dedup-isc-ft-v107-score
float64 0.32
1
| uid
stringlengths 32
32
| text
stringlengths 0
32.5k
| paper_id
stringlengths 1
14
| original_image_filename
stringlengths 5
222
|
|---|---|---|---|---|
0.518286 | c45b26f95afa4cfdbe1fe2305f83eebd | (a) Squeezing of the mechanical quadrature fluctuations ⟨ΔQ̂^2 ⟩ (in decibel) as a function of the sideband parameter κ/ω_m, for different value of the coupling g: g=0.01ω_m (red line), g=0.05ω_m (yellow line), g=0.3ω_m (cyan line). The other values for the parameters are: γ=10^-4ω_m, η=1, n̅=10. Solid lines represent the fluctuations ⟨ΔQ̂^2⟩_ fb of the unconditional state obtained via Markovian feedback through the Hamiltonian ℋ̂_ fb = ξ_mP̂ I_Y. Dashed curves represent the fluctuations ⟨ΔQ̂^2⟩_c of the conditional states [or equivalently of the unconditional state obtained via the optimal feedback Hamiltonian in Eq. (<ref>)]. The shaded area marks the excluded region beyond the threshold value Eq. (<ref>). Dashed black lines are the predictions of the adiabatic theory Eq. (<ref>). (b) Effects of measurement-plus-feedback for an ideal BAE measurement. Black arrows describe how quadratures influence each other. By measuring the output phase quadrature, information is extracted from Q̂ (outgoing arrow), while no information can be extracted from P̂ (incoming arrow) where all the backaction goes. Ideal Markovian feedback drives both Q̂ and Ŷ. (c) Markovian feedback restricted to the mechanical mode (mechanical-limited feedback) and including non-QND terms; these open new paths (dashed arrows) where both backaction and conditioning can spread. The mechanical mode is also subject to (unmonitored) noise from a thermal environment (Q̂_in and P̂_in), which for convenience is not shown | 2012.15221 | Fig1New.jpg |
0.45025 | d8f961e832f94066855631267f12d0e2 | Top views of the lowest energy structures of C_18 supported on Ir(111). Carbon and iridium atoms are shown in grey and blue colors, respectively. The brightness of the colors follow the coordinate perpendicular to the surface. Side views are further shown for structures 0 and 4. | 2012.15222 | C18.jpg |
0.400806 | 1a2d5c286279487d964424d8fc013b29 | The anatase TiO_2(001)-(1×4) surface reconstruction, featuring rows of protruding titanium atoms. In the structure search context we consider two versions of the problem, having either 2 or 3 atomic layers optimized on top of a fixed bulk layer. One unit cell of the global minimum for these problems, labeled the 2-layer and 3-layer TiO_2(001)-(1×4) problems respectively, are shown in the figure. For the 2 layer problem 27 atoms are optimized, whereas 39 atoms are optimized in the 3 layer problem. | 2012.15222 | TiO2_structures.jpg |
0.423151 | 80635f0c3e9c4f9395b9d188148d07ca | One dimensional GPR example, comparing the normal Gaussian kernel (a,b,c) with the double Gaussian kernel (d,e,f) used in GOFEE. The target function are represented by the thick green line, the GPR model by the black line and the training data by red dots. The thin lines represent functions sampled from the model distributions. a and d) Prior to training, both models just predict the mean of the data, however functions sampled for the two kernels are visually dissimilar, with the double Gaussian kernel resulting in short scale wiggles on top of the long scale variations. b) and e) For the trained models, the single Gaussian model tends to overshoot the target, and underestimate short scale uncertainties Even with only a weight of β=0.01, the short scale Gaussian in the double Gaussian model remedies these flaws to a large extent. c) and f) Adding further data, as orange dots, requiring even more short scale variation, only reinforces this point. | 2012.15222 | covariance.jpg |
