這就是一場電影的電影的電影這就是一場電影的電影的電影這就是一場電影的電影這就是一場電影的電影這就是一場電影的電影這就是一場電影的電影這就是一場電影的電影這就是一場電影的電影這就是一場電影的電影這就是一場電影的電影這就是一場電影的電影這就是一場電影的電影這就是一場電影的電影這就是一場電影的電影這就是一場電影的電影這就是一場電影這就是一場電影的電影這就是一場電影的電影這就是一場電影的電影這就是一場電影這就是一場電影這就是一場電影這就是一場電影這就是一場電影這就是一場電影這就是一場電影這就是一場電影這就是一場電影這就是一場電影這就是一場電影然后说什么每个字都在这里然后那种字幕可以用这个技术为了一个确实的形式说什么所以我们可以从确实的形式上然后把形式放在一个确实的形式上用这个确实的形式鞋子是一块儿的用链锋的链锋用链锋的链锋的链锋这个确实的形式上和形式上和形式上和形式上和形式上和形式上和形式上和形式上和形式上和形式上这种形式我们得到一个确实的形式用链锋和众用链锋和众用链锋和众用链锋和众用链锋和众用链锋和众用链锋和众用链锋和众用链锋和众用链锋和众用链锋和众用链锋和众用链锋在旁边看那边,有两个人的影响会设定了怎么说的?让我们看一看,让我们看三个绿色的影响会设定了让我们看一看,让我们看三个绿色的影响会设定了让我们看一看,让我们看一看,让我们看一看让我们看一看,让我们看一看让我们看一看,让我们看一看让我们看一看,让我们看一看让我们看一看,让我们看一看让我们看一看让我们看一看让我们看一看让我们看一看让我们看一看让我们看一看让我们看一看让我们看一看让我们看一看让我们看一看和以调整的这种团团一起拿出来的抽象和以它的特技和以它的特技和以它的特技和以它的特技和以它的特技和以它的特技和以它的特技和以它的特技和以它的特技和以它的特技和以他的特技和以他的特技和以他的特技和以他的特技和以他的特技和以他的特技和以他的特技和以他的特技还有以他的特技和以他的特技还有以他的特技和以他的特技还有以他的特技和以他的特技还有以他的特技还有以他的特技还有以他的特技和以他的特技最高的一切是比较比较比较的而 result 没有比较比较这种方法是不太大但我没有认为一个箱子的箱子或箱子在这种箱子没有比较比较而这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子在这种箱子你会做这种箱子你会要去到很快的我希望你会到最后的箱子然后牙利能够得到用了一条影片你会拿出的箱子像看这种箱子在最后的箱子这种箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子在最后的箱子真是为了我给你拿出来的箱子拿出来的箱子我给你拿出来的箱子我给你拿出来的箱子我给你拿出来的箱子我给你拿出来的箱子我给你拿出来的箱子我给你拿出来的箱子我给你拿出来的箱子我给你拿出来的箱子我给你拿出来的箱子

Dear Fellow Scholars, this is two-minute papers with Kato Ejolna Ifehir. In our earlier episodes, when it came to learning techniques, we almost always talked about supervised learning. This means that we give the algorithm a bunch of images and some additional information, for instance, that these images depict dogs or cats. Then, the learning algorithm is exposed to new images that it had never seen before and has to be able to classify them correctly. This kind of like a teacher sitting next to a student providing supervision. Then, the exam comes with new questions. This is supervised learning, and as you have seen from more than 180 episodes of two-minute papers, there is no doubt that this is an enormously successful field of research. However, this means that we have to label our datasets, so we have to add some additional information to every image we have. This is a very laborious task, which is typically performed by researchers, or through crowdsourcing, both of which takes a lot of funding and hundreds of work hours. But if we think about it, we have a ton of videos on the internet. You always hear these mind-melting new statistics on how many hours of video footage is uploaded to YouTube every day. Of course, we could hire all the employees in the world to annotate these videos frame-by-frame to tell the algorithm that this is a guitar, this is an accordion, or a keyboard, and we would still not be able to learn on most of what's uploaded. But it would be so great to have an algorithm that can learn on unlabeled data. However, there are learning techniques in the field of unsupervised learning, which means that the algorithm is given a bunch of images or any media and is instructed to learn on it without any additional information. There is no teacher to supervise the learning. The algorithm learns by itself. And in this work, the objective is to learn both visual and audio-related tasks in an unsupervised manner. So for instance, if we look at this layer of the visual subnetwork, we'll find neurons that get very excited when they see, for instance, someone playing on accordion. And each of the neurons in this layer belong to different object classes. I surely have something like this for papers. And here comes the katoi goes crazy part one. This technique not only classifies the frames of the videos, but it also creates semantic heat maps, which show us which part of the image is responsible for the sounds that we hear. This is insanity. To accomplish this, they ran a vision subnetwork on the video part and a separate audio subnetwork to learn about the sounds. And at the last step, all this information is fused together to obtain katoi goes crazy part two. This makes the network able to guess whether the audio and the video stream correspond to each other. It looks at a man with a fiddle, listens to a sound clip, and will say whether the two correspond to each other. Wow! The audio subnetwork also learned the concept of human voices, the sound of water, wind, music, live concerts, and much much more. And the answer is yes, it is remarkably close to human level performance on sound classification. And all this is provided by the two networks that were trained from scratch and no supervision is required. We don't need to annotate these videos. Nailed it. And please don't get this wrong, it's not like DeepMind has suddenly invented unsupervised learning. Not at all. This is a field that has been actively researched for decades. It's just that we rarely see really punchy results like these ones here. Truly incredible work. If you enjoyed this episode and you feel that 8 of these videos a month is worth a dollar, please consider supporting us on Patreon. Details are available in the video description. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Zsolnai-Fehir, creating a photorealistic image with fur and hair is hard. It is typically done by using light simulation programs where we use the laws of physics to simulate the path of millions and millions of light rays as they bounce off of different objects in the scene. This typically takes from minutes to hours if we are lucky. However, in the presence of materials like hair and fur, this problem becomes even more difficult because fur fibers have inner scattering media. This means that we not only have to bounce these rays off of the surface of objects, but also have to simulate how light is transmitted between these inner layers. And initially, we start out with a noisy image, and this noise gets slowly eliminated as we compute more and more rays for the simulation. SPP means samples per pixel, which is the number of rays we compute for each pixel in our image. You can see that with previous techniques, using 256 samples per pixel leads to a very noisy image, and we need to spend significantly more time to obtain a clear, converged image. And this new technique enables us to get the most out of our samples, and if we render an image with 256 SPP, we get a roughly equivalent quality to a previous technique using around 6 times as many samples. If we had a film studio and someone walked up on us and said that we can render the next guardians of the Galaxy film 6 times cheaper would surely be all over it. This would save us millions of dollars. The main selling point is that this work introduces a multi-scale model for rendering hair and fur. This means that it computes near and far-field scattering separately. The far-field scattering model contains simplifications, which means that it's way faster to compute. This simplification is sufficient if we look at the model from afar, or we look closely at the hair model that is way thinner than human hair strands. The near-field model is more faithful to reality, but also more expensive to compute. And the final, most important puzzle piece is stitching together the two. Whenever we can get away with it, we should use the far-field model and compute the expensive near-field model only when it makes a difference visually. And one more thing, as these hamsters get closer or further away from the camera, we need to make sure that there is no annoying jump when we are switching models. And as you can see, the animations are buttery smooth. And when we look at it, we see beautiful rendered images, and if we didn't know a bit about the theory, we would have no idea about the multi-scale wizardry under the hood. Excellent work. The paper also contains a set of deck compositions for different lightpaths. For instance, here you can see a fully rendered image on the left, and different combinations of light reflection and transmission events. For instance, R stands for one light reflection, TT for two transmission events, and so on. The S in the superscript denotes light scattering events. Adding up all the possible combinations of these T's and R's, we get the photorealistic image on the left. That's really cool, loving it. If you would like to learn more about light simulations, I'm holding a full, master-level course on it at the Technical University of Vienna. And the entirety of this course is available free of charge for everyone. I got some feedback from you fellow scholars that you watched it, and enjoyed it quite a bit. Give it a go. As always, the details are available in the video description. Thanks for watching, and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Karo Ejona Efehir. This topic has been perhaps the most highly anticipated by you Fellow Scholars, and I am extremely excited to show you the first joint paper between Deep Mind and Blizzard on creating an AI program to play Starcraft 2. Hell yeah! Unfortunately we have a paper where every detail is meticulously described, so there's much less room for misunderstandings. And before we start, note that this is a preliminary work, so please don't expect superhuman performance. However difficult you thought this problem was, you'll see in a minute that it's way more complex than most people would think. But before we start, what is Starcraft 2? It is a highly technical strategy game which will be a huge challenge to write a formidable AI for because of three reasons. One, we have imperfect information with a partially observed map. If we wish to see what the opponent is up to, we have to devote resources to scouting, which may or may not be successful depending on the vigilance of the other player. Two, we need to select and control hundreds of units under heavy time pressure. One wrong decision and we can quickly lose most of our units and become unable to recover from it. And three, perhaps the most important part. Long-term strategies need to be developed where a poor decision in the early game can lead to a crushing defeat several thousands of actions later. These cases are especially difficult to identify and learn. However, we ran a bit too far ahead to the gameplay part. What needs to be emphasized is that there is a step number one before that. And that step number one is making sure that the AI is able to communicate and interact with the game, which requires a significant engineering effort. In this paper, a Python-based interface is described to make this happen. It is great to have companies like DeepMind and OpenAI who are devoted to lay down the foundations for such an interface, which is a herculean task. This work would likely have never seen the light of day if AI research would only take place in academia. Huge respect and much thanks for the DeepMind guys for making this happen. To play the game, Deep Reinforcement Learning is used, which you heard about earlier in the series. This is a powerful learning algorithm where a neural network is used to process the video input and is combined with a reinforcement learner. With reinforcement learning, we are observing the environment around us and choose the next action to maximize the score or reward. However, defining score was very easy in Atari Breakout because we knew that if the number of our lives drops to zero, we lost. And if we break a lot of breaks, our score improves. Simple. Not so much in StarCraft 2 because how do we know exactly if we are winning? What is the score we are trying to maximize? In this work, there are two definitions for score. One that we get to know at the very end that describes whether we won, had a tie or lost. This is the score that ultimately matters. However, this information is not available throughout the game to drive the reinforcement learner, so there is an intermediate score that is referred to as Blizzard Score in the paper, which involves a weighted sum of current resources and upgrades as well as our units and buildings. This sounds good for a first approximation since it is monotonically increasing if we win battles and manage our resources well and decreases when we are losing. However, there are many matches where the player with the more resources does not have enough time to spend it and ultimately loses a deciding encounter. It remains to be seen whether this is exactly what we need to maximize to beat a formidable human player. There are also non-trivial engineering decisions on how to process the video stream. The current system uses a set of feature layers, which encode relevant information for the AI, such as terrain height, the camera location, heatpoints for the units on the screen and much, much more. There is a huge amount of information that the convolutional neural network has to make sense of. And I think it is now easy to see that starting out with throwing the AI in the deep water and expecting it to perform well on a full one versus one match at this point is a for loan effort. The paper describes a set of mini-games where the algorithm can learn different aspects of the game in isolation, such as picking up mineral shards scattered around the map, defeating enemy units in small skirmishes, building units or harvesting resources. In these mini-games, the AI has reached the level of a novice human player which is quite amazing given the magnitude and the complexity of the problem. The authors also encourage the community to create more mini-games for the AI to train on. I really love the openness and the community-effort aspects of this work. And with only just scratched the surface, there is so much more in the paper with a lot more non-trivial design decisions and the database with tens of thousands of recorded games. And the best part is that the source code for this environment is available right now for the fellow tinkerers out there. I've put a link to this in the video description. This is going to be one heck of a challenge for even the brightest AI researchers of our time. I can't wait to get my hands on the code and also I am very excited to read some follow-up papers on this. I expect there will be many of those in the following months. In the meantime, as we know, OpenAI is also working on Dota with remarkable results and there's lots of discussion whether a Dota 5 vs 5 or a StarCraft 2 1 vs 1 game is more complex for the AI to learn. If you have an opinion on this, make sure to leave a comment below this video, which is more complex. Why? This also signals that there's going to be tons of fun to be had with AI and video games this year. Stay tuned. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karojona Ifeher. In this series, we talk a lot about photorealistic rendering, which is one of the most exciting areas in computer graphics research. Photo-realistic rendering means that we put virtual objects in a scene, assign material models to them, and then run a light simulation program to create a beautiful image. This image depicts how these objects would look like in reality. This is particularly useful for the film industry, because we can create highly realistic scenes and set them up in a way that we couldn't do in real life. We can have any possible object we can imagine, light sources that we wouldn't ever be able to buy, change the time of the day, or even the planet we are on. Practically we have an infinite budget. That's amazing. However, creating such an image takes a long time, often in the order of hours, two days. You can see me render an image, and even if the footage is sped up significantly, you can see that this is going to take a long, long time. SPP means samples per pixel, so this is the number of light rays we compute for every pixel. The more SPP, the cleaner, more converged image we get. This technique performs spatial temporal filtering. This means that we take a noisy input video and try to eliminate the noise and try to guess how the final image would look like. And it can create almost fully converged images from extremely noisy inputs. Well, as you can see, these videos are created with one sample per pixel, which is as noisy as it gets. These images with the one sample per pixel can be created extremely quickly, in less than 10 milliseconds per image, and this new denoiser also takes around 10 milliseconds to reconstruct the final image from the noisy input. And yes, you heard it right, this is finally a real-time result. This all happens through decoupling the direct and indirect effect of light sources and denoising them separately. In the meantime, the algorithm also tries to estimate the amount of noise in different parts of the image to provide more useful information for the denoising routines. The fact that the entire pipeline runs on the graphics card is a great testament to the simplicity of this algorithm. Here you see the term SVGF, you see the results of the new technique. So we have these noisy input images with one SPP and look at that. Wow! This is one of those papers that looks like magic, and know your own networks or learning algorithms have been used in this work. Not so long ago, I speculated or more accurately hoped that real-time photorealistic rendering would be a possibility during my lifetime. In just a few years later, this paper appears. We know that the rate of progress in computer graphics research is just staggering, but this is too much to handle. Super excited to see where the artists will take this, and of course, I'll be here to show you the coolest follow-up works. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is 2 Minute Papers with Karoijona Efehir. It is time for some minds to be blown. Dota 2 is a multiplayer online battle arena game with a huge cult following and world championship events with a prize pool of over 20 million dollars. In this game, players form two teams and control a hero each and use their strategy and special abilities to defeat the other team. Even AI recently created an AI for this game that is so good that they challenged the best players in the world. Now note that this program is not playing the full feature set of the game, but a version that is limited to one versus one encounters with several other elements of the game disabled. Since lots of strategies involved, we always discuss in these episodes that long term planning is the Achilles heel of these learning algorithms. A small blender in the early game can often snowball out of control by the end of the match and it is hard for the AI and sometimes to even humans to identify these cases. And this game is a huge challenge because unlike chess and go, it has lots of incomplete information and even the simplified one versus one mode involves a reasonable amount of long-term planning. It also involves attacks, trickery and deceiving an opponent and can be imagined as a strategy game that also requires significant technical prowess to pull off the most spectacular moves. This game is also designed in a way that new and unfamiliar situations come up all the time, which require lots of experience and split second decision making to master. This is a true test for any kind of AI. And note that this AI wasn't told anything about the game, not even the rules and was just instructed to try to find a way to win. The algorithm was trained in 24 hours and during this time it not only learned the rules and objectives of the game but it also pulls off remarkable tactics. For instance, other players were very surprised that the bot didn't take the bait, which typically means a smart tactic involving giving up a smaller battle in favor of winning a bigger objective. The AI has a ton of experience playing the game and typically sees through these shenanigans. In this game there are also neutral units that we call creep. When killed they grant precious gold and experience to our opponent so we typically try to deny that. If these units encounter an obstacle they go around it. So players develop the technique by the name Creep Blocking, which is the art of holding them up by the hero character to minimize the distance travelled by them in a unit of time. And the AI has not only learned about the existence of this technique by itself but it also executes it with stunning precision which is quite remarkable. And again during the training phase it had never seen any human play the game and do something like this. The other remarkable thing is that when a player disappears in the darkness, the AI predicts what he could be doing, plans around it and strikes where the player is expected to show up. If you remember, DeepMind's initial go algorithm contained a bootstrapping step where it was fed a large amount of games by players to grasp the basics. The truly remarkable thing is that none of that happened here. The algorithm was trained for only 24 hours and it only played against itself. When it finally played against Dendi, the reigning world champion, the first match was such a treat and I was shocked to see that the AI has outplayed him. In the second game, the player tried to create a situation that he thought the AI hasn't encountered before by giving up some creep to it. The program ruthlessly took advantage of this mistake and defeated him almost immediately. Open AI is bought not only one but apparently also broke the will of Dendi who tapped out after two matches. I feel like someone being hit by a sledgehammer. I didn't even know this was being worked on. This is such a remarkable achievement. Usually the first argument I hear is that of course the AI can play non-stop without bathroom breaks or sleep. While admittedly this is also true for some players, the algorithm was only trained for 24 hours. Note that the steel means a stupendous amount of games played but in terms of training time. That these algorithms typically take from weeks to months to train properly 24 hours is nothing. The second argument that I often hear is that the AI should of course win every time because it has close to zero reaction time and can perform thousands of actions every second. For instance if we would play a game where the goal is to perform the most amount of actions per minute, clearly humans with biological limitations would stand no chance against the computer program. However in this case the number of actions that this algorithm performs in a minute is comparable to that of a human player. This means that these results stem from superior technical abilities and planning and not from the fact that we are talking about the computer. We can look at this result from two different directions. One could be saying well no big deal because this is only a highly limited and hamstrung version of the game which is way less complex than a fully fleshed 5 vs 5 team match. Or two we could say that the algorithm had shown a remarkable aptitude for learning highly sophisticated technical maneuvers and longer term strategy in a difficult game and the rest is only a matter of time. In fact in 5 vs 5 there is even more room for a highly intelligent program to shine and create new tactics that we have never thought of. I would bet that if anything we are going to be even more surprised by the 5 vs 5 results later. We are still lacking in details a bit but I have contacted the open AI guys who noted that there will be more information available in the next few days. Whenever something new appears I'll be here to cover it for you fellow scholars. If you are new to the series and enjoyed this episode make sure to subscribe and click the bell icon for two super fun science videos a week. And if you find yourself interested in Dota 2 and admittedly it's hard not to and would like to catch up a bit on the basics. Make sure to visit Day 9's channel who has a really nice playlist about the fundamentals of the game. There's a link in the description for it, check it out. If you go to his channel make sure to leave him a kind scholarly comment. Let the world see how courteous the two minute papers listeners are. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karojona Ifeher. This paper does not contain the usual fireworks that you are used to in two-minute papers, but I feel that this is a very important story that needs to be told to everyone. In computer science, we encounter many interesting problems, like finding the shortest path between two given streets in a city or measuring the stability of a bridge. Up until a few years ago, these were almost exclusively solved by traditional handcrafted techniques. This means a class of techniques that were designed by hand by scientists and are often specific to the problem we have at hand. Different problem, different algorithm. And fast forward to a few years ago, we witnessed an amazing resurgence of neural networks and learning algorithms. Many problems that were previously thought to be unsolvable crumbled quickly one after another. Now it is clear that the age of AI is coming, and clearly there are possible applications of it that we need to be very cautious with. Since we design these traditional techniques by hand, the failure cases are often known because these algorithms are simple enough that we can look under the hood and make reasonable assumptions. This is not the case with deep neural networks. We know that in some cases neural networks are unreliable, but it is remarkably hard to identify these failure cases. For instance, earlier we talked about this technique by the name PIX2PIX, where we could make a crude drawing of a cat and it would translate it to a real image. It works spectacularly in many cases, but Twitter was also full of examples with really amusing failure cases. Beyond the unreliability, we have a much bigger problem, and that problem is adversarial examples. In an earlier episode, we discussed an adversarial algorithm, wherein in an amusing example, they added a tiny bit of barely perceptible noise to this image to make the deep neural network misidentify a bus for an ostrich. We can even train a new neural network that is specifically tailored to break the one we have, opening up the possibility of targeted attacks against it. To alleviate this problem, it is always a good idea to make sure that these neural networks are also trained on adversarial inputs as well. But how do we know how many possible other adversarial examples exist that we haven't found yet? The paper discusses a way of verifying important properties of neural networks. For instance, it can measure the adversarial robustness of such a network, and this is super useful, because it gives us information whether there are possible forged inputs that could break our learning systems. The paper also contains a nice little experiment with airborne collision avoidance systems. The goal here is avoiding mid-air collisions between commercial aircrafts while minimizing the number of alerts. As a small-scale thought experiment, we can train a neural network to replace an existing system, but in this case, such a neural network would have to be verified, and it is now finally a possibility. Now, make no mistake, this does not mean that there are any sort of aircraft safety systems deployed in the industry that are relying on neural networks. No, no, absolutely not. This is a small scale, what if kind of experiment that may prove to be a first step towards something really exciting? This is one of those incredible papers that even without the usual visual fireworks makes me feel that I am a part of the future. This is a step towards the future where we can prove that a learning algorithm is guaranteed to work in mission-critical systems. I would also like to note that even if this episode is not meant to go viral on the internet, it is still an important story to be told. Normally, creating videos like this would be a financial suicide, but we are not hurt by this at all because we get stable support from you on Patreon. And that's what it is all about. Worrying less about views and spending more time talking about what's really important. Absolutely amazing. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Kato Ijona Ifeher. A bit more than two years ago, the Deep Mind Guys implemented an algorithm that could play Atari Breakout on a superhuman level by looking at the video feed that you see here. And the news immediately took the world by storm. This original paper is a bit more than two years old and has already been referenced in well over a thousand other research papers. That is one powerful paper. This algorithm was based on a combination of a neural network and reinforcement learning. The neural network was used to understand the video feed and reinforcement learning is there to come up with the appropriate actions. This is the part that plays the game. Reinforcement learning is very suitable for tasks where we are in a changing environment and we need to choose an appropriate action based on our surroundings to maximize some sort of score. This score can be, for instance, how far we've gotten in a labyrinth or how many collisions we have avoided with the helicopter or any sort of score that reflects how well we are currently doing. And this algorithm works similarly to how an animal learns new things. It observes the environment, tries different actions and sees if they worked well. If yes, it will keep doing that. If not, well, let's try something else. Pavlov's dog with the bell is an excellent example of that. There are many existing works in this area and it performs remarkably well for a number of problems and computer games, but only if the reward comes relatively quickly after the action. For instance, in breakout, if we miss the ball, we lose a life immediately, but if we hit it, we'll almost immediately break some breaks and increase our score. This is more than suitable for a well-built reinforcement learner algorithm. However, this earlier work didn't perform well on any games that required long-term planning. If Pavlov gave his dog a treat for something that it did two days ago, the animal would have no clue as to which action led to this tasty reward. And this work subject is a game where we control this green character and our goal is to push the boxes onto the red dots. This game is particularly difficult, not only for algorithms, but even humans because of two important reasons. One, it requires long-term planning, which, as we know, is a huge issue for reinforcement learning algorithms. Just because a box is next to a dot doesn't mean that it is the one that belongs there. This is a particularly nasty property of the game. And two, some mistakes we make are irreversible. For instance, pushing a box in a corner can make it impossible to complete the level. If we have an algorithm that tries a bunch of actions and sees if they stick, well, that's not going to work here. It is now hopefully easy to see that this is an obscenely difficult problem. And the deep-mind guys just came up with imagination-augmented agents as a solution for it. So what is behind this really cool name? The interesting part about this novel architecture is that it uses imagination, which is a routine to cook up not only one action, but entire plans consisting of several steps, and finally, choose one that has the greatest expected reward over the long term. It takes information about the present and imagines possible futures and chooses the one with the most handsome reward. And as you can see, this is only the first paper on this new architecture and it can already solve a problem with seven boxes. This is just unreal, absolutely amazing work. And please note that this is a fairly general algorithm that can be used for a number of different problems. This particular game was just one way of demonstrating the attractive properties of this new technique. The paper contains more results and is a great read. Make sure to have a look. Also, if you've enjoyed this episode, please consider supporting two-minute papers on Patreon. Details are available in the video description. Have a look. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Kato Ejona Efehir. Style Transfer is an amazing area in machine learning and AI research where we take two images. Image number one is an input photograph and Image number two is the desired style. And the output of this process is the content of Image number one with the style of Image number two. This first paper opened up an incredible new area of research. As a result, a ton of different variants have emerged in the last two years. Feet forward style transfer for close to real-time results, temporal leak adherence style transfer for videos and much much more. And this one not only outperforms previously existing techniques but also broadens the horizon of possible style transfer applications. And obviously, a human would be best at doing this because a human has an understanding of the objects seen in these images. And now, hold on to your papers because the main objective of this method is to create semantically meaningful results for style transfer. It is meant to do well with input image pairs that may look completely different visually but have some semantic components that are similar. For instance, a photograph of a human face and the drawing of a virtual character is an excellent example of that. In this case, this learning algorithm recognizes that they both have noses and uses this valuable information in the style transfer process. As a result, it has three super cool applications. First, the regular photo to style transfer that we all know and love. Second, it is also capable of swapping the style of two input images. Third, and hold on to your papers because this is going to be even more insane. Style or sketch to photo. And we have a plus one here as well, so fourth, it also supports color transfer between photographs, which will allow creating amazing time lapse videos. I always try to lure you fellow scholars into looking at these papers, so make sure to have a look at the paper for some more results on this. And you can see here that this method was compared to several other techniques. For instance, you can see the cycle consistency paper and patch match. And this is one of those moments when I get super happy because more than 170 episodes into the series, we can not only appreciate the quality of these new results, but we also had previous episodes about both of these algorithms. As always, the links are available in the video description. Make sure to have a look, it's going to be a lot of fun. The source code of this project is also available. We also have a ton of episodes on computer graphics in the series. Make sure to have a look at those as well. Every now and then, I get emails from viewers who say that they came for the AI videos. In just in case, watched a recent episode on computer graphics and were completely hooked. Give it a try. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karojona Ifei. This is a piece of elastic cloth modeled from more than a million tiny triangles and its interaction with seven million colored grains of sand. This is super challenging because of two things. One, we have to compute the elastic deformations when these millions of tiny elements collide and two, all this while maintaining two-way coupling. This means that the cloth has an effect on the sand, but the effect of the sand is also simulated on the cloth. In elastic deformations, there are potential interactions between distant parts of the same material and self-collisions may also occur. Previous state-of-the-art techniques were either lacking in these self-collision effects, or the ones that were able to process that also included the fracturing of the material. With this novel work, it is possible to simulate both elastic deformations as you can see here, but it also supports simulating plasticity as you can see here with the cloth pieces sliding off of each other. Beautiful. This new technique also supports simulating a variety of different types of materials, knitted cloth ponchos, shag carpets, twisting cloth, hair, tearing fiber, and more. And it does all this with a typical execution time between 10 to 90 seconds per frame. In these black screens, you see the timing information and the number of particles and triangles used in these simulations. And you will see that there are many scenes where millions of triangles and particles are processed in very little time. It is very rare that we can implement only one technique that takes care of so many kinds of interactions while still obtaining results very quickly. This is insanity. This paper is absolutely top-tier bank for the buck, and I am really excited to see some more elastic plastic simulations in all kinds of digital media in the future. You know our motto, a couple more papers down the line and having something like this in real-time applications may become a reality. Really cool. If you enjoyed this episode and you feel that 8 of these videos a month is worth a dollar, please consider supporting us on Patreon. One dollar per month really doesn't break the bank, but it is a great deal of help for us in keeping the series going. And your support has always been absolutely amazing, and I am so grateful to have so many devoted fellow scholars like you in our ranks. Details are in the video description. Have a look. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karojona Ifeher. In this series, we talk a lot about photorealistic rendering and making sure that the appearance of our virtual objects is simulated properly. A lot of works on how things look. However, in order to create a more complete sensorial experience, we also have to simulate how these things sound. And today, we are going to have a look at a really cool piece of work that simulates the sound of virtual elastic rods made of aluminum, steel, oak tree and rubber. And of course, before you ask, this also means that there will be sound simulations of everyone's favorite toy, the Walking Slinky. As for all papers that have anything to do with sound synthesis, I recommend using a pair of headphones for this episode. The sound emerging from these elastic rods is particularly difficult to simulate because of the fact that sound frequencies vary quite a bit over time and the objects themselves are also in motion and subject to deformations during the simulation. And as you will see with the Slinky, we potentially have tens of thousands of contact events in the meantime. Let's have a look at some results. For the fellow scholars who are worried about the validity of these Star Wars sounds, I know you're out there, make sure to watch the video until the end. The authors of the paper proposed a dipole model to create these simulations. Dipoles are typically used to approximate electric and magnetic fields in physics and in this case it is really amazing to see an application of it for sound synthesis. For instance, in most cases, these sound waves are typically symmetric around 2D cross sections of these objects which can be described by a dipole model quite well. Also, it is computationally quite effective and can eliminate these lengthy pre-computation steps that are typically present in previous techniques. There are also comparisons against the state of the art and we can hear how much richer the sound of this new technique is. And as you know all too well, I love all papers that have something to do with the real world around us. And the reason for this is that we can try the very best kind of validation for these algorithms and this is when we let reality be our judge. Some frequency plots are also available to validate the output of the algorithm against the real world sound samples from the lab. It is really amazing to see that we can use science to breathe more life in our virtual worlds. Watch in and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karojona Ifeher. In the film industry, we can often see footage of a human walking on the moon, fighting underwater or appearing in any environment without actually going there. To do this, a piece of footage of the actor is recorded in front of a green screen and then the background of the scene is changed to something else. This process is called green screen keying and in theory, this sounds simple enough that make no mistake, this is a challenging problem. Here's why. Issue number one is that separating the foreground from the background is non-trivial and is not a fully automatic process. Let's call this semi-automatic because the compositing artist starts drawing these separation masks and even though there is some help from pre-existing software, it still takes quite a bit of manual labor. For instance, in this example, it is extremely difficult to create a perfect separation between the background and the hair of the actor. Our eyes are extremely keen on catching such details, so even the slightest inaccuracies are going to appear as glaring mistakes. This takes a ton of time and effort from the side of the artist and we haven't even talked about tracking the changes between frames as we are talking about video animations. I think it is now easy to see that this is a hugely relevant problem in the post-production of feature films. And now onto issue number two, which is subtracting indirect illumination from this footage. This is a beautiful light transport effect where the color of different diffuse objects bleed onto each other. In this case, the green color of the background bleeds onto the karate uniform. That is normally a beautiful effect, but here it is highly undesirable because if we put this character in a different environment, it won't look like it belongs there. It will look more like one of those super fake Photoshop disasters that we see everywhere on the internet. And this technique offers a novel solution to this key-ing problem. First, we are asked to scribble on the screen and mark the most dominant colors of the scene. This we only have to do once, even though we are processing an entire video. As a result, we get an initial map where we can easily fix some of the issues. This is very easy and intuitive, not like those long sessions spent with pixel by pixel editing. These colors are then propagated to the entirety of the animation. The final results are compared to a ton of already existing methods on the market, and this one smokes them all. However, what is even more surprising is that it is also way better than what an independent artist produced which took 10 times that long. Other comparisons are also made for removing indirect illumination, which is also referred to as color and mixing in the paper. It is also shown that the algorithm is not too sensitive to this choice of dominant colors, so there is room for amazing follow-up papers to make the process a bit more automatic. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two minute papers with Karojona Ifeher. Whenever we take an image with our camera and look at it after an event, we often feel that many of them are close to perfect. If only it was less blurry or the focus distance was a bit further away. But the magic moment is now gone, and there's nothing to do other than cursing at the blurry footage that we are left with when showing it to our friends. However, if we have access to light fields, we can change some camera parameters after the photo was taken. This includes changing the focal distance or even slightly adjusting the viewpoint of the camera. How cool is that? This can be accomplished by a light field camera, which is also referred to as a planoptic camera. This tries to record not only light intensities, but the direction of incoming light as well. Earlier, this was typically achieved by using an array of cameras that's both expensive and cumbersome. And here comes the problem with using only one light field camera. Because of the increased amount of data that they have to record, current light field cameras are only able to take three frames per second. That's hardly satisfying if we wish to do this sort of post editing for videos. This work offers a novel technique to remedy this situation by attaching a standard camera to this light field camera. The goal is that the standard camera has 30, so tons of frames per second, but with little additional information and a light field camera, which has only a few frames per second, but with a ton of additional information. If we stitch all this information together in a smart way, maybe it is a possibility to get full light field editing for videos. Earlier, we have talked about interpolation techniques that can fill some of the missing frames in videos. And this way, we can fill in maybe every other frame in a footage, or we can be a bit more generous than that. However, if we are shown three frames per second and we have to create a smooth video by filling the blanks, would almost be like asking an algorithm to create a movie from a comic book. This would be awesome, but we're not there yet. Too much information is missing. This teaching process works with a bit more information than this, and the key idea is to use two convolutional neural networks to fill in the blanks. One is used to predict flows, which describe the movements and rotations of the objects in the scene, and one to predict the final appearance of the objects. Basically, one for how they move and one for how they look. And the results are just absolutely incredible. It is also blazing fast and takes less than a tenth of a second to create one of these new views. Here, you can see how the final program is able to change the focal distance of any of the frames in our video, or we can even click on something in the image to get it in focus. And all this is done after the video has been taken. The source code of this project is also available. With some more improvements, this could be tremendously useful in the film industry, because the directors could adjust their scenes after the shooting and not just sigh over the inaccuracies and missed opportunities. And this is just one of the many possible other applications. Absolutely amazing. If you have enjoyed this episode, don't forget to subscribe to Two Minute Papers and also make sure to click the bell icon to never miss an episode. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karojiwana Ifehir. This work is about transferring our gestures onto a virtual human's face in a way that is physically correct. This means that not only the changes in the facial geometry are transferred to a digital character. No, no, no. Here is how it works. This piece of work uses a really cool digital representation of our face that contains not only geometry, but there's also information about the bone and flesh and muscle structures as well. This means that it builds on a physically accurate model which synthesizes animations, where this human face is actuated by the appropriate muscles. We start out with a surface scan of the user, which through a registration step is then converted to a set of expressions that we wish to achieve. The inverse physics module tries to guess exactly which muscles are used and how they are used to achieve these target expressions. The animation step takes information of how the desired target expressions evolve in time and some physics information, such as gravity or wind, and the forward physics unit computes the final simulation of the digital character. So while we are talking about the effects of gravity and wind, here you can see how this can create more convincing outputs because these characters really become a part of their digital environment. As a result, the body mass index of a character can also be changed in both directions, slimming or fattening the face. Lip enhancement is also a possibility. If we had super high resolution facial scans, maybe a follow-up work could simulate the effects of Botox injections. How could would that be? Also, one of my favorite features of this technique is that it also enables artistic editing. By means of drawing, we can also specify a map of stiffness and mass distributions, and if we feel cruel enough, we can create a barely functioning human face to model and animate virtual zombies. Imagine what artists could do with this, especially in the presence of super high resolution textures and photorealistic rendering. Oh my! Another glimpse of the future of computer graphics and animation. Make sure to have a look at the paper for more applications. For instance, they also demonstrate the possibility of modifying the chin and the jawbone. They even have some result in simulating the effect of Bell's palsy, which is the paralysis of facial muscles on one side. While we are at this high note of illnesses, if you enjoy this episode and would like to support us, you can pick up really cool perks like early access for all of these episodes on Patreon. The link is available in the video description. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Kato Ejolna Ifehir. In earlier episodes, we've seen plenty of video footage about hair simulations and rendering. And today, we are going to look at a cool new technique that produces self-shadowing effects for hair and fur. In this image pair, you can see this drastic difference that shows how prominent this effect is in the visual appearance of hair. Just look at that. Beautiful. But computing such a thing is extremely costly. Since we have a dense piece of geometry, for instance hundreds of thousands of hair strands, we have to know how each one occludes the other ones. This would take hopelessly long to compute. To even get a program that executes in a reasonable amount of time, we clearly need to simplify the problem further. An earlier technique takes a few planes that cut the hair volume into layers. These planes are typically regularly spaced outward from the light sources, and it is much easier to work with a handful of these volume segments than with the full geometry. The more planes we use, the more layers we obtain, and the higher quality results we can expect. However, even if we can do this in real time, we will produce unrealistic images when using around 16 layers. Well, of course, we should then crank up the number of layers some more. If we do that, for instance by now, using 128 layers, we can expect better quality results, but we'll be able to process an image only twice a second, which is far from competitive. And even then, the final results still contain layering artifacts, and are not very close to the ground truth. There has to be a better way to do this. And with this new technique called Deep Opacity Maps, these layers are chosen more wisely, and this way we can achieve higher quality results with using only three layers, and it runs easily in real time. It is also more memory efficient than previous techniques. The key idea is that if we look at the hair from the light sources point of view, we can record how far away different parts of the geometry are from the light source. Then we can create the new layers further and further away according to this shape. This way, the layers are not plain or anymore. They adapt to the scene that we have at hand, and contain significantly more useful occlusion information. As you can see, this new technique blows all previous methods away and is incredibly simple. I have found an implementation from Philip Rideout. The link to this is available in the video description. If you have found more, let me know and I'll include your findings in the video description for the fellow tinkerers out there. The paper is ample in comparisons. Make sure to have a look at that too. And sometimes I get some messages saying, Karoi, why do you bother covering papers from so many years ago? It doesn't make any sense. And here, you can see that part of the excitement of two minute papers is that the next episode can be about absolutely anything. The series has been mostly focusing on computer graphics and machine learning papers, but don't forget that we also have an episode on whether we are living in a simulation or the stunning Kruger effect and so much more. I've put a link to both of them in the video description for your enjoyment. The other reason for covering older papers is that a lot of people don't know about them, and if we can help, just a tiny bit to make sure these incredible works see more widespread adoption with Donar job well. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Károly Zsolnai-Fehér. Everyone who watches this series knows that among many other scientific topics I am severely addicted to fluid simulations. And today it's time to relapse. And this time we are going to run wind tunnel tests on hummingbirds. Typically when a new engine, airplane, or even a new phone is being designed, we are interested in knowing how the heat flow and dissipation will look like, preferably before we are designing an object. To do so, we often run some virtual wind tunnel tests and optimize our design until we are happy with the results. Then we can proceed to build these new contractions. Simulating the pressure distribution and the aerodynamic forces is a large topic. However, visualizing these results is at least as well studied and difficult as writing a simulator. What is it exactly that we are interested in? Even if we have an intuitive particle-based simulation, millions and millions of particles, it is clearly impossible to show the path for every one of them. Grid-based simulations are often even more challenging to visualize well. So how do we choose what to visualize and what not to show on the screen? And in this paper, we can witness a new way of visualizing velocity and vorticity fields. And this visualization happens through clapsh maps. This is a mathematical transformation where we create a sphere and a set of points on this sphere correspond to vortex lines and their evolution over time. However, if instead of only points, we pick an entire region on this sphere, as you can see the north and south pole regions here, we obtain vortex tubes. These vortex tubes provide an accurate representation of the vorticity information within the simulation, and this is one of the rare cases where the validity of such a solution can also be shown. Such a crazy idea, loving it. And with this, we can get a better understanding of the airflow around the wings of the hummingbird, but we can also learn more from pre-existing NASA aircraft data sets. Have a look at these incredible results. Publishing a paper at the CIGARF conference is an incredible feat that typically takes a few brilliant guys and several years of unbelievably hard work. Well, apparently this is not such a challenge for Albert Churn, who was also the first author of this and the Schrodinger smoke paper just a year ago that we reported on. He's doing incredible work at taking a piece of mathematical theory and showing remarkable applications of it in new areas where we would think it doesn't belong at all. The link is available in the video description, both for this and the previous works, make sure to have a look. There's lots of beautifully written mathematics to be read there that seems to be from another world. It's a truly unique experience. The paper reports that the source code is also available, but I was unable to find it yet. If you have found a public implementation, please let me know and I'll update the video description with your link. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Karoizona Ifeher. Today, we are going to talk about a new endeavor to teach some more common sense to learning algorithms. If you remember, in an earlier episode, we talked about an excellent work by Andrei Carpethy, who built an algorithm that looked at an input image and described in a full, well-formed sentence what is depicted there. By the way, he recently became director of AI at Tesla. Before that, he worked at OpenAI freshly after graduating with a PhD. Now that is a scholarly career, if I've ever seen one. Reading about this earlier work was one of those moments when I really had to hold onto my papers, not to fall out of the chair, but of course, as it should be with every new breakthrough. The failure cases were thoroughly discussed. One of the motivations for this new work is that we could improve the results by creating a video database that contains a ton of commonly occurring events that would be useful to learn. These events include moving and picking up or holding, poking, throwing, pouring, or plugging in different things, and much more. The goal is that these neural algorithms would get tons of training data for these, and would be able to distinguish whether a human is showing them something, or just moving things about. The already existing video databases are surprisingly sparse in this sort of information, and in this new, freshly published dataset, we can learn on a hundred, thousand labeled videos to accelerate research in this direction. I love how many of these works are intertwined, and how follow-up research works try to address the weaknesses of previous techniques. Some initial results with learning on this dataset are also reported to kick things off, and they seem quite good if you look at the results here, but since this was not the focus of the paper, we shouldn't expect superhuman performance. However, as almost all papers in research are stepping stones, two more follow-up papers down the line, this will be an entirely different discussion. I'd love to report back to you on the progress later. Super excited for that. Thanks for watching, and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karoizona Ifehir. This paper is from the Google Deep Mind Guys and is about teaching neural networks to be capable of relational reasoning. This means that we can present the algorithm with an image and ask it relatively complex relational questions. For instance, if we show it this image and ask, what is the color of the object that is closest to the blue object, it would answer red. This is a particularly difficult problem because all the algorithm has access to is a bunch of pixels. In computer code, it is near impossible to mathematically express that in an image something is below or next to something else, especially in three-dimensional scenes. Beyond the list of colors, this requires a cognitive understanding of the entirety of the image. This is something that we humans are amazingly good at, but computer algorithms are dreadful for this type of work, and this work almost feels like teaching common sense to a learning algorithm. This is accomplished by augmenting an already existing neural network with a relational network module. This is implemented on top of a recurrent neural network that we call long short-term memory or LSTM that is able to process sequences of information, for instance, an input sentence. The more seasoned fellow scholars know that we have talked about LSTM's in earlier episodes, and of course, as always, the video description contains these episodes for your enjoyment. Make sure to have a look, you'll love it. As you can see in this result, this relational reasoning also works for three-dimensional scenes as well. The aggregated results in the paper show that this method is not only leaps and bounds beyond the capabilities of already existing algorithms, but, and now, hold on to your papers. In many cases, it also shows superhuman performance. I love seeing these charts in machine learning papers, where several learning algorithms and humans are benchmarked on the same tasks. This paper was barely published, and there is already a first unofficial public implementation, and two research papers have already referenced it. This is such a great testament to the incredible pace of machine learning research these days, to say that it is competitive would be a huge understatement. Achieving high-quality results in relational reasoning is an important cornerstone for achieving general intelligence, and even though there is still much, much more to do, today is one of those days when we can feel that we are a part of the future. The failure cases are also reported in the paper and are definitely worthy of your time and attention. When I asked for permissions to cover this paper in the series, all three scientists from DeepMind happily answered yes within 30 minutes. That's unbelievable. Thanks guys! Also, some of these questions sound like ones that we would get in the easier part of an IQ test. I wouldn't be very surprised to see a learning algorithm complete a full IQ test with flying colors in the near future. If you enjoyed this episode and you feel that eight of these videos a month is worth a dollar, please consider supporting us on Patreon. This way, we can make better videos for your enjoyment. We have recently reached a new milestone, which means that part of these funds will be used to empower research projects. Details are available in the video description. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Car Oizro Mnefe Hair. Dear Fellow Scholars, this is Two Minute Papers with Car Oizro Mnefe Hair. Close enough. As you have probably noticed, today we are going to talk about text to speech or TTS in short. TTS means that we write a piece of text and a computer synthesized voice will read it aloud for us. This is really useful for reading the news or creating audiobooks that don't have any official voice overs. This work was done by researchers at Princeton University and Adobe and is about text-based audio narration editing. This one is going to be crazy good. The Adobe guys like to call this the photoshop of voice overs. In a normal situation, we have access to a waveform and if we wish to change anything in a voice over, we need to edit it. Editing waveforms by hand is extremely difficult. Traditional techniques often can't even reliably find the boundaries between words and letters let alone edit them. And with this technique, we can cut, copy and even edit this text and the waveforms will automatically be transformed appropriately using the same voice. Had it struck squarely, it would have killed him. Had it struck squarely, it would have saved him. We can even use new words that have never been uttered in the original narration. We leave the eventuality to time and law. We leave the eventuality to time and believe. It solves an optimization problem where the similarity, smoothness and the pace of the original footage is to be matched as closely as possible. One of the excellent new features is that we can even choose from several different voicings for the new word and insert the one that we deemed the most appropriate. For expert users, the pitch and duration is also editable. It's always important to have a look at a new technique and make sure that it works well in practice. But in science, this is only the first step. There has to be more proof that a new proposed method works well in a variety of cases. In this case, a theoretical proof by means of mathematics is not feasible, therefore a user study was carried out where listeners were shown, synthesized and real audio samples and had to blindly decide which was which. The algorithm was remarkably successful at deceiving the test subjects. Make sure to have a look at the paper in the description for more details. This technique is traditional in a sense that it doesn't use any sort of neural networks. However, there are great strides being made in that area as well, which I am quite excited to show you in future episodes. And due to some of these newer video and audio editing techniques, I expect that within the internet forums, fake news is going to be an enduring topic. I hope that in parallel with better and better text and video synthesis, there will be an arms race with other methods that are designed to identify these cases. A neural detective, if you will. And now, if you excuse me, I'll give this publicly available TTS one more try and see if I can retire from narrating videos. Thanks for watching and for your generous support, and I'll see you next time. Yup, exact same thing. Bet you didn't even notice it.

Dear Fellow Scholars, this is two-minute papers with Kato Ejolenei Fehr. This paper is about creating stunning cloth simulations that are rich in yarn-to-yarn contact. Normally, this is a challenging problem because finding and simulating all the possible contacts between tens of thousands of interlinked pieces of geometry is a prohibitively long process. Also, due to the many different kinds of possible loop configurations, these contacts can take an awful lot of different shapes which all need to be taken into consideration. Since we are so used to see these garments moving about in real life, if someone writes a simulator that is off just by a tiny bit will immediately spot the difference. I think it is now easy to see why this is a highly challenging problem. This technique optimizes this process by only computing some of the forces that emerge from these yarns pulling each other and only trying to approximate the rest. The good news is that this approximation is carried out with temporal coherence. This means that these contact models are retained through time and are only rebuilt when it is absolutely necessary. The regions marked with red in these simulations show the domains that are found to be undergoing significant deformation, therefore we need to focus most of our efforts in rebuilding the simulation model for these regions. Look at these results. This is unbelievable. There is so much detail in these simulations and all this was done seven years ago. In research and technology this is an eternity. This just blows my mind. The results are also compared against the expensive reference technique as well. And you can see that the differences are miniscule, but the new improved technique offers a 4-5 time speedup over that. For my research project I also run many of these simulations myself and many of these tasks take several all-nighters to compute. If someone would say that each of my all-nighters would now count as 5, I'd be absolutely delighted. If you haven't subscribed to the series, please make sure to do so and please also click the bell icon to never miss an episode. We have tons of awesome papers to come in the next few episodes. Looking forward to seeing you there. Thanks for watching and for your generous support. Bye.

Dear Fellow Scholars, this is two-minute papers with Kato Ejona Ifehir. From many fond childhood memories, most of us are quite fond of the colorful physical appearance of bubbles and fuel-water mixtures. We also know surprisingly little about this peculiar phenomenon, where the color of an object changes when we either turn our head or change the lighting. This happens in a very colorful manner and physicists like to call this iridescence organiocromism, what is even less known is that if we try to use a light simulation program to make an image with leather, we'll be surprised to see that it also shows a pronounced goniocromatic effect. An even more less known fact is that quite a few birds, insects, minerals, seashells, and even some fruits are iridescent as well. I've added links to some really cool additional readings to the video description for your enjoyment. This effect is caused by materials that scatter different colors of light in different directions. A white incoming light is therefore scattered not in one direction, but in a number of different directions sorted by their colors. This is why we get these beautiful rainbow colored patterns that we all love so much. Now that we know what iridescence is, the next step is obviously to infuse our light simulation programs to have this awesome feature. This paper is about simulating this effect with micro facets which are tiny microstructures on the surface of rough objects. And with this, it is now suddenly possible to put a thin iridescent film onto a virtual object and create a photorealistic image out of it. If you're into math and would like to read about some tasty spectral integration in the frequency spaced with Fourier transforms, this paper is for you. If you're not a mathematician, also make sure to have a look because the production quality of this paper is through the roof. The methodology, derivations, comparisons are all really crisp. Loving it. If you have a look, you will get the glimpse of what it takes to create a work of this quality. This is one of the best papers in photorealistic rendering I've seen in a while. In the meantime, I'm getting more and more messages from you fellow scholars who tell their stories on how they chose to turn their lives around and started studying science because of this series. Wow, that's incredibly humbling and I really don't know how to express my joy for this. I always say that it's so great to be a part of the future and I am delighted to see that some of you want to be a part of the future and not only as an observer but as a research scientist. This sort of impact is stronger than the absolute best case scenario I have ever dreamed of for the series. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Kato Ejolene Fahir. This paper is about the absolute favorite thing of computer graphics researchers, destroying virtual objects in the most creative ways. This is the only place on earth where words like the formable bodies and cutting can be used in the same sentence and be delighted about it. This time around we are going to cut and dismember every virtual object that stands in our way, and then some more. In these animations we have complex 3D geometry and the objective is to change these geometries in a way that remains physically correct even in the presence of complex cut surfaces. When such a cut happens, traditional techniques typically delete and duplicate parts of the geometry close to the cut. This is a heavily simplified solution that leads to inaccurate results. Other techniques try to rebuild parts of the geometry that are affected by the cut. This is what computer graphics researchers like to call remashing, and it works quite well, but it takes ages to perform. Also, it still has drawbacks, for instance quantities like temperature and deformations also have to be transferred to the new geometry, which is non-trivial to execute properly. In this work, a new technique is proposed that is able to process really complex cuts without creating new geometry. No remashing takes place, but the mass and stiffness properties of the materials are retained correctly. Also, the fact that it minimizes the geometric processing overhead leads to a not only simpler, but a more efficient solution. There is so much visual detail in the results that I could watch this video 10 times and still find something new in there. There are also some horrifying game of thrones kind of experiments in this footage. Watch out! Ouch! The presentation of the results and the part of the video that compares against the previous technique is absolutely brilliant, you have to see it. The paper is also remarkably well written, make sure to have a look at that too. The link is available in the video description. I am really itching to make some longer videos where we can go into some of these derivations and build a strong intuitive understanding of them. That sounds like a ton of fun, and if this could ever become a full-time endeavor, I am more than enthused to start doing more and work on bonus videos like that. If you enjoyed this episode, don't forget to hit the like button and subscribe to the series. Normally it's up to YouTube to decide whether you get a notification or not, so make sure to click the bell icon as well to never miss a two minute paper's episode. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Katojol Naifahir. This is one of those new, absolutely insane papers from the Google DeepMind guys. You're going to see a follow-up work to an algorithm that looks at a bunch of images, and from that, it automatically learns the concept of birds, human faces, or coral reefs. So much so that we are able to write a new sentence, and it will generate a new, close to photorealistic image from this written description. This network is capable of creating images that are significantly different than the ones it has been trained on. This already sounds like science fiction. Completely unreal. This work goes by the name Pixels CNN. We'll discuss a follow-up work to that in a moment. The downside of this method is that these images are generated pixel by pixel, and many of these pixels depend on their neighborhood. For instance, if I start to draw one pixel of the beak of a bird, the neighboring pixels have to adhere to this constraint, and have to be the continuation of the beak. Clearly, these images have a lot of structure. This means that we cannot do this process in parallel, but create these new images one pixel at a time. This is an extremely slow and computationally expensive process, and hence, the original paper showed results with 32 by 32 and 64 by 64 images at most. As we process everything sequentially, the execution time of the algorithm scales linearly with the number of pixels we can generate. It is like a factory where there are a ton of assembly lines, but only one person to run around and operate all of them. Here, the goal was to start generating different regions of these images independently, but only in cases when these pixels are not strongly correlated. For instance, doing this with neighbors is an ogre. This is possible, but extremely challenging, and the paper contains details on how to select these pixels, and when we can pretend them to be independent. And now, feast your eyes upon these spectacular results. If we are looking for a yellow bird with a black head, orange eyes, and an orange bill, we are going to see much more detailed images. The complexity of the new algorithm scales with the number of pixels not linearly, but in a logarithmic manner, which is basically the equivalent of winning the jackpot in terms of parallelization, and it often results in a more than 100 times speed up. This is a factory that's not run by one guy, but one that works properly. The lead author, Scott Reed, has also published some more amazing results on Twitter as well. In these examples, we can see the evolution of the final image that is generated by the network. It is an amazing feeling to be a part of the future. And note that there is a ton of challenges with the idea. This is one of those typical cases when the idea is only the first step, and execution is king. Make sure to have a look at the paper for more details. According to our regular schedule, we try our best to put out two videos every week. That's eight episodes a month. If you feel that eight of these episodes is worth a dollar for you, please consider supporting us on Patreon. This way, we can create more elaborate episodes for you. The channel is growing at a remarkable rate, and your support has been absolutely amazing. I am honored to have an audience like you fellow scholars. We are quite close to hitting our next milestone, and this milestone will be about giving back more to the scientific community and empowering other research projects. I've put a link to our Patreon page with the details in the video description. Thanks for watching, and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karoizhou and Aifahir. We've seen a lot of fluid and smoke simulations throughout the series. In each of these cases, the objective was to maximize the realism of these animations often to the point where they are indistinguishable from reality. However, there are cases where creating photorealistic footage is not the main objective. Artists often seek to imbue these fluid and smoke simulations with their own distinctive style and this style needs not to be photorealistic. It can be cartoonish, black and white, or take a variety of different color schemes. But unfortunately, to obtain such an effect, we have to sit down, get a bunch of papers, and draw the entirety of the animation frame by frame. And of course, to accomplish this, we also need to be physicists and know the underlying laws of fluid dynamics. That's not only borderline impossible, but extremely laborious as well. It would be really cool to have an algorithm that is somehow able to learn our art style and apply it to a fluid or smoke simulation sequence. But the question is, how do we exactly specify this style? Have a look at this really cool technique. I love the idea behind it. First, we compute a classical smoke simulation, then we freeze a few frames and get the artist to colorize them. After that, the algorithm tries to propagate this artistic style to the entirety of the sequence. Intuitively, this is artistic style transfer for fluid animations, but without using any machine learning techniques. Here, we are doing patch-based regenerative morphing. This awesome term refers to a technique that is trying to understand the direction of flows and that vac the colored regions according to it in a way that is both visually and temporarily coherent. Visually coherent means that it looks as close to plausible as we can make it, and temporarily coherent means that we are not looking only at one frame, but a sequence of frames, and the movement through these neighboring frames has to be smooth and consistent. These animation sequences were created from 8 to 9 colorized frames, and whatever you see happening in between was filled in by the algorithm. And again, we are talking about the artistic style here, not the simulation itself. A fine, handcrafted work in the world dominated by advanced learning algorithms. This paper is a bit like a beautiful handmade automatic timepiece in the era of quartz watches. If you enjoyed this episode, make sure to leave a like on the video, and don't forget to subscribe to get a glimpse of the future on the channel twice a week. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Kato Ejolnai-Fahir. Creating applications for mobile Android and iOS devices is a laborious endeavor, which most of the time includes creating a graphical user interface. These are the shiny, front-end interfaces that enable the user to interact with the back-end of our applications. So what about an algorithm that learns how to create these graphical user interfaces and automates part of this process? This piece of work takes one single input image that we can trivially obtain by making a screenshot of the user interface, and it almost immediately provides us with the code that is required to recreate it. What an amazing idea! The algorithm supports several different target platforms. For instance, it can give us code for iOS and Android devices. This code we can hand over to a compiler, which will create an executable application. This technique also supports HTML as well for creating websites with a desired user interface. Under the hood, a domain-specific language is being learned, and using this, it is possible to have a concise text representation of a user interface. Note that by no means the only use of domain-specific languages. The image of the graphical user interface is learned by a classical convolutional neural network, and this text representation is learned by a technique machine learning researchers like to call long short-term memory. LSTM in short. This is a neural network variant that is able to learn sequences of data and is typically used for language translation, music composition, or learning all the novels of Shakespeare, and writing new ones in his style. If you are wondering why these examples are suspiciously specific, we've had an earlier episode about this, I've put a link to it in the video description. Make sure to have a look, you're going to love it. Also, this year it will have its 20th year anniversary. Live long and prosper, little LSTM. Now I already see the forums go up in flames, sweeping generalizations, far-eaching statements on front-end developers around the world getting fired and all that. I start out by saying that I highly doubt that this work would mean the end of front-end development jobs in the industry. However, what I do think is that with a few improvements, it can quickly prove its worth by augmenting human labor and cutting down the costs of implementing graphical user interfaces in the future. This is another testament to the variety of tasks modern machine learning algorithms can take care of. The author also has a GitHub repository with a few more clarifications stating that the source code of the project and the dataset will be available soon. Tinkerers rejoice. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Karojola Ifehir. After around 160 episodes into Two Minute Papers, I think it is no secret to anyone that I am helplessly addicted to fluid simulations, so you can already guess what this episode is going to be about. I bet you will be as spellbound by this beautiful footage of wet sand simulations as I was when I first seen it. Before you ask, yes, I have attempted to prepare some slow motion action too. As you remember, simulating the motion of fluids involves solving equations that tell us how the velocity and the pressure evolves in time. Now the 3D world we live in is a continuum and we cannot solve these quantities everywhere because that would take an infinite amount of time. To alleviate this, we can put a grid in our virtual world and obtain these quantities only in these grid points. The higher the resolution the grid is, the more realistic the animations are, but the computation time also scales quite poorly. It is really not a surprise that we have barely seen any wet sand simulations in the visual effects industry so far. Here we have an efficient algorithm to handle these cases and as you will see, this is not only extremely expensive to compute, but nasty stability issues also arise. Have a look at this example here. These are sand simulations with different cohesion values. The means the strength of intermolecular forces that hold the material together. The higher cohesion is, the harder it is to break the sand up, the bigger the clumps are. This is an important quantity for our simulation because the higher the water saturation of this block of sand, the more cohesive it is. Now if we try to simulate this effect with traditional techniques on a coarse grid, will encounter a weird phenomenon. Namely, the longer our simulation runs, the larger the volume of the sand becomes. An excellent way to demonstrate this phenomenon is using these R-glasses, where you can clearly see that after only a good couple turns, the amount of sand within is significantly increased. This is particularly interesting because normally in classical fluid simulations, if our grid resolution is insufficient, we typically encounter water volume dissipation, which means that the total amount of mass in the simulation decreases over time. Here, we have the exact opposite, like in a magic trick, after every turn, the volume gets inflated. That's a really peculiar and no less challenging problem. This issue can be alleviated by using a finer grid, which is, as we know, extremely costly to compute, or the authors propose the volume fixing method to take care of this without significantly increasing the execution time of the algorithm. Make sure to have a look at the paper, which is certainly my kind of paper. Lots of beautiful physics and a study on how to solve these equations so that we can obtain an efficient wet sand simulator. And also, don't forget, a fluid paper a day keeps the obsessions away. In the meantime, a word about the two minute paper shirts. I am always delighted to see you fellow scholars sending over photos of yourselves, proudly posing with your newly obtained shirts for the series. Thanks so much and please keep them coming. They are available through two minute papers.com for the US and the EU and Worldwide link is also available in the video description. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Kato Ejolene Fahir. Today we are going to talk about a rigorous scientific topic, none other than the creation of the perfect selfie photo. By definition, selfies are made by us, which means that these are typically short-range photos, and due to the perspective distortion of the camera lens, we often experience unpleasant effects like the heavy magnification of the nose and the forehead. To get this, this technique enables us to take a photo, and after that, edit the perceived camera distance for it without changing anything else. Basically, algorithmic beautification. This technique works the following way. We analyze the photo and try to figure out how distant the camera was when the photo was taken. Then, we create a digital model of the perspective camera and create a 3D model of the face. This is a process that mathematicians like to call fitting. It means that if we know the optics of perspective cameras, we can work backwards from the input photo that we have and find an appropriate setup that would result in this photo. Then we will be able to adjust this distance to even out the camera lens distortions. But that's not all, because as we have the digital 3D model of the face, we can do even more. For instance, we can also rotate it around in multiple directions. To build such a 3D model, we typically try to locate several well recognizable hotspots on the face, such as the chin, eyebrows, nose stem, the region under the nose, eyes, and lips. However, as these hotspots lead to a poor 3D representation of the human face, the authors added a few more of these hotspots to the detection process. This still takes less than 5 seconds. Earlier, we also talked about a neural network based technique that judged our selfie photos by assigning a score to them. I would absolutely love to see how that work would react to a before and after photo that comes from this technique. This way, we can formulate this score as a maximization problem, and as a result, we could have an automated technique that truly creates the perfect selfie photo through these warping operations. The best kind of evaluation is when we let reality be our judge and use images that were taken closer or farther away and compare the output of this technique against them. These true images bear the ground truth label throughout this video. The differences are often barely perceptible and to provide a better localization of the error, some different images are shown in the paper. If you are into stereoscopy, there's also an entire section about that as well. The authors also uploaded an interactive version of their work online that anyone can try free of charge. So as always, your scholarly before and after selfie experiments are more than welcome in the comment section. Whether you're already subscribed to the series or just subscribing now, which you should absolutely do, make sure to click the bell icon to never miss an episode. We have lots of amazing works coming up in the next few videos. Hope to see you there again. Thanks for watching and for your generous support, and I'll see you next time.

Dear fellow scholars, this is two-minute papers with Karojona Ifehir. This episode is about simulating a beautiful phenomenon in nature, the buckling and coiling effect of honey. Hmmmm. This effect is due to the high viscosity of materials like honey, which means that they are highly resistant against deformation. Water, however, is much less viscous as it is held together by weaker intermolecular forces, therefore it is easier to deform, making it so easy to pour it into a glass. We had an earlier episode on honey buckling, and as every season fellow scholar already knows, the link is available in the video description. One key difference of this work is that the older solution was built upon a Lagrangian approach, which means that the simulation consists of computing the velocities and the pressure that acts on these particles. It is a particle-based simulation. Here, a solution is proposed for the Eulerian approach, which means that we do not compute these quantities everywhere in the continuum of space, but we use a fine 3D grid and we compute these quantities only in these grid points. No particles to be seen anywhere. There are mathematical techniques to try to guess what happens between these individual grid points and this process is referred to as interpolation. So normally, in this grid-based approach, if we wish to simulate such a buckling effect will be sorely disappointed, because what we will see is that the surface details rapidly disappear due to the inaccuracies in the simulation. The reason for this is that the classical grid-based simulators utilize a technique that mathematicians like to call operator splitting. This means that we solve these fluid equations by taking care of advection, pressure, and viscosity separately. These are the required quantities, separate solutions. This is great because it eases the computational complexity of the problem, however, we have to pay a price for it in the form of newly introduced inaccuracies. For instance, some kinetic and shear forces are significantly dampened, which leads to a loss of detail for buckling effects with traditional techniques. This paper introduces a new way of efficiently solving these operators together in a way that these coupling effects are retained in the simulation. The final solution not only looks stable, but is mathematically proven to work well for a variety of cases, and it also takes into consideration collisions with other solid objects correctly. I absolutely love this, and anyone who is in the middle of creating a new movie with some fluid action going on has to be all over this new technique. And the paper is absolutely amazing. It contains crystal clear writing, many paragraphs are so tight that I'd find it almost impossible to cut even one word from them, yet it is still digestible and absolutely beautifully written. Make sure to have a look, as always, the link is available in the video description. These amazing papers are stories that need to be told to everyone, not only to experts to everyone, and before creating these videos I always try my best to be in contact with the authors of these works. And nowadays many of them are telling me that they were really surprised by the influx of views they got after they were showcased in the series. Writing papers that are featured in two-minute papers takes a ridiculous amount of hard work, and after that the researchers make them available for everyone free of charge. And now I am so glad to see them get more and more recognition for their hard work. Absolutely amazing. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karoji Olaifa here. This paper is about designing and creating furniture with pieces that are geometrically interlocked. Such pieces not only have artistic value, but such structures can also enhance the integrity and sturdiness of a piece of furniture. This piece of work takes a simple 2D drawing of this interlocking structure and assembles the required pieces for us to build a 3D model from them. This drawing can be done with one of the most user friendly modeler program out there, Google Sketchup. This can be used even by novices. From these disassembled parts, it is highly non-trivial to create a 3D printable model. For instance, it is required that these pieces can be put together with one translational motion. Basically, all we need is one nudge to put two of these pieces together. If you ever had a new, really simple piece of furniture from IKEA, had a look at the final product at the shop and thought, well, I only have 10 minutes to put this thing together. But anyway, how hard can it be? And you know, three hours of curzing later, the damn thing is still not completely assembled. If you had any of those experiences before, this one push assembly condition is for you. And the algorithm automatically finds a sequence of motions that assembles our target 3D shape and because we only have 2D information from the input, it also has to decide how, and where to extrude, thicken, or subtract from these volumes. The third space of possible motions is immense. And we have to take into consideration that we don't even know if there is a possible solution for this puzzle at all. If this is the case, the algorithm finds out and proposes changes to the model that make the construction feasible. And if this wasn't enough, we can also put this digital furniture model into a virtual world where gravitational forces are simulated to see how stable the final result is. Here, the proposed yellow regions indicate that the stability of this table could be improved via small modifications. It is remarkable to see that a novice user who has never done a minute of 3D modeling can create such a beautiful and resilient piece of furniture. Really, really nice work, loving it. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Kato Ejone Fahir. Today, we are going to talk about explosions. To be precise, imagine that we already have the physics simulation data for an explosion on our computer, but we would like to visualize it on our screen. This requires a light simulation program that is able to create an image of this virtual scene that looks exactly the same as it would in reality. We have had plenty of earlier episodes on light transport, and as you know all too well, it is one of my favorite topics. I just can't get enough of it. I've put a link to these related episodes in the video description. If we wish to render a huge smoke plume, we perform something that computer graphics people call volumetric light transport. This means that a ray of light doesn't necessarily bounce off of the surface of materials, but it can penetrate their surfaces and scatter around inside of them. A technique that can deal with this is called volumetric path tracing, and if we wish to create an image of an explosion using that, well, better pack some fast food because it is likely going to take several hours. The explosion in this image took 13 hours, and it is still not rendered perfectly. But this technique is able to solve this problem in 20 minutes, which is almost 40 times quicker. Unbelievable. The key idea is that this super complicated volumetric explosion data can be reimagined as a large batch of point light sources. If we solve this light transport problem between these point light sources, we get a solution that is remarkably similar to the original solution with path tracing. However, solving this new representation is much simpler. But that's only the first step. If we have a bunch of light sources, we can create a grid structure around them, and in these grid points, we can compute shadows and illumination in a highly efficient manner. What's more, we can create multiple of these grid representations. They all work on the very same data, but some of them are finer, and some of them are significantly sparser, more coarse. Another smart observation here is that even though sharp, high frequency illumination details need to be computed on this fine grid, which takes quite a bit of computation time, it is sufficient to solve the coarse, low frequency details on one of these sparser grids. The results look indistinguishable from the ground truth solutions, but the overall computation time is significantly reduced. The paper contains detailed comparisons against other techniques as well. Most of these scenes are rendered using hundreds of thousands of these point light sources, and as you can see, the results are unbelievable. If you would like to learn even more about light transport, I am holding a master level course on this at the Vienna University of Technology in Austria. I thought that the teachings should not only be available for those 30 people who sit in the room who can afford a university education. It should be available for everyone. So we made the entirety of the lecture available for everyone, free of charge, and I am so glad to see that thousands of people have watched it, and to this day I get many messages that they enjoyed it, and now they see the world differently. It was recorded live with the students in the room, and it doesn't have the audio quality of two minute papers. However, what it does well is it conjures up the atmosphere of these lectures, and you can almost feel like one of the students sitting there. If you are interested, have a look. The link is available in the video description, and make sure to read this paper too. It's incredible. Thanks for watching, and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karojona Ifeher. We already know that by using a computer we can simulate fluids and we can simulate hair. But what about simulating both of them at the same time? This paper is about liquid hair interaction and simulating the dynamics of wet hair. Our season Fellow Scholars immediately know that this episode is going to be ample in amazing slow motion footage. I hope I didn't mess up with any of them, you will see soon if this is the case or not. Before we start talking about it, I'd like to note the following remarkable features. The authors uploaded a supplementary video in 4K resolution, executable files for their technique for all three major operating systems, data assets, and they also freshly revealed the full source code of the project. Pellia, I feel in heaven, a big two-minute paper style head tip to the authors for this premium quality presentation. If this paper were a car, it would definitely be a Maserati or a Mercedes. This technique solves the equations for liquid motion along every single hair strand computes the cohesion effects between the hairs and it can also simulate the effect of water dripping off the hair. Feast your eyes on these absolutely incredible results. The main issue with such an idea is that the theory of large and small scale simulations are inherently different and in this case we need both. The large scale simulator would be a standard program that is able to compute how the velocity and pressure of the liquid evolves in time. However, we also wish to model the water droplets contained within one tiny hair strand. With a large scale simulator, this would take a steepenously large amount of time and resources so the key observation is that a small scale fluid simulator program would be introduced to take care of this. However, these two simulators cannot simply coexist without side effects. As they are two separate programs that work on the very same scene, we have to make sure that as we pass different quantities between them, they will still remain intact. This means that a drop of water that gets trapped in a hair strand has to disappear from the large scale simulator and has to be re-edited to it when it rips out. This is a remarkably challenging problem. But with this, we only scratch the surface. Make sure to have a look at the paper that has so much more to offer, it is impossible to even enumerate the list of contributions within in such a short video. The quality of this paper simply left me speechless and I would encourage you to take a look as well. And while this amazing footage is rolling, I would like to let you know that two minute papers can exist because of your support through Patreon. Supporters of the series gain really cool perks like watching every single one of these episodes in Early Access. I am super happy to see how many of you decided to support the series and in return, we are able to create better and better videos for you. Thank you again, you fellow scholars are the most amazing audience. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Károly Zsolnai-Fehér. In this piece of work, we seek to control digital characters in real time. It happens the following way. We specify a target trajectory, and the algorithm has to synthesize a series of motions that follows that path. To make these motions as realistic as possible, this is typically accomplished by unleashing a learning algorithm on a large database that contains a ton of motion information. Previous techniques did not have a good understanding of these databases, and they often synthesized motions from pieces that corresponded to different kinds of movements. This lack of understanding results in stiff and natural output motion. Intuitively, it is a bit like putting together a sentence from a set of letters that were cut out one by one from different newspaper articles. It is a fully formed sentence, but it lacks the smoothness and the flow of a properly aligned piece of text. This is a neural network-based technique that introduces a face function to the learning process. This face function augments the learning with the timing information of a given motion. With this face function, the neural network recognizes that we are not only learning periodic motions, but it knows when these motions start and when they end. The final technique takes very little memory, runs in real time, and it accomplishes smooth walking, running, jumping, and climbing motions, and so much more over a variety of terrains with flying colors. In a previous episode, we have discussed a different technique that accomplished something similar with a low and high level controller. One of the major selling points of this technique is that this one offers a unified solution for terrain traversal with using only one neural network. This has the potential to make it really big on computer games and real-time animation. It is absolutely amazing to witness this and be a part of the future. Make sure to have a look at the paper, which also contains the details of a terrain-fitting step to make this learning algorithm capable of taking into consideration a variety of obstacles. I would also like to thank Claudio Panacci for his amazing work in translating so many of these episodes to Italian. This makes two-minute paper successful for more people around the globe and the more people we can reach the happier I am. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Ejona Ifeher. Earlier, we have talked about a few amazing algorithms to teach digital creatures to walk. And this time, we are interested in controlling the joints of a digital character to not only walk properly, but take into consideration its surroundings. This new version can navigate in 3D with static and dynamic moving obstacles or even dribble a ball toward a target. Loving the execution and the production value of this paper. This is accomplished by an efficient system that consists of two controllers that are represented by learning algorithms. One, the low-level controller is about maintaining balance and proper limb control by manipulating the joint positions and velocities appropriately. This controller operates on a fine time scale, 30 times per second, and is trained via a four-layer neural network. Two, the high-level controller can accomplish bigger overarching goals, such as following a path or avoiding static and dynamic obstacles. We don't need to run this so often, therefore, to save resources, this controller operates on a coarse time scale only twice each second and is trained via a deep convolutional neural network. It also has support for a small degree of transfer learning. Transfer learning means that after successfully learning to solve a problem, we don't have to start from scratch for the next one, but we can reuse some of that valuable knowledge and get a head start. This is a heavily researched area and is likely going to be one of the major next frontiers in machine learning research. Now, make no mistake, it is not like transfer learning is suddenly to be considered a solved problem, but in this particular case, it is finally a possibility. Really cool. I hope this brief expose fired you up too. This paper is a bomb, make sure to have a look, as always, the link is available in the video description. And by the way, with your support on Patreon, we will soon be able to spend part of our budget on empowering research projects. How amazing is that? The new two-minute paper shirts are also flying off the shelves. Happy to hear you're enjoying them so much. If you're interested, hit up two-minute papers.com if you are located in the US. The EU and Worldwide stores link is available in the video description. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two minute papers with Karo Ejona Ifeher. I just finished reading this paper and I fell out of the chair. And I can almost guarantee you that the results in this work are so insane, you will have to double or even triple check to believe what you're going to see here. This one is about image translation, which means that the input is an image and the output is a different version of this input image that is changed according to our guidelines. Imagine that we have a Monet painting and we'd like to create a photograph of this beautiful view. There we go. What if we'd like to change this winter landscape to an image created during the summer? There we go. If we are one of those people on the internet forums who just love to compare apples to oranges, this is now also a possibility. And have a look at this. Imagine that we like the background of this image, but instead of the zebras, we would like to have a couple of horses. No problem coming right up. This algorithm synthesizes them from scratch. The first important thing we should know about this technique is that it uses generative adversarial networks. This means that we have two neural networks battling each other in an arms race. The generator network tries to create more and more realistic images and these are passed to the discriminator network which tries to learn the difference between real photographs and fake forged images. During this process, the two neural networks learn and improve together until they become experts at their own craft. However, this piece of work introduces two novel additions to this process. One, in earlier works, the training samples were typically paired. This means that the photograph of a shoe would be paired to a drawing that depicts it. This additional information helps the training process a great deal and the algorithm would be able to map drawings to photographs. However, a key difference here is that without such pairings, we don't need these labels. We can use significantly more training samples in our data sets, which also helps the learning process. If this is executed well, the technique is able to pair anything to anything else, which results in a remarkably powerful algorithm. Key difference number two. A cycle consistency loss function is introduced to the optimization problem. This means that if we convert a summer image to a winter image and then back to a summer image, we should get the very same input image back. If our learning system obeys to this principle, the output quality of the translation is going to be significantly better. This cycle consistency loss is introduced as a regularization term. Our season fellow scholars already know what it means, but in case you don't, I've put a link to our explanation in the video description. The paper contains a ton more results and fortunately, the source code for this project is also available. Multiple implementations in fact. Just as a side note, which is draw dropping by the way, there's some rudimentary support for video. Amazing piece of work. Bravo! Now you can also see that the rate of progress in machine learning research is completely out of this world. No doubt that it is the best time to be a research scientist it's ever been. If you've liked this episode, make sure to subscribe to the series and have a look at our Patreon page where you can pick up cool perks like watching every single one of these episodes in early access. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is 2 Minute Papers with Karo Ejona Ifehir. Have a look at these beautiful images. Representing these materials that you see here takes more than 25GB of storage. You could store several feature length movies on your hard drive using the same amount of storage. And this technique is able to compress all these 25GB into 45Mb without introducing any significant perceptible difference. That is close to a whopping 500 times more efficient representation. Hmm... This improved representation not only helps easing the storage requirements of these assets, but it also makes the rendering times, or in other words, the process of creating these images via light simulation programs typically more than twice as fast to process. That is a ton of money and time saved for the artists. An important keyword in this piece of work is anisotropic scattering. So what does this mean exactly? The scattering part means that we have to imagine these materials not as a surface, but as a volume in which rays of light bounce around and get absorbed. If we render a piece of cloth made of velvet, twill, or a similar material, there are lots of microscopic differences in the surface, so much so that it is insufficient to treat them as a solid surface, such as wood or metals. We have to think about them as volumes. This is the scattering part. The anisotropy means that light can scatter unevenly in this medium. These rays don't bounce around in all directions with equal probability. This means that there is significant forward and backward scattering in this media, making it even more difficult to create more optimized algorithms that simplify these scattering equations. If you look below here, you'll see these colorful images that researchers like to call different images. It basically means that we create one image with the perfectly accurate technique as a reference. As expected, this reference image probably takes forever to compute, but is important to have as a yardstick. Then, we compute one image with the proposed technique that is usually significantly faster. So we have these two images, and sometimes the differences are so difficult to see, we have no way of knowing where the inaccuracies are. So what we do is subtract the two images from each other and assign a color coding for the differences. As the error may be spatially varying, this is super useful because we can recognize exactly where the information is lost. The angrier the colors are, the higher the error is in a given region. As you can see, the proposed technique is significantly more accurate in representing this medium than a naive method using the same amount of storage. This paper is extraordinarily well written. It is one of the finest pieces of craftsmanship I've come along in a long while, and yes, it is a crime not having a look at it. Also, if you like this episode, make sure to subscribe to the series and check out our other videos. We have more than 150 episodes for you ready to go right now. You'll love it, and there will be lots of fun to be had. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Ejona Ifehir. Let's have a look at this majestic technique that is about style transfer for photos. Style transfer is a magical algorithm where we have one photograph with content and one with an interesting style, and the output is a third image with these two photos fused together. This is typically achieved by a classical machine learning technique that we call a convolution on your own network. The more layers these networks contain, the more powerful they are, and the more capable they are in building an intuitive understanding of an image. We had several earlier episodes on visualizing the inner workings of these neural networks, as always, the links are available in the video description. Don't miss out, I'm sure you'll be as amazed by the results as I was when I first seen them. These previous neural style transfer techniques work amazingly well if we are looking for a painterly result. However, for photo style transfer, the close-ups here reveal that they introduce unnecessary distortions to the image. They won't look realistic anymore. But not with this new one. Have a look at these results. This is absolute insanity. They are just right in some sense. There is an elusive quality to them. And this is the challenge. We not only have to put what we are searching for into words, but we have to find a mathematical description of these words to make the computer executed. So what would this definition be? Just think about this. This is a really challenging question. The author decided that the photorealism of the output image is to be maximized. Well this sounds great, but who really knows a rigorous mathematical description of photorealism. One possible solution would be to stipulate that the changes in the output color would have to preserve the ratios and distances of the input style colors. Similar rules are used in linear algebra and computer graphics to make sure shapes don't get distorted as we are tormenting them with rotations, translations, and more. We like to call these operations affine transformations. So the fully scientific description would be that we add a regularization term that stipulates that these colors only undergo affine transformations. And we've used one more new word here. What does this regularization term mean? This means that there are a ton of different possible solutions for transferring the colors and we are trying to steer the optimizer towards solutions that adhere to some additional criterion in our case the affine transformations. In the mathematical description of this problem, these additional stipulations appear in the form of a regularization term. I am so happy that you fellow scholars have been watching too many papers for so long that we can finally talk about techniques like this. It's fantastic to have an audience that has this level of understanding of these topics. Love it. Just absolutely love it. The source code of this project is also available. Also, make sure to have a look at this still, an absolutely amazing new science journal from the Google Brain team. That this is no ordinary journal because what they are looking for is not necessarily novel techniques, but novel and intuitive ways of explaining already existing works. There is also an excellent write-up on research that can almost be understood as a manifesto for this journal. A worthy reading deed. I love this new initiative and I am sure we'll hear about this journal a lot in the near future. Make sure to have a look. There is a link to all of these in the video description. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two minute papers with Karo Ejona Efeira. How could it be to be able to place a character representing us in a digital film or a computer game? Of course, it would clearly be an extremely laborious task to digitize the 3D geometry and the Albedo map of our face. This Albedo map means a texture, a colored pattern that describes how our skin reflects and absorbs light. Capturing such a representation is clearly a very lengthy and expensive process, so get this completely crazy idea. This technique creates this full digital representation of any face from no more than one simple photograph. We can even get historical figures in our digital universe, all we need is one photograph of them. And now, feast your eyes on these incredible results. After taking a photograph, this technique creates two of these Albedo maps. One is a complete, low-frequency map which records the entirety of the face but only contains the rough details. The other Albedo map contains finer details, but in return, it's incomplete. Do you remember the textures synthesis methods that we discussed earlier in the series? The input was a tiny patch of image with a repetitive structure and after learning the statistical properties of these structures, it was possible to continue them indefinitely. The key insight is that we can also do something akin to that here as well. We take this incomplete Albedo map and try to synthesize the missing details. Pretty amazing idea indeed. The authors of the paper invoke a classical learning algorithm, a convolutional neural network to accomplish that. The deeper the neural network we use, the more high frequency details appear on the outputs or in other words, the crisper the image we get. In the paper, you will find a detailed description of their crowdsource user study that was used to validate this technique, including the user interface and the questions being asked. There are also some comparisons against patch match. One of the landmark techniques for texture synthesis that we have also talked about in an earlier episode. It's pretty amazing to see these two-minute papers knowledge-based grow and get more and more intertwined. I hope you're enjoying the process as much as I do. Also, due to popular requests, the two-minute paper's t-shirts are now available. This time, we are using a different service for printing these shirts. Please give us some feedback on how you liked it. I've put my email address in the video description. If you attach a photo of yourself wearing some cool two-minute papers merch, we'll be even more delighted. Just open two-minute papers.com and you'll immediately have access to it. This link will bring you to the service that ships to the US. The link for shipping outside the US is available in the video description. Thanks for watching and for your generous support. I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Karoizhou, Naifahir. Today, we are going to discuss a great piece of work that shows us how efficient and versatile neural network based techniques have become recently. Here, the input is a bunch of RGBD images, which are photographs endowed with depth information, and the output can be a full 3D reconstruction of a scene and much, much more, which we'll see in a moment. This task is typically taken care of by handcrafting descriptors. A descriptor is a specialized representation for doing useful tasks on images and other data structures. For instance, if we seek to build an algorithm to recognize black and white images, a useful descriptor would definitely contain the number of colors that are visible in an image and the list of these colors. Again, these descriptors have been typically handcrafted by scientists for decades, new problem, new descriptors, new papers. But not this time, because here, super effective descriptors are proposed automatically via a learning algorithm, a convolutional neural network, and Xiaomi's networks. This is incredible. Creating such descriptors took extremely smart researchers and years of work on a specific problem and were still often not as good as these ones. By the way, we have discussed Xiaomi's networks in an earlier episode, as always, the link is available in the video description. And as you can imagine, several really cool applications emerge from this. One, when combined with RANDSAC, a technique used to find order in noisy measurement data, it is able to perform 3D scene reconstructions from just a few images, and it completely smokes the competition. Two, pause estimation with bounding boxes. Even a sample of an object, the algorithm is able to recognize not only the shape itself, but also its orientation when given a scene cluttered with other objects. Three, correspondence search is possible. This is really cool. This means that a semantically similar piece of geometry is recognized on different objects. For instance, the algorithm can learn the concept of a handle and recognize the handles on a variety of objects, such as on motorcycles, carriages, chairs, and more. The source code of this project is also available. Yuhu! Neural networks are rapidly establishing supremacy in a number of research fields, and I am so happy to be alive in this age of incredible research progress. Make sure to subscribe to the series and click the bell icon. Some amazing works are coming up in the next few episodes, and there will be lots of fun to be had. Thanks for watching, and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Károly Zsolnai-Fehér. This piece of work is an amazing application of deep neural networks that performs semantic scene completion from only one depth image. This depth image is the colorful image that you see here, where the colors denote how far away different objects are from our camera. We can create these images inexpensively with commodity hardware, for instance Microsoft Kinect has a depth sensor that is suitable for this task. The scene completion part means that from this highly incomplete depth information, the algorithm reconstructs the geometry for the entirety of the room. Even parts that are completely missing from our images are things that are occluded. The output is what computer graphics researchers like to call a volumetric representation or a voxel array, which is essentially a large collection of tiny LEGO pieces that build up the scene. But this is not all because the semantic part means that the algorithm actually understands what we are looking at and thus is able to classify different parts of the scene. These classes include walls, windows, floors, sofas, and other furniture. Previous works were able to do scene completion and geometry classification, but the coolest part of this algorithm is that it not only does these steps way better, but it does them both at the very same time. This work uses a 3D convolutional neural network to accomplish this task. The 3D part is required for this learning algorithm to be able to operate on this kind of volumetric data. As you can see, the results are excellent and are remarkably close to the ground truth data. If you remember, not so long ago, I flipped out when I've seen the first neural network based techniques that understood 3D geometry from 2D images. That technique used a much more complicated architecture, a generative adversarial network, which also didn't do scene completion and on top of that, the resolution of the output was way lower, which intuitively means that the LEGO pieces were much larger. This is insanity. The rate of progress in machine learning research is just stunning. Probably even for you, season fellow scholars who watch too many papers and have high expectations. We've had plenty of previous episodes about the inner workings of different kinds of neural networks. I've put some links to them in the video description. Make sure to have a look if you wish to brush up on your machine learning concfurbit. The authors also published a new data set to solve these kind of problems in future research works. And it is also super useful because the output of their technique can be compared to ground truth data. When new solutions pop up in the future, this data set can be used as a yardstick to compare results with. The source code for this project is also available. Tinkerers rejoice. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karojejona Ifeher. This paper is about interactively modeling and editing, climbing plants. This is one of my favorite kind of works, mundane sounding topic, immaculate execution. There are so many cool things about this paper I don't even know where to start. But first, let's talk about the modeling part. We can, for instance, plant a seed and we can not only have a look at how it grows as time goes by, but we can also influence the grow variability and shoot growth rates. Branches can be added and removed at will at any point in time. We can also add attractors, regions that are set to be more likely for the plant to grow towards. With these techniques, we can easily create any sort of artistic effect, be it highly artificial looking vines and branches, or some long forgotten object overgrown with climbing plants. However, a model is just 3D geometry. What truly makes these models come alive is animation, which is also executed with flying colors. The animations created with this technique are both biologically and physically plausible. So what do these terms mean exactly? Biologically plausible means that the plants grow according to the laws of nature, and physically plausible means that if we start tugging at it, branches start moving, bending and breaking according to the laws of physics. Due to its responsive and interactive nature, the applications of this technique are typically in the domain of architectural visualization, digital storytelling, or any sort of real-time application. And of course, the usual suspects, animated movies, and game developers can use this to create more immersive digital environments with ease. And don't forget about me, Karoy, who would happily play with this basically all day long. If you are one of our many fellow scholars who are completely addicted to two-minute papers, make sure to check out our Patreon page where you can grab cool perks like watching these episodes in early access, or deciding the order of upcoming episodes, and more. Also, your support is extremely helpful, so much so that even the price of a cup of coffee per month helps us to create better videos for you. We write some reports from time to time to assess the improvements we were able to make with your support. The link is in the video description, or you can just click the letter P on the end link screen in a moment. Thanks for watching, and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Karoizhou, naifahir. Today, we are going to talk some more about fluid simulations and fluid guiding. Hell yeah! As you know, all too well, it is possible to simulate the loss of fluid motion on a computer, make a digital scene, and create absolutely beautiful videos, such as the ones you see here. Newer and newer research papers show up to both extend the possible scenarios that we can simulate, and there are also other works to speed up already existing solutions. This piece of work introduces a technique that mathematicians like to call the primal dual optimization method. This helps us accomplish two really cool things. One is fluid guiding. Fluid guiding is a problem where we are looking to exert control over the fluid, while keeping the fluid flow as natural as possible. I've written my master thesis on the very same topic and can confirm that it's one hell of a problem. The core of the problem is that if we use the loss of physics to create a fluid simulation, we get what would happen in reality as a result. However, if we wish to guide this piece of fluid towards a target shape, for instance to form an image of our choice, we have to both retain natural fluid flows, but still creates something that would be highly unlikely to happen according to the loss of physics. For instance, a splash of water is unlikely to suddenly form a human face of our choice. The proposed technique helps this ambivalent goal of exerting a bit of control over the fluid simulation, while keeping the flows as natural as possible. There are already many existing applications of fluids and smoke in movies, where an actor fires a gun and the fire and smoke plumes are added to the footage in post-production. However, with a high-quality fluid guiding technique, we could choose target shapes for these smoke plumes and explosions that best convey our artistic vision. And number two, it also accomplishes something that we call separating boundary conditions, which prevents imprecisions where small fluid volumes are being stuck to walls. The guiding process is also followed by an upsand-link step, where we take a core simulation and artificially synthesize sharp, high-frequency details onto it. Computing the more detailed simulation would often take days without such synthesizing techniques, kind of like with wavelet turbulence, which is an absolutely incredible paper that was showcased in none other than the very first two-minute paper's episode. Link is in the video description box. Don't watch it. It's quite embarrassing. And all this leads to eye-poppingly beautiful solutions. Wow, I cannot get tired of this. In the paper, you will find much more about breaking dams, tornado simulations, and applications of the primal dual optimization method. Normally, to remain as authentic to the source materials as possible, I don't do any kind of slow motion and other similar shenanigans, but this time I just couldn't resist it. Have a look at this, and I hope you'll like the results. If you feel the alluring call of fluids, I've put some resources in the video description, including a gentle description I wrote on the basics of fluid simulation and fluid control, with source code, both on the CPU and GPU, and the link to Doyub Kim's amazing book that I'm currently reading. Highly recommended. If you also have some online tutorials and papers that help you solidify your understanding of the topic, make sure to leave a link in the comments. I'll include the best ones in the video description. If you would like to see more episodes like this one, make sure to subscribe to two-minute papers. We would be more than happy to have you along on our journey of science. Thanks for watching, and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two minute papers with Karo Zsolnai-Fehir. You are going to love this killer paper. In the classical case of 3D fabrication, we first create the 3D geometry in our modeling software on a computer. Then, after we are done, we send this model to a 3D printer to create a physical copy of it. If we don't like an aspect of this printed model, we have to go back to the computer and adjust accordingly. If there are more fundamental issues, we may even have to start over. And get this, with this piece of work, we can have a low fidelity wireframe version printed immediately as we make changes within the modeling software. This process we can refer to as real time or on the fly 3D printing. In this work, both the hardware design and the algorithm that runs the printer is described. This approach has a number of benefits and, of course, a huge set of challenges. For instance, we can immediately see the result of our decisions and can test whether a new piece of equipment would correctly fit into the scene we are designing. Sometimes, depending on the geometry of the final object, different jobs need to be reordered to get a plan that is physically plausible to print. In this example, the bottom branch was designed by the artist first and the top branch afterwards. But their order has to be changed, otherwise, the bottom branch would block the way to the top branch. The algorithm recognizes these cases and reorders the printing jobs correctly. Quite remarkable. And, an alternative solution for rotating the object around for better reachability is also demonstrated. Because of the fact that the algorithm is capable of this sort of decision-making, we don't even need to wait for the printer to finish a given step and can remain focused on the modeling process. Also, the handle of the teapot here collides with the body. Because of the limitations of wireframe modeling, such cases have to be detected and omitted. Connecting patches and adding differently sized holes to a model are also highly non-trivial problems that are all addressed in the paper. And this piece of work is also a testament to the potential of solutions where hardware and software is designed with each other in mind. I can only imagine how many work hours were put in this project until the final working solution was obtained. Incredible work indeed. We really just scratched the surface in this episode. Make sure to go to the video description and have a look at the paper for more details. It's definitely worth it. Also, if you enjoyed this two-minute paper's episode, make sure to subscribe to the series and if you are subscribed, click the bell icon to never miss an episode. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Karo Ejol Naifahir. It's been quite a while since we've talked about a paper on fluid simulations. Since my withdrawal symptoms are already kicking in, today I simply have to talk to you about this amazing paint simulator program that runs in real-time and on our mobile devices. These handheld devices typically have a lower image resolution compared to desktop computers, therefore it is indeed a challenge to put together a solution that artists can use to create detailed paintings with. And to accomplish this, this piece of work offers several killer features. For instance, the paint pigment concentration can be adjusted. The direction of the brushstrokes is also controllable. And third, this technique is powered by a viscoelastic shallow water simulator that also supports simulating multiple layers of paint. This is particularly challenging as the inner paint layers may have already dried when adding a new wet layer on top of them. This all has to be simulated in a way that is physically plausible. But we are not done yet. With many different kinds of paint types that we are using, the overall outlook of our paintings are dramatically different depending on the lighting conditions around them. To take this effect into consideration, this technique also has an intuitive feature where the effect of virtual light sources is also simulated and the output is changed interactively as we tilt the tablet around. And get this, gravity is also simulated and the paint trickles down depending on the orientation of our tablet according to the laws of physics. Really cool. The paper also shows visual comparisons against similar algorithms. And clearly, artists who work with these substances all day know exactly how they should behave in reality. So the ultimate challenge is always to give it to them and ask them whether they have enjoyed the workflow and found the simulation faithful to reality. Let the artist speed the judge. The user study presented in the paper revealed that the artists loved the user experience and they expressed that it's second to none for testing ideas. I am sure that with a few improvements, this could be the ultimate tool for artists to unleash their creative potential while sitting outside and getting inspired by nature. Thanks for watching and for your generous support. I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Kato Ejolene Ifehir. In this piece of work, we are interested in creating a 3D virtual tour for an apartment. However, for this apartment, no 3D information is available. Instead, the input for the algorithm is something that we can obtain easily. In this case, a 2D floor plan and a set of images that we shot in the apartment. From this information, we would create a 3D floor plan that is not only faithful to the real one in terms of geometry, but the photos with the correct viewpoints are also to be assigned to the correct walls. In order to accomplish this, one has to overcome a series of challenging problems. For instance, we have to estimate the layout of each room and find the location of the camera in each of these images. Also, to obtain high quality solutions, the goal is to extract as much information from the inputs as possible. The authors recognized that the floor plans provide way more information than we take for granted. For instance, beyond showing the geometric relation of the rooms, it can also be used to find out the aspect ratios of the floor for each room. The window tour ratios can also be approximated and matched between the photos and the floor plan. This additional information is super useful when trying to find out which room is to be assigned to which part of the 3D floor plan. On just looking at the photos, we also have access to a large swath of learning algorithms that can reliably classify whether we are looking at a bathroom or a living room. There are even more constraints to adhere to in order to aggressively reduce the number of physical configurations. Make sure to have a look at the paper for details. There are lots of cool tricks described there. As always, there is a link to it in the video description. For instance, since the space of possible solutions is still too vast, a branch and bound type algorithm is proposed to further decimate the number of potential solutions to evaluate. And as you can see here, the comparisons against ground truth floor plans reveal that these solutions are indeed quite faithful to reality. The authors also kindly provided a dataset with more than 200 full apartments with well over a thousand photos and annotations for future use in follow-up research works. Creating such a dataset and publishing it is incredibly laborious and could easily be a paper on its own, and here we also get an excellent solution for this problem as well. In a separate work, the authors also published a different version of this problem formulation that reconstructs the exterior of buildings in a similar manner. There is so much to explore. The links are available in the video description. Make sure to have a look. In case you're wondering, it's still considered a crime not doing that. I hope you have enjoyed this episode and I find it so delightful to see this unbelievably rapid growth on the channel. Earlier I thought that even too would be amazing, but now we have exactly 8 times as many subscribers as one year ago. Words fail me to describe the joy of showing these amazing works to such a rapidly growing audience. This is why I always say at the end of every episode, thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Kato Ejone Fehr. In the world of Digital 3D modeling, it often occurs that we are looking for surfaces that are not perfectly smooth but have some sort of surface detail. Wrinkles and gravings, grain on a wooden table are excellent examples of details that we can add to our models and computer graphics people like to collectively call these things displacement maps. Artists often encounter cases where they like the displacements on one object but the object itself is not really interesting. However, it could be that there is a different piece of geometry these details would look great on. Consider this problem solved because in this piece of work, the input is two 3D models. One, with interesting geometric details and the other is the model onto which we transfer these surface details. The output will be our 3D geometric shape with two of these models fused together. The results look absolutely amazing. I would love to use this right away in several projects. The first key part is the usage of metric learning. Wait, technical term. What does this mean exactly? Metric learning is a classical technique in the field of machine learning where we are trying to learn distances between things where distance is mathematically ill-defined. Let's make it even simpler and go with an example. For instance, we have a database of human faces and we would like to search for faces that are similar to a given input. To do this, we specify a few distances by hand. For instance, we could say that a person with a beard is a short distance from one with a mustache and the larger distance from one with no facial hair. If we hand many examples of these distances to a learning algorithm, it will be able to find people with similar beards. And in this work, this metric learning is used to learn the relationship between objects with and without these rich surface details. This helps in the transferring process. As to creating the new displacements on the new model, there are several hurdles to overcome. One, we cannot just grab the displacements and shove them onto a different model because it can potentially look different, have different curvatures and sizes. The solution to this would be capturing the statistical properties of the surface details and use this information to synthesize new ones on the target model. Note that we cannot just perform this texture synthesis in 2D like we do for images because as we project the result to a 3D model, it introduces severe distortions to the displacement patterns. It is a bit like putting a rubber blanket onto a complicated object. Different regions of the blanket will be distorted differently. Make sure to have a look at the paper where the authors present quite a few more results and of course the intricacies of this technique are also described in detail. I hope some public implementations of this method will appear soon. I would be quite excited to use this right away and I am sure there are many artists who would love to create these wonderfully detailed models for the animated films and computer games of the future. In the meantime, we have a completely overhauled software and hardware pipeline to create these videos. We have written down our joyful and perilous story of it on Patreon. If you are interested in looking a bit behind the curtain as to how these episodes are made, make sure to have a look it is available in the video description. Thanks for watching and for your generous support and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Kato Ejona Ifehir. Not so long ago, we talked about a technique that enabled us to render stunningly high-quality cloth models in real time. It supported level of detail, self-shadowing, and lots of other goodies that make computer game developers, and of course, my humble self, super happy. And today, we are going to talk about a technique that is able to create these highly detailed cloth geometries for our digital characters. I have really fond memories of attending to the talk of the Oscar Award winner Steve Martiner on this paper a few years ago in Switzerland, and I remember being so spellbound by it that I knew this was a day I will never forget. I am sure you'll love it too. In this piece of work, the goal is to create a digital garment model that is as detailed and realistic as possible. We start out with an input 3D geometry that shows the rough shape of the model, then we pick a knitting pattern of our choice. After that, the points of this knitting pattern are moved so that they correctly fit this 3D geometry that we specified. And now comes the coolest part. What we created so far is an ad hoc model that doesn't really look and behave like a real piece of cloth. To remedy this, a physics-based simulation is run that takes this ad hoc model and the output of this process will be a realistic rest shape for these yarn curves. And here you can witness how the simulated forces pull the entire piece of garment together. We start out with dreaming up a piece of cloth geometry and this simulator gradually transforms it into a real world version of that. This is a step that we call yarn level relaxation. Wow. These final results look not only magnificent, but in a physical simulation they also behave like real garments. It's such a joy to look at results like this, loving it. Again, I would like to note that we are not talking about the visualization of the garment but creating a realistic piece of geometry. The most obvious drawback of this technique is its computation time. It was run on a very expensive system and still took several hours of number crunching to get this done. However, I haven't seen an implementation of this on the graphics card yet, so if someone can come up with an efficient way to do it, in an ideal case we may be able to do this in several minutes. I also have to notify you about the fact that it is considered a crime not having a look at the paper in the video description. It does not suffice to say that it is well written. It is so brilliantly presented, it's truly a one-of-a-kind work that everyone has to see. If you enjoyed this episode, make sure to subscribe to Two Minute Papers with Be Happy to have you in our Growing Club of Fellow Scholars. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is 2 Minute Papers with Kato Ejo and Afehir. We are currently more than 130 episodes into the series and we still haven't talked about this algorithm. How could we go on for so long without patch match? So let's do this right now. You'll love this one. This technique helps us to make absolutely crazy modifications to previously existing photographs and it is one of the landmark papers for all kinds of photo manipulation which is still widely used to this day. Consider the following workflow. We have this image as an input. Let's mark the roof line for hole filling or image in painting as the literature refers to it and the hole is now filled with quite sensible information. Now we mark some of the pillars to reshape the object and then we pull the roof upward. The output is a completely redesigned version of the input photograph. Wow! Absolutely incredible. And the whole thing happens interactively, almost in real time. But if we consider the hardware improvement since this paper was published, it is safe to say that today it runs in real time even on a mediocre computer. And in this piece of work, the image completion part works by adding additional hints to the algorithm. For instance, marking the expected shape of an object that we wish to cut out and have it filled with new data. Moreover, showing the shape of the building that we wish to edit also helps the technique considerably. These results are so stunning. I remember when I had first seen them. I had to recheck over and over again because I could hardly believe my eyes. This technique offers not only these high quality results but it is considerably quicker than its competitors. To accomplish image in painting, most algorithms look for regions in the image that are similar to the one that is being removed and borrow some information from there for the filling process. Here, one of the key ideas that speed up the process is that when good correspondences are found, if we are doing another look-up, we shouldn't restart the patch matching process, but we should try to search nearby because that's where we are most likely to find useful information. You know what the best part is? What you see here is just the start. This technique does not use any of the modern machine learning techniques, so in the era of these incredibly powerful deep neural networks, I can only imagine the quality of solutions will be able to obtain in the near future. We are living amazing times indeed. Thanks for watching and for your generous support. Bye.

Dear Fellow Scholars, this is two minute papers with Karo Ejonei Fehere. This one is going to be absolutely amazing. This piece of work is aimed to help a machine build a better understanding of images and 3D geometry. Imagine that we have a large database with these geometries and images, and we can search and compare them with arbitrary inputs and outputs. What does this mean exactly? For instance, it can handle a text input such as school bus and automatically retrieve 3D models, sketches and images that depict these kinds of objects. This is great, but we said that it supports arbitrary inputs and outputs, which means that we can use the 3D geometry of a chair as an input and obtain other similar looking chairs from the database. This technique is so crazy it can even take a sketch as an input and provide excellent quality outputs. We can even give it a heat map of the input and expect quite reasonable results. Typically, these images and 3D geometries contain a lot of information, and to be able to compare which is similar to which we have to compress this information into a more concise description. This description offers a common ground for comparisons. We like to call these embedding techniques. Here, you can see an example of a 2D visualization of such an embedding of word classes. The retrieval from the database happens by compressing the user provided input and putting it into this space and fetching the results that are the closest to it in this embedding. Before the emergence of powerful learning algorithms, these embeddings were typically done by hand. But now, we have these deep neural networks that are able to automatically create solutions for us that are in some sense optimal, meaning that according to a set of rules, it will always do better than we would by hand. We get better results by going to sleep and leaving the computer on overnight, then we would have working all night using the finest algorithms from 10 years ago. Isn't this incredible? The interesting thing is that here, we are able to do this for several different representations. For instance, a piece of 3D geometry or 2D color image is being embedded into the very same vector space opening up the possibility of doing these amazing comparisons between completely different representations. The results speak for themselves. This is another great testament to the power of convolution on your own networks and as you can see, the rate of progress in AI and machine learning research is absolutely stunning. Also, big thumbs up for the observant fellow scholars out there who noticed the new outro music and some other minor changes in the series. If you are among those people, you can consider yourself a hardcore 2 minute paper scholar. High five! Thanks for watching and for your generous support. I'll see you next time.

Dear Fellow Scholars, this is Two Minute Papers with Karo Ejona Efehir. Today, we are going to talk about an algorithm that is capable of filling holes in space and time. Classical image in painting or in other words, filling holes in images is something that we've explored in earlier episodes. There are several incredible techniques to take care of that. But with this piece of work, it is possible to generalize such a technique for video and fill holes for not only one image, but a series of images, like removing a nambraela to create an unaccluded view to the beach, or removing a waving person from a video of us jogging. These results are truly incredible, and even though this method was published long, long ago, it still enjoys a great deal of reverence among computer graphics practitioners. Not only that, but this algorithm also serves the basis of the awesome content to where field feature in Adobe Photoshop CS5. In this problem formulation, we have to make sure that our solution has spatio-temporal consistency. What does this mean exactly? This means that holes can exist through space and time, so multiple frames of a video may be missing, or there may be regions that we wish to cut out not only for one image, but for the entirety of the video. The field in regions have to be consistent with their surroundings if they are looked at as an image, but there also has to be a consistency across the time domain, otherwise we would see a disturbing flickering effect in the results. It is a really challenging problem indeed, because there are more constraints that we have to adhere to. However, a key observation is that it is also easier, because in return, we have access to more information that comes from the previous and next frames in the video. For instance, here you can see an example of retouching old footage by removing a huge pesky artifact. And clearly, we know the fact that Charlie Chaplin is supposed to be in the middle of the image, only because we have this information from the previous and next frames. All this is achieved by an optimization algorithm that takes into consideration that consistency has to be enforced through the spatial and the time domain at the same time. It can also be used to fill in completely missing frames of a video, or it also helps where we have parts of an image missing after being removed by an image stabilizer algorithm. Video editors do this all the time, so such a restoration technique is super useful. The source code of this technique is available, I've put a link to it in the video description. Thanks for watching and for your generous support, and I'll see you next time.

Dear Fellow Scholars, this is two-minute papers with Kato Ejona Ifehir. Neural style transfer is an incredible technique where we have two input photographs, and the output would be a combination of these two, namely the content of one and the artistic style of the other fused together. When the first paper appeared on this topic, the news took the world by storm, and lots of speculative discussions emerged, as to what this could be used for, and how it would change digital arts and the video game industry. It is great fun to use these algorithms, and we have also witnessed a recent proliferation of phone apps that are able to accomplish this, which is super cool because of two reasons. One, the amount of time to go from a published research paper to industry-wide application has never been so small, and two, the first work required a powerful computer to accomplish this, and took several minutes of strenuous computation, and now, less than two years later, it's right in your pocket and can be done instantly. Talk about exponential progress in science and research. Absolutely amazing. And now, while we feast our eyes upon these beautiful results, let's talk about the selling points of this extension of the original technique. The paper contains a nice formal explanation of the weak points of the existing style transfer algorithms. The intuition behind the explanation is that the neural networks think in terms of neuron activations, which may not be proportional to the calorie intensities in the source image styles, therefore their behavior often becomes inconsistent or different than expected. The authors propose thinking in terms of histograms, which means that the output image should rely on statistical similarities with the source images. And as we can see, the results look outstanding, even when compared to the original method. It is also important to point out that this proposed technique is also more arc-directible. Make sure to have a look at the paper for more details on that. As always, I've put a link in the video description. This extension is also capable of texture synthesis, which means that we give it a small image patch that shows some sort of repetition, and it tries to continue it indefinitely, in a way that seems completely seamless. However, we have to be acutely aware of the fact that in the computer graphics community, texture synthesis is considered a subfield of its own with hundreds of papers, and one has to be extremely sure to have a clear cut selling point over the scale of the art. For the more interested fellow scholars out there, I've put a survey paper on this in the video description. Make sure to have a look. Thanks for watching and for your generous support. That's CU, next time.

Dear Fellow Scholars, this is two-minute papers with Kato Ijola Ifehir. Not so long ago, Google DeepMind introduced a novel learning algorithm that was able to reach superhuman levels in playing many Atari games. It was a spectacular milestone in AI research. Interestingly, while these learning algorithms are being improved at the staggering pace, there's a parallel subfield where researchers endeavor to break these learning systems by slightly changing the information they are presented with. Fraudulent pampering with images or video feeds, if you will. Imagine a system that is designed to identify what is seen in an image. In an earlier episode, we discussed an adversarial algorithm where, in an amusing example, they added a tiny bit of barely perceptible noise to this image to make the deep neural network misidentify a bus for an ostrich. Machine learning researchers like to call these evil forged images adversarial samples. And now, this time around, OpenAI published a super fun piece of work to fool these game learning algorithms by changing some of their input visual information. As you will see in a moment, it is so effective that by only using a tiny bit of information, it can turn a powerful learning algorithm into a blabbering idiot. The first method adds a tiny bit of noise to a large portion of the video input, where the difference is barely perceptible, but it forces the learning algorithm to choose a different action that it would have chosen otherwise. In the other one, a different modification was used that has a smaller footprint, for instance in Pong, adding a tiny fake ball to the game to coerce the learner into going down when it was originally planning to go up. The algorithm is able to learn game specific knowledge for almost any other game to fool the player. Despite the huge difference in the results, I love the elegant mathematical formulation of the two noise types, because despite the fact that they do something radically different, their mathematical formulation is quite similar. Mathematicians like to say that we are solving the same problem while optimizing for different target norms. One deep mind's deep-queue learning, two other high-quality learning algorithms are also fooled by this technique. In the white box formulation, we have access to the inner workings of the algorithm. But interestingly, a black box formulation is also proposed, where we know much less about the target system, but we know the game itself and we train our own system and look for weaknesses in that. When we found these weak points, we used this knowledge to break other systems. I can only imagine how much fun there was to be had for the authors when they were developing these techniques. Super excited to see how these arms raised of creating more powerful learning algorithms and in response, more powerful adversarial techniques to break them develops. In the future, I feel that the robustness of a learning algorithm, or in other words, its resilience against adversarial attacks will be just as important of a design factor as how powerful it is. There are a ton of videos published on the authors' website, make sure to have a look. And also, if you wish to support the series, make sure to have a look at our Patreon page. We kindly thank you for your contribution. It definitely helps keeping the series running. Thanks for watching and for your generous support. I'll see you next time.

Dear Fellow Scholars, this is 2 Minute Papers with Kato Ejonei Fahir. Have a look at this magnificent idea. The input is a digital 3D model of an object and the viewpoint of our choice and the output is an easy to follow step-by-step breakdown on how to draw it. Automated drawing tutorials. I wish tools like this were available back when I was a child. Awesome. This technique offers a way to create the scaffoldings to help achieving the correct perspective and positioning for the individual elements of the 3D model, something that novice artists often struggle with. This problem is particularly challenging because we have a bunch of competing solutions and we have to decide which one should be presented to the user. To achieve this, we have to include a sound mathematical description of how easy a drawing process is. The algorithm also makes adjustments to the individual parts of the model to make them easier to draw without introducing severe distortions to the shapes. The proposed technique uses graph theory to find a suitable ordering of the drawing steps. Beyond the scientific parts, there are a lot of usability issues to be taken into consideration. For instance, the algorithm should notify the user when a given guide is not to be used anymore and can be safely erased. novice, apprentice and adapt users are also to be handled differently. To show the validity of this solution, the authors made a user study where they tested this new tutorial type against the most common existing solution and found that the users were not only able to create more accurate drawings with it, but they were also enjoying the process more. I commend the authors for taking into consideration the overall experience of the drawing process, which is an incredibly important factor. If the user enjoyed the process, he'll surely come back for more and of course, the more we show up, the more we learn. Some of these tutorials are available on the website of the authors, as always, I've linked it in the video description. If you're in the mood to draw, make sure to give it a go and let us know how it went in the comment section. Hell, even I am now in the mood to give this a try. If I disappear for a while, you know where I am. Thanks for watching and for your generous support, and I'll see you next time.