Unnamed: 0
int64 2
9.3k
| sentence
stringlengths 30
941
| aspect_term_1
stringlengths 1
32
⌀ | aspect_term_2
stringlengths 2
27
⌀ | aspect_term_3
stringlengths 2
23
⌀ | aspect_term_4
stringclasses 25
values | aspect_term_5
stringclasses 7
values | aspect_term_6
stringclasses 1
value | aspect_category_1
stringclasses 9
values | aspect_category_2
stringclasses 9
values | aspect_category_3
stringclasses 9
values | aspect_category_4
stringclasses 2
values | aspect_category_5
stringclasses 1
value | aspect_term_1_polarity
stringclasses 3
values | aspect_term_2_polarity
stringclasses 3
values | aspect_term_3_polarity
stringclasses 3
values | aspect_term_4_polarity
stringclasses 3
values | aspect_term_5_polarity
stringclasses 3
values | aspect_term_6_polarity
stringclasses 1
value | aspect_category_1_polarity
stringclasses 3
values | aspect_category_2_polarity
stringclasses 3
values | aspect_category_3_polarity
stringclasses 3
values | aspect_category_4_polarity
stringclasses 1
value | aspect_category_5_polarity
stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
8,814 | Finally it's better to do some experiments on machine translation or speech recognition and see how the improvement on BLEU or WER could get. [experiments-NEU], [IMP-NEU] | experiments | null | null | null | null | null | IMP | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,817 | Regarding the latter methods: what is described in the paper sounds like competent engineering details that those performing such a task for launch in a real service would figure out how to accomplish, and the specific reported details may or may not represent the 'right' way to go about this versus other choices that might be made.[details-NEU], [EMP-NEU] | details | null | null | null | null | null | EMP | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,818 | The final threshold for 'successful' speedups feels somewhat arbitrary -- why 16ms in particular? [null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,819 | In any case, these methods are useful to document, but derive their value mainly from the fact that they allow the use of the completion/correction methods that are the primary contribution of the paper.[contribution-NEU], [EMP-NEU] | contribution | null | null | null | null | null | EMP | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,820 | While the idea of integrating the spelling error probability into the search for completions is a sound one, the specific details of the model being pursued feel very ad hoc, which diminishes the ultimate impact of these results.[idea-NEU], [EMP-NEU] | idea | null | null | null | null | null | EMP | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,821 | Specifically, estimating the log probability to be proportional to the number of edits in the Levenshtein distance is really not the right thing to do at all.[null], [EMP-NEG] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEG | null | null | null | null |
8,822 | Under such an approach, the unedited string receives probability one, which doesn't leave much additional probability mass for the other candidates -- not to mention that the number of possible misspellings would require some aggressive normalization. [approach-NEU], [EMP-NEU] | approach | null | null | null | null | null | EMP | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,823 | Even under the assumption that a normalized edit probability is not particularly critical (an issue that was not raised at all in the paper, let alone assessed), the fact is that the assumptions of independent errors and a single substitution cost are grossly invalid in natural language.[null], [EMP-NEG] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEG | null | null | null | null |
8,824 | For example, the probability p_1 of 'pkoe' versus p_2 of 'zoze' as likely versions of 'poke' (as, say, the prefix of pokemon, as in your example) should be such that p_1 >>> p_2, not equal as they are in your model.[model-NEG], [EMP-NEG] | model | null | null | null | null | null | EMP | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,825 | Probabilistic models of string distance have been common since Ristad and Yianlios in the late 90s, and there are proper probabilistic models that would work with your same dynamic programming algorithm, as well as improved models with some modest state splitting.[models-NEU], [NOV-NEU] | models | null | null | null | null | null | NOV | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,826 | And even with very simple assumptions some unsupervised training could be used to yield at least a properly normalized model.[model-NEU], [EMP-NEU] | model | null | null | null | null | null | EMP | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,827 | It may very well end up that your very simple model does as well as a well estimated model, but that is something to establish in your paper, not assume.[model-NEG], [EMP-NEG] | model | null | null | null | null | null | EMP | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,828 | That such shortcomings are not noted in the paper is troublesome, particularly for a conference like ICLR that is focused on learned models, which this is not. [shortcomings-NEG], [APR-NEG] | shortcomings | null | null | null | null | null | APR | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,829 | As the primary contribution of the paper is this method for combining correction with completion, this shortcoming in the paper is pretty serious.[contribution-NEU, shortcoming-NEG], [EMP-NEG] | contribution | shortcoming | null | null | null | null | EMP | null | null | null | null | NEU | NEG | null | null | null | null | NEG | null | null | null | null |