0.438303 | bea3a3afa471403eb5e30c0266bcfc40 | Workflow in GOFEE. (i) Random initial structures are generated, evaluated with DFT and added to a structural database. (ii) The surrogate model is trained based on the database (iii) The population is initialized/updated. (iv) A number, N_c, of new candidate structures are generated by applying stochastic changes to the population. (v) All N_c candidates are relaxed in the acquisition function. (vi) The single most promising candidate according to the acquisition function is selected. (vii) A single evaluation, using the target energy expression, is performed for this structure, along with a second evaluation for the structure perturbed slightly along the force. The search is carried out by adding these two newly evaluated structures to the training database and repeating steps (ii)-(vii) N times in total. | 2012.15222 | flow_diagram.jpg |
0.471416 | 3e00c31d531d46108ee8107d9edb4a23 | The population scheme, visualized with principal component analysis (PCA), with the first two principal components capturing 83% of the variance. a) First, all evaluated structures, with an energy above Δ E_pop of the best structure found so far, are ignored, as represented by black circles. b) Secondly, The remaining structures are clustered in the feature space using the k-means++ algorithm. The resulting clusters are colored, and cluster centers drawn as black circles. Thirdly, the population is created by selecting the lowest energy structure in each cluster, drawn as black crosses. c) The population resulting from applying the scheme to this dataset, consisting of the 500 first structures of a GOFEE search. As the population structures can lie on cluster edges, some could in principle be very similar. The diversity of the population is however not found to suffer greatly from this effect, as exemplified in c). The adequacy of the PCA visualization is also visible in the fairly reasonable positions of the cluster centers, and only slight smearing of the cluster borders. | 2012.15222 | population.jpg |
0.418208 | e21d115d58394603ba071a2d435fe587 | Benchmark of GOFEE, on the 3-layer TiO_2(001)-(1×4) problem. Blue curve: GOFEE is used as presented in this paper, with dual-point, the double Gaussian kernel, the same κ_relax=κ=2 used for relaxation and candidate selection and with the repulsive prior. Green curve: uses the normal Gaussian covariance. Orange curve: Omitting the repulsive prior mean function. | 2012.15222 | success_bench_cov_prior.jpg |
0.46636 | 92d702794482413182aca834e8c8e492 | Benchmark of GOFEE, on the 3-layer TiO_2(001)-(1×4) problem. Blue curve: GOFEE as presented in this paper. Red curve: The dual-point evaluation is ommitted in the search. | 2012.15222 | success_bench_dualpoint.jpg |
0.47533 | 0df8c87fcb7640eca9a1757622cb0376 | Benchmark of GOFEE, on the 3-layer TiO_2(001)-(1×4) problem. Blue curve: GOFEE as presented in this paper. Purple curve: Relaxations are carried out with κ=0 in the acquisition function. | 2012.15222 | success_bench_kappa_relax.jpg |
0.522051 | 8a831422d1364afda63f124de7a3a584 | The geometries of the four lowest energy structures of C_24, along with success curves for finding each structure, based on 50 independent restarts of GOFEE. The colored success curves represent searches run with the new population and improved local convergence schemes, whereas the gray curves represent searches without these additions. Structures 1 and 3 are found in almost 100 of searches after only 1500 DFT evaluations, whereas the global minimum and structure 2 is found in ≈80 of searches after 3000 DFT evaluations. These might be more difficult because they both feature two four membered rings, which are typically not energetically favored compared to hexagonal and pentagonal rings. In comparison, structures 1 and 3 contain only one and none, respectively. | 2012.15222 | success_c24.jpg |
0.414925 | c286dc77d18046df8cfec2fc4a44cd5b | Importance of the new candidate structures. a) Success curves for the 2-layer TiO_2(001)-(1×4) problem, resulting from searches with different numbers, N_c, of new candidates generated in each search iteration. The gain in performance, from e.g. quadrupling N_c, is remarkable, especially in the range of relatively small N_c. b) The full, green line shows the success curve for the 3-layer TiO_2(001)-(1×4) problem. The dashed, dark blue line shows the fraction of the searches, that, based on the acquisition function, would select the global minimum, were it supplied as one of the new candidates. The remaining dashed lines only counts this artificial selection of the global minimum successful if the predicted uncertainty, σ_sur(x_GM), thereon, is less than the various values given in the figure. This supplies information about how certain the model is about the GM, when it is selected over the other candidates. | 2012.15222 | success_candidates.jpg |