8,830 | Some other comments: Your presentation of completion cost versus edit cost separation in section 3.3 is not particularly clear, partly since the methods are discussed prior to this point as extension of (possibly corrected) prefixes.[presentation-NEG, section-NEG], [PNF-NEG, EMP-NEG] | presentation | section | null | null | null | null | PNF | EMP | null | null | null | NEG | NEG | null | null | null | null | NEG | NEG | null | null | null |
8,831 | In fact, it seems that your completion model also includes extension of words with end point prior to the end of the prefix -- which doesn't match your prior notation, or, frankly, the way in which the experimental results are described.[experimental results-NEG], [EMP-NEG] | experimental results | null | null | null | null | null | EMP | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,832 | The notation that you use is a bit sloppy and not everything is introduced in a clear way.[notation-NEG], [PNF-NEG, CLA-NEG] | notation | null | null | null | null | null | PNF | CLA | null | null | null | NEG | null | null | null | null | null | NEG | NEG | null | null | null |
8,833 | For example, the s_0:m notation is introduced before indicating that s_i would be the symbol in the i_th position (which you use in section 3.3).[notation-NEG], [CLA-NEG] | notation | null | null | null | null | null | CLA | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,834 | Also, you claim that s_0 is the empty string, but isn't it more correct to model this symbol as the beginning of string symbol?[null], [EMP-NEG] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEG | null | null | null | null |
8,835 | If not, what is the difference between s_0:m and s_1:m?[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,836 | If s_0 is start of string, the s_0:m is of length m+1 not length m.[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,838 | (you don't need them, but also why number if you never refer to them later?[null], [PNF-NEU] | null | null | null | null | null | null | PNF | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,839 | ) Also the dynamic programming for Levenshtein is foundational, not required to present that algorithm in detail, unless there is something specific that you need to point out there (which your section 3.3 modification really doesn't require to make that point).[algorithm-NEG], [SUB-NEG] | algorithm | null | null | null | null | null | SUB | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,840 | Is there a specific use scenario for the prefix splitting, other than for the evaluation of unseen prefixes?[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,841 | This doesn't strike me as the most effective way to try to assess the seen/unseen distinction, since, as I understand the procedure, you will end up with very common prefixes alongside less common prefixes in your validation set, which doesn't really correspond to true 'unseen' scenarios.[null], [EMP-NEG] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEG | null | null | null | null |
8,843 | You never explicitly mention what your training loss is in section 5.1.[section-NEG], [CLA-NEG] | section | null | null | null | null | null | CLA | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,844 | Overall, while this is an interesting and important problem, and the engineering details are interesting and reasonably well-motivated, the main contribution of the paper is based on a pretty flawed approach to modeling correction probability, which would limit the ultimate applicability of the methods.[problem-POS, main contribution-NEG], [EMP-NEG] | problem | main contribution | null | null | null | null | EMP | null | null | null | null | POS | NEG | null | null | null | null | NEG | null | null | null | null |
8,850 | The paper is well explained, and it's also nice that the runtime is shown for each of the algorithm blocks.[paper-POS], [CLA-POS, EMP-POS] | paper | null | null | null | null | null | CLA | EMP | null | null | null | POS | null | null | null | null | null | POS | POS | null | null | null |
8,851 | Could imagine this work giving nice guidelines for others who also want to run query completion using neural networks.[work-POS], [IMP-POS] | work | null | null | null | null | null | IMP | null | null | null | null | POS | null | null | null | null | null | POS | null | null | null | null |
8,852 | The final dataset is also a good size (36M search queries).[dataset-POS], [SUB-POS] | dataset | null | null | null | null | null | SUB | null | null | null | null | POS | null | null | null | null | null | POS | null | null | null | null |
8,853 | My major concerns are perhaps the fit of the paper for ICLR as well as the thoroughness of the final experiments.[experiments-NEU], [APR-NEU] | experiments | null | null | null | null | null | APR | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,854 | Much of the paper provides background on LSTMs and edit distance, which granted, are helpful for explaining the ideas.[null], [EMP-POS] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | POS | null | null | null | null |
8,855 | But much of the realtime completion section is also standard practice, e.g. maintaining previous hidden states and grouping together the different gates.[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,856 | So the paper feels directed to an audience with less background in neural net LMs.[null], [IMP-NEG] | null | null | null | null | null | null | IMP | null | null | null | null | null | null | null | null | null | null | NEG | null | null | null | null |
8,857 | Secondly, the experiments could have more thorough/stronger baselines.[experiments-NEU, baselines-NEU], [EMP-NEU, CMP-NEU] | experiments | baselines | null | null | null | null | EMP | CMP | null | null | null | NEU | NEU | null | null | null | null | NEU | NEU | null | null | null |
8,858 | I don't really see why we would try stochastic search. And expected to see more analysis of how performance was impacted as the number of errors increased, even if errors were introduced artificially, and expected analysis of how different systems scale with varying amounts of data.[analysis-NEU], [EMP-NEG] | analysis | null | null | null | null | null | EMP | null | null | null | null | NEU | null | null | null | null | null | NEG | null | null | null | null |
8,859 | The fact that 256 hidden dimension worked best while 512 overfit was also surprising, as character language models on datasets such as Penn Treebank with only 1 million words use hidden states far larger than that for 2 layers.[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,864 | The experiments show robustness to these types of noise.[experiments-POS], [EMP-NEU] | experiments | null | null | null | null | null | EMP | null | null | null | null | POS | null | null | null | null | null | NEU | null | null | null | null |
8,865 | Review: The claim made by the paper is overly general, and in my own experience incorrect when considering real-world-noise.[claim-NEG], [EMP-NEG] | claim | null | null | null | null | null | EMP | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,866 | This is supported by the literature on data cleaning (partially by the authors), a procedure which is widely acknowledged as critical for good object recognition.[literature-NEU, procedure-NEU], [EMP-NEU] | literature | procedure | null | null | null | null | EMP | null | null | null | null | NEU | NEU | null | null | null | null | NEU | null | null | null | null |
8,867 | While it is true that some image-independent label noise can be alleviated in some datasets, incorrect labels in real world datasets can substantially harm classification accuracy.[datasets-NEU, accuracy-NEU], [EMP-NEG] | datasets | accuracy | null | null | null | null | EMP | null | null | null | null | NEU | NEU | null | null | null | null | NEG | null | null | null | null |
8,868 | It would be interesting to understand the source of the difference between the results in this paper and the more common results (where label noise damages recognition quality).[results-NEU], [EMP-NEU, CMP-NEU] | results | null | null | null | null | null | EMP | CMP | null | null | null | NEU | null | null | null | null | null | NEU | NEU | null | null | null |
8,869 | The paper did not get a chance to test these differences, and I can only raise a few hypotheses.[paper-NEG], [CMP-NEG] | paper | null | null | null | null | null | CMP | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,870 | First, real-world noise depends on the image and classes in a more structured way. For instance, raters may confuse one bird species from a similar one, when the bird is photographed from a particular angle.[null], [CLA-NEG] | null | null | null | null | null | null | CLA | null | null | null | null | null | null | null | null | null | null | NEG | null | null | null | null |
8,872 | Another possible reason is that classes in MNIST and CIFAR10 are already very distinctive, so are more robust to noise.[null], [EMP-POS] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | POS | null | null | null | null |
8,873 | Once again, it would be interesting for the paper to study why they achieve robustness to noise while the effect does not hold in general.[paper-NEU], [SUB-NEU] | paper | null | null | null | null | null | SUB | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,874 | Without such an analysis, I feel the paper should not be accepted to ICLR because the way it states its claim may mislead readers.[analysis-NEG, paper-NEG], [SUB-NEG, APR-NEG] | analysis | paper | null | null | null | null | SUB | APR | null | null | null | NEG | NEG | null | null | null | null | NEG | NEG | null | null | null |
8,875 | Other specific comments: -- Section 3.4 the experimental setup, should clearly state details of the optimization, architecture and hyper parameter search.[Section-NEG, architecture-NEU], [EMP-NEU, CLA-NEG] | Section | architecture | null | null | null | null | EMP | CLA | null | null | null | NEG | NEU | null | null | null | null | NEU | NEG | null | null | null |
8,876 | For example, for Conv4, how many channels at each layer?[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,877 | how was the net initialized? [null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,878 | which hyper parameters were tuned and with which values?[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,879 | were hyper parameters tuned on a separate validation set?[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,880 | How was the train/val/test split done, etc.[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,882 | -- Section 4, importance of large datasets.[Section-NEU], [EMP-POS] | Section | null | null | null | null | null | EMP | null | null | null | null | NEU | null | null | null | null | null | POS | null | null | null | null |
8,883 | The recent paper by Chen et al (2017) would be relevant here.[null], [SUB-NEU] | null | null | null | null | null | null | SUB | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,884 | -- Figure 8 failed to show for me.[Figure-NEG], [PNF-NEG] | Figure | null | null | null | null | null | PNF | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,885 | -- Figure 9,10, need to specify which noise model was used. [Figure-NEG, model-NEU], [EMP-NEG] | Figure | model | null | null | null | null | EMP | null | null | null | null | NEG | NEU | null | null | null | null | NEG | null | null | null | null |
8,890 | Naive multitask learning with deep neural networks fails in many practical cases, as covered in the paper. [paper-NEU], [EMP-NEG] | paper | null | null | null | null | null | EMP | null | null | null | null | NEU | null | null | null | null | null | NEG | null | null | null | null |
8,891 | The one concern I have is perhaps the choice of distinct of Atari games to multitask learn may be almost adversarial, since naive multitask learning struggles in this case; but in practice, the observed interference can appear even with less visually diverse inputs.[null], [EMP-NEG] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEG | null | null | null | null |
8,892 | Although performance is still reduced compared to single task learning in some cases,[performance-NEG], [EMP-NEG] | performance | null | null | null | null | null | EMP | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,893 | this paper delivers an important reference point for future work towards achieving generalist agents, which master diverse tasks and represent complementary behaviours compactly at scale.[reference-POS, future work-POS], [IMP-POS] | reference | future work | null | null | null | null | IMP | null | null | null | null | POS | POS | null | null | null | null | POS | null | null | null | null |
8,894 | I wonder how efficient the approach would be on DM lab tasks, which have much more similar visual inputs, but optimal behaviours are still distinct. [approach-NEU], [IMP-NEU] | approach | null | null | null | null | null | IMP | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,900 | ** REVIEW SUMMARY ** The paper reads well, has sufficient reference.[paper-POS], [CLA-POS] | paper | null | null | null | null | null | CLA | null | null | null | null | POS | null | null | null | null | null | POS | null | null | null | null |
8,901 | The idea is simple and well explained.[idea-POS], [EMP-POS] | idea | null | null | null | null | null | EMP | null | null | null | null | POS | null | null | null | null | null | POS | null | null | null | null |
8,902 | Positive empirial results support the proposed regularizer.[empirial results-POS], [EMP-POS] | empirial results | null | null | null | null | null | EMP | null | null | null | null | POS | null | null | null | null | null | POS | null | null | null | null |
8,905 | In related work, I would cite co-training approaches.[related work-NEU], [CMP-NEU, SUB-NEU] | related work | null | null | null | null | null | CMP | SUB | null | null | null | NEU | null | null | null | null | null | NEU | NEU | null | null | null |
8,906 | In effect, you have two view of a point in time, its past and its future and you force these two views to agree, see (Blum and Mitchell, 1998) or Xu, Chang, Dacheng Tao, and Chao Xu.[null], [CMP-NEU] | null | null | null | null | null | null | CMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,907 | A survey on multi-view learning. arXiv preprint arXiv:1304.5634 (2013).[null], [CMP-NEU] | null | null | null | null | null | null | CMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,908 | I would also relate your work to distillation/model compression which tries to get one network to behave like another. On that point, is it important to train the forward and backward network jointly or could the backward network be pre-trained?[work-NEU], [CMP-NEU, EMP-NEU] | work | null | null | null | null | null | CMP | EMP | null | null | null | NEU | null | null | null | null | null | NEU | NEU | null | null | null |
8,909 | In section 2, it is not obvious to me that the regularizer (4) would not be ignored in absence of regularization on the output matrix.[section-NEU], [EMP-NEG] | section | null | null | null | null | null | EMP | null | null | null | null | NEU | null | null | null | null | null | NEG | null | null | null | null |
8,910 | I mean, the regularizer could push h^b to small norm, compensating with higher norm for the output word embeddings.[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,911 | Could you comment why this would not happen?[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,912 | In Section 4.2, you need to refer to Table 2 in the text.[Section-NEU, Table-NEU, text-NEU], [PNF-NEU] | Section | Table | text | null | null | null | PNF | null | null | null | null | NEU | NEU | NEU | null | null | null | NEU | null | null | null | null |
8,913 | You also need to define the evaluation metrics used.[evaluation metrics-NEU], [EMP-NEU] | evaluation metrics | null | null | null | null | null | EMP | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,914 | In this section, why are you not reporting the results from the original Show&Tell paper?[section-NEU], [EMP-NEU] | section | null | null | null | null | null | EMP | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,915 | How does your implementation compare to the original work?[implementation-NEU], [CMP-NEU] | implementation | null | null | null | null | null | CMP | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,916 | On unconditional generation, your hypothesis on uncertainty is interesting and could be tested.[hypothesis-POS], [EMP-POS] | hypothesis | null | null | null | null | null | EMP | null | null | null | null | POS | null | null | null | null | null | POS | null | null | null | null |
8,917 | You could inject uncertainty in the captioning task for instance, e.g. consider that multiple version of each word e.g. dogA, dogB, docC which are alternatively used instead of dog with predefined substitution rates.[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,918 | Would your regularizer still be helpful there?[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,919 | At which point would it break?[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,923 | I think the fact that the authors demonstrate the viability of training VDFFNWSCs that could have, in principle, arbitrary nonlinearities and normalization layers, is somewhat valuable and as such I would generally be inclined towards acceptance,[acceptance-POS], [REC-POS] | acceptance | null | null | null | null | null | REC | null | null | null | null | POS | null | null | null | null | null | POS | null | null | null | null |
8,924 | even though the potential impact of this paper is limited because the training strategy proposed is (by deep learning standards) relatively complicated, requires tuning two additional hyperparameters in the initial value of lambda as well as the step size for updating lambda, and seems to have no significant advantage over just using skip connections throughout training.[potential impact-NEG, strategy-NEG], [IMP-NEG] | potential impact | strategy | null | null | null | null | IMP | null | null | null | null | NEG | NEG | null | null | null | null | NEG | null | null | null | null |
8,925 | So my rating based on the message of the paper would be 6/10. [rating-NEU], [REC-NEU] | rating | null | null | null | null | null | REC | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,927 | As long as those issues remain unresolved, my rating is at is but if those issues were resolved it could go up to a 6.[rating-NEU], [REC-NEU] | rating | null | null | null | null | null | REC | null | null | null | null | NEU | null | null | null | null | null | NEU | null | null | null | null |
8,928 | +++ Section 3.1 problems +++ - I think the toy example presented in section 3.1 is more confusing than it is helpful because the skip connection you introduce in the toy example is different from the skip connection you introduce in VANs.[section-NEG], [EMP-NEG] | section | null | null | null | null | null | EMP | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,929 | In the toy example, you add (1 - alpha)wx whereas in the VANs you add (1 - alpha)x.[example-NEG], [EMP-NEG] | example | null | null | null | null | null | EMP | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,930 | Therefore, the type of vanishing gradient that is observed when tanh saturates, which you combat in the toy model, is not actually combated at all in the VAN model.[null], [EMP-NEG] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEG | null | null | null | null |
8,931 | While it is true that skip connections combat vanishing gradients in certain situations, your example does not capture how this is achieved in VANs.[example-NEG], [EMP-NEG] | example | null | null | null | null | null | EMP | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,932 | - The toy example seems to be an example where Lagrangian relaxation fails, not where it succeeds.[null], [EMP-NEG] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEG | null | null | null | null |
8,933 | Looking at figure 1, it appears that you start out with some alpha < 1 but then immediately alpha converges to 1, i.e. the skip connection is eliminated early in training, because wx is further away from y than tanh(wx).[figure-NEG], [EMP-NEG] | figure | null | null | null | null | null | EMP | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,934 | Most of the training takes place without the skip connection.[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,935 | In fact, after 10^4 iterations, training with and without skip connection seem to achieve the same error.[null], [EMP-NEG] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEG | null | null | null | null |
8,936 | It appears that introducing the skip connection was next to useless and the model failed to recognize the usefulness of the skip connection early in training.[null], [EMP-NEG] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEG | null | null | null | null |
8,937 | - Regarding the optimization algorithm involving alpha^* at the end of section 3: It looks to me like a hacky, unprincipled method with no guarantees that just happened to work in the particular example you studied.[section-NEG], [EMP-NEG] | section | null | null | null | null | null | EMP | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,938 | You motivate the choice of alpha^* by wanting to maximize the reduction in the local linear approximation to mathcal{C} induced by the update on w.[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,939 | However, this reduction grows to infinity the larger the update is.[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,940 | Does that mean that larger updates are always better?[null], [EMP-NEU] | null | null | null | null | null | null | EMP | null | null | null | null | null | null | null | null | null | null | NEU | null | null | null | null |
8,942 | If we wanted to reduce the size of the objective according to the local linear approximation, why wouldn't we choose infinitely large step sizes?[approximation-NEG], [EMP-NEG] | approximation | null | null | null | null | null | EMP | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,943 | Hence, the motivation for the algorithm you present is invalid.[motivation-NEG], [EMP-NEG] | motivation | null | null | null | null | null | EMP | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |
8,944 | Here is an example where this algorithm fails: consider the point (x,y,w,alpha,lambda) (100, sigma(100), 1.0001, 1, 1).[example-NEG], [EMP-NEG] | example | null | null | null | null | null | EMP | null | null | null | null | NEG | null | null | null | null | null | NEG | null | null | null | null |