0.481806 | 4470427c04304cf89b98b2941e204a02 | The effect κ on the performance on searches on the 2-layer TiO_2(001)-(1×4) system. a) success curves, each based on 75 independent searches, for different values of κ, with values increasing when going from blue to red. b) Snapshots of the success rates for each curve, in a), N_SP=200, 400 and 800 single-point evaluations into the search. | 2012.15222 | success_kappa.jpg |
0.407602 | f2a77693bccf47e79add89ba7d6eccd0 | Sketch of the TPR scheme, with N_GED=3 for simplicity. First, the predicted uncertainty of the selected structure is compared to the threshold value σ_TPR, to determine, whether to continue with the TPR scheme. If moving on with the scheme, GED steps are taken, in an attempt to reach the lowest energy structure among the previously evaluated structures within the basin. Each GED step is carried out by moving to the nearest structure, among the N_GED closest, which has a lower energy than the present one. When no lower energy neighbors are found, a TPR trajectory is initialized by carrying out two relaxation steps in the target potential, from the final structure. Alternatively, if the GED steps at some point reaches a structure which is part of an ongoing TPR trajectory, this relaxation is continued. | 2012.15222 | tpr.jpg |
0.420937 | 7df7a07b8296452dbc6e98559460f661 | Transverse profiles for (a) the characteristic speeds and (b) primitive variables of a photospheric waveguide representative of pores and sunspot umbrae. These profiles are generated by specifying the transverse distributions of the adiabatic sound (c_ s) and cusp (c_ T) speeds, for which a continuous transition layer (TL) connects a uniform interior to a uniform exterior (see Equation <ref>). This TL is located between r_ i = R-l/2 and r_ e = R+l/2, where R is the mean waveguide radius and l the TL width. For illustration purposes, here l/R is arbitrarily chosen to be 0.2. | 2012.15426 | Fig01.jpg |
0.473435 | a96565711dfd4fffa608e3226a842cea | Dependence on the magnetic Reynolds number (R_ m) of the dispersive properties of slow surface kink modes in a photospheric waveguide. Plotted are (a) the damping-time-to-period ratio τ/P, (b) the axial phase speed /k, and (c) the damping rate in units of k. A number of combinations of the transition layer width (l/R) and axial wavenumber (kR) are examined as labeled. The area shaded yellow corresponds to the range of R_ m that is likely to be relevant for pores and sunspot umbrae. See the text for details. | 2012.15426 | Fig02.jpg |
0.479557 | 0fb1f973a1fb4353b7a8b65130f6484b | Dependence on the magnetic Reynolds number (R_ m) of the terms involved in the energy balance (see Equation <ref> and the associated definitions). A number of combinations of the transition layer width (l/R) and axial wavenumber (kR) are examined as labeled. The area shaded yellow corresponds to the range of R_ m that is likely to be relevant for pores and sunspot umbrae. The diamonds represent -2, taken from a small subset of the values presented in Figure <ref>c. See the text for details. | 2012.15426 | Fig03.jpg |
0.417767 | f8a753fc791b4b07866984546b8c8a5c | Radial distributions of the s̅_ res term for a number of representative values of the magnetic Reynolds number (R_ m) at a given pair of [l/R, kR] = [0.1, 0.7]. Here s̅_ res is evaluated by applying the bar operation (Equation <ref>) to the definition of (Equation <ref>). The shaded area represents where the equilibrium quantities are nonuniform. See the text for details. | 2012.15426 | Fig04.jpg |
0.512124 | f733af380efd40bba5b51294dc06ed84 | Dispersive properties of resonantly damped slow surface kink modes, shown by the dependence on the dimensionless layer width (l/R) of (a) the damping-time-to-period ratios (τ/P), (b) the axial phase speeds (/k), and (c) the damping rates () in units of k. A number of axial wavenumbers (kR) are examined as represented by the different colors. The results from self-consistent resistive computations are given by the solid curves. For each pair of [l/R, kR], the explicit dispersion relation (Equation <ref>) in the thin-boundary limit is also solved. The dashed curves represent the computations where both the cusp and resonances are accounted for, whereas the dotted curves take into account only the cusp resonance. | 2012.15426 | Fig05.jpg |
0.482347 | d3c2d66ad57546108947eb94a1b34b66 | Radial distributions of the real part of the Fourier amplitude of the radial flow speed (Reṽ_r) in the intervals embracing (a) the cusp, and (b) the dissipative layers. A number of magnetic Reynolds numbers () are examined as labeled, whereas [l/R, kR] is fixed at [0.2, 2]. The eigenfunctions are rescaled such that = 1 at r = 1.25 R. | 2012.15426 | Fig06.jpg |
0.453691 | 6a557fe161634ed8a5c1c89804af6cd7 | Dependence on the transition layer width (l/R) of the ratios of the hatted terms in Equation (<ref>). The axial wavenumber is fixed at kR = 2. For a given l/R, a number of magnetic Reynolds numbers () are examined to evaluate the hatted terms as represented by the symbols of different colors. With solid curves we connect the symbols representing , , and Ŝ_ ideal for the same but different l/R. The dashed curve represents -2, the values of which are direct outputs of our code. See the text for details. | 2012.15426 | Fig07.jpg |
0.469889 | 645e6cb561054cc4abc3b6f1f62fc99c | Radial profiles of some relevant eigen-functions found in self-consistent resistive computations for resonantly damped slow surface kink modes. Two different values for the axial wavenumber are examined, one corresponding to kR = 0.3 (the left column) and the other to kR = 2 (right). The vertical dash-dotted lines mark the boundaries of the transition layer, which corresponds to a layer-width-to-radius-ratio of l/R = 0.2. The black (red) curves correspond to the real (imaginary) parts. The top row pertains to the Eulerian perturbation to total pressure (), while the rest correspond to the radial (ṽ_r), azimuthal (ṽ_θ), and axial (ṽ_z) speeds. The eigen-functions are normalized such that =1 where its modulus attains maximum. Different values for the magnetic Reynolds number () are adopted in the two columns, see the text for details. | 2012.15426 | Fig08.jpg |
0.498792 | d1fe3a33d3a84de28c09933433bf2ced | Similar to Figure <ref> except that the dependence on the axial wavenumber (kR) is examined. Here the dimensionless transition layer width is fixed at l/R = 0.2. | 2012.15426 | Fig09.jpg |
0.451325 | 18ccc757aeaa4b4a886611bd0ecec432 | Dependence on the axial wavenumber (k) of the optimal damping-time-to-period ratio (τ/P)_min and the transition-layer-width-to-radius ratio (l/R)_min where the optimal damping is reached. The values for (τ/P)_min and (l/R)_min are derived from self-consistent resistive computations. At an (l/R)_min thus derived, the explicit dispersion relation in the TB limit (Equation <ref>) is also solved, yielding (τ/P)^ TB for comparison. See the text for details. | 2012.15426 | Fig10.jpg |
0.499207 | 2a47459a6c8549d3945628e1a28dc71a | Comparsion of 3-qubit systems using different methods for benchmark case. | 2012.15427 | 1Statetransfer_level8_bench_BF.jpg |
0.469321 | c57c245f43034a80a7d411788bfe5969 | Best fidelity under different time durations. | 2012.15427 | 4Opencontrol_level4_time.jpg |
0.456735 | a4b13d48fcd8434a90fc7b2982b8d423 | Illustration of the curriculum learning process with a flow of tasks. | 2012.15427 | Fixedtasks.jpg |
0.472841 | 0d7ee0e3910446a6ad8fc4062ac86599 | Framework of CDRL approach for quantum control. (a) Curriculum construction and management. The curriculum agent generates a task for training the RL agent and utilizes the performance of the RL agent to determine whether a stop criterion is satisfied. Once the previous task has been achieved, knowledge is transferred between different tasks by reusing past samples. Meanwhile, measures are taken to excite the RL agent before the subsequent task is scheduled. (b) Train the RL agent for one task. At each time step t, the RL agent observes its current state s_t (step 1) and suggests a control action a_t (step 2), which is mapped to the control fields {u_m(t)} (step 3). Then, the quantum system takes the proposed control strategy and obtains the next state s_t, with fidelity F_t and reward signal r_t (step 4). The transition e_t=(s_t,a_t,r_t,s_t+1,F_t) is put into a large memory buffer (step 5). Finally, a batch of transitions is selected from the buffer and then fed into the networks to update its weights or parameters (step 6). | 2012.15427 | overall_framework.jpg |
0.473088 | 46bc940845004fb9915b8642607ea873 | Illustration of transferring knowledge between two tasks. | 2012.15427 | transferknowledge.jpg |
0.435889 | 9dad86e5c2d94baf855814edf7b3a4fb | The mass difference between extremal hairy dyonic and magnetic black holes for different electric charge q. The upper charge limit q_ max≃72 or 𝒪(2π/e^2) is obtained by requiring an exponentially decreasing behavior for f(r) at large r. | 2012.15430 | dyonmasscharge.jpg |
0.41666 | 7d94eba8eb2b45f2a5f52ae909756432 | The black hole masses as functions of horizon radii for both the hairy and RN magnetic black holes. The left endpoints of the two curves correspond to the extremal black hole cases. | 2012.15430 | masscomp.jpg |
0.398719 | f92ddcc673af41d69bd49f713cf7faf5 | RF-Visual Grasping. Using the RFID's location, crops out the RGB-D data near the tagged object location, and feeds the cropped RGB-D data and RFID location probability as input to the Deep-NN that outputs the Q-value estimations. The rotated inputs in 16 different directions are separately fed to DNN to estimate Q-values of different grasping rotations. | 2012.15436 | Net_new.jpg |
0.437978 | 56e1509ea0fc4f25a68a64a3e58992ca | System Overview. uses RGB-D (green arrows) to create a 3D model of the environment and extract occluded regions, and it plugs the location of the RFID-tagged object (red arrows) into it. If the object is behind obstacles, it performs RF-guided active exploration and trajectory optimization to obtain a better view of the object by maneuvering around obstacles. Once a stopping criterion is met, it proceeds to grasping. Here, it applies RF-based attention to the RGB-D information and relies on a model-free deep-reinforcement network to discover optimal grasping affordances. uses the RFID's locationto close the loop on the grasping task and determine whether it has been successful or whether it needs to make another grasping attempt. | 2012.15436 | overview_final.jpg |
0.40593 | c9ccf4865fb445afb28ca2af088dcce6 | RFID-based Perception. uses a customized reader that can identify and accurately localize RFID-tagged objects through occlusions. | 2012.15436 | rfid.jpg |
0.483923 | 5699a4fd44834a26a506d0e3fc8736d2 | Citrus tree | 2012.15438 | CITRUS-macrobench.jpg |
0.458307 | bcb5e16272aa444ea51840fd55885383 | Skip list | 2012.15438 | SKIPLISTLOCK-macrobench.jpg |
0.511715 | 5bff950db36b4a72b07489317a280e71 | An example of using bundled references in a linked list. The path made of solid lines represents the state of the linked list after all update operations take place. Edges are labeled with their respective timestamps. | 2012.15438 | bundle-example.jpg |
0.471631 | cc0cdc716000463eb5a70a5a448db44e | CT, 0-90-10 | 2012.15438 | citrus0.jpg |
0.47854 | 4fab5590478648bc9b6ee3ea76e1780c | CT, 10-80-10 | 2012.15438 | citrus10.jpg |
0.474962 | 363d6f3e6bc7476b8bd224082d5ef1a9 | CT, 2-88-10 | 2012.15438 | citrus2.jpg |
0.468797 | cd6d63a9efd547ab9b7446a3b2b9bb22 | CT, 50-40-10 | 2012.15438 | citrus50.jpg |
0.461442 | 5e342d88f060459aade9c34296009943 | CT, 90-0-10 | 2012.15438 | citrus90.jpg |
0.449809 | b283f99c07fb4e36a067fd39206b995d | SL, 0-90-10 | 2012.15438 | skiplistlock0.jpg |
0.441825 | 0139cb4f016e45ea937985567bc92e10 | SL, 10-80-10 | 2012.15438 | skiplistlock10.jpg |
0.495103 | ef95fd0b7e874dae95e1cc59a3a1b3be | SL, 2-88-10 | 2012.15438 | skiplistlock2.jpg |
0.473667 | 865c339dfc024145929edb5990becdac | SL, 50-40-10 | 2012.15438 | skiplistlock50.jpg |
0.465196 | 4effc561436c4cf5979bd470a5e59c07 | SL, 90-0-10 | 2012.15438 | skiplistlock90.jpg |
0.397622 | e899aee5e5a24be5af99d7a91bd61d62 | Framework of SID method based on CenterNet <cit.>. | 2012.15439 | CenterNet.jpg |
0.415487 | 670f35f57c0b4184a1cfbbb9c50bd489 | Framework of SID method based on FCOS <cit.>. | 2012.15439 | FCOS.jpg |
0.436289 | c2723601ca83459787e6b2da06a5cf1a | An example of incremental object detection. The challenge is to retrain the model to maintain the detection of the old classes whilst learning the new classes. During each incremental step, only new class data and annotations for new class objects are provided. | 2012.15439 | incremental_learning.jpg |
0.47899 | 4213f8d461674d7bb2c4a50cde444db9 | The top graph shows the prediction accuracy of training and test sets for 400 epochs, whereas the bottom graph indicates the loss for DeepTake on prediction of three classes of low-, mid-, and high- takeover time. | 2012.15441 | AccuracyVSLoss.jpg |
0.394044 | 73e2c4258f8345b9bb7eb8fde0d24a42 | The left figure shows the accuracy result of training and test sets for 400 epochs whereas the right figure indicates the loss for our DNN model on prediction of three class driver's reaction time. | 2012.15441 | Accuracy_vs_Loss.jpg |
0.363041 | f7dda7fead3c430e8a0f48b77cacd22e | Confusion matrix for the prediction of takeover behavior.The results are averaged over 10 fold cross validation splits. (a) Binary class takeover intention takeover(TK) vs. Not Takeover(NTK), (b) 3-Class classification results of takeover time, (c) 3-class classification of takeover quality. | 2012.15441 | CM_ALL.jpg |
0.387405 | aa8dc664bfc2438db51612e6d782f4d5 | Confusion matrix for the prediction of five classes of driver takeover time. | 2012.15441 | CM_FiveCL_RT.jpg |
0.487049 | 85bc5820a5344810938ba01939cd1562 | Confusion matrix for the prediction of three classes of driver's takeover time. | 2012.15441 | CM_ThreeCL_RT.jpg |
0.421308 | a567aec1f3334658b66861f77c9be2e6 | DeepTake uses data from multiple sources (pre-driving survey, vehicle data, non-driving related tasks (NDRTs) information, and driver biometrics) and feeds the preprocessed extracted features into deep neural network models for the prediction of takeover intention, time and quality. | 2012.15441 | DeepTake-overview-cropped.jpg |
0.416828 | 8b122d7f33e74ad580d69afb360186b4 | DeepTake uses data from multiple sources (pre-driving survey, vehicle data, non-driving related tasks (NDRTs) information, and driver biometrics) and feeds the preprocessed extracted features into deep neural network models for the prediction of takeover intention, time and quality. | 2012.15441 | DeepTake-overview-revised.jpg |
0.417187 | f17e2cfb129342a78803fb9fe0a7e37e | A schematic view of an example of a takeover situation used in our study, consisting of: 1) takeover timeline associated with participantsâ course of action; 2) system status; and 3) takeover situation. The vehicle was driven in the automated mode to the point after the TOR initiation and transitioning preparation period. The ego vehicle is shown in red and the lead car is white. When the Ego vehicle reaches its limits, the system may initiate (true alarm) or fail (no alarm) to initiate the TOR, and the driver takes the control back from the automated system. | 2012.15441 | Fig2-scenario.jpg |
0.44671 | 3bc4e089d380491ea37df59f61df33a5 | The ROC curve comparison of our DeepTake and six ML classification algorithms for classification of takeover behavior: (a) takeover intention, (b) takeover time, and (c) takeover quality. The ROC curve shows the average performance of each classifier and the shadowed areas represent the 95% confidence interval. The macro AUC associated with each classifier is shown where AUC value of 0.5 refers to a chance.[Best viewed in color] | 2012.15441 | ROC_ALL_test_4.jpg |
0.48949 | 2bba2febd1324a8d8e4b2769354e344a | User study setup. This custom driving simulator consists of a 30-inch monitor, a Logitech G29 steering wheel, and 10.5-inch Apple iPad Air on which the non-driving tasks are displayed. For switching between the automated and manual control of the vehicle, the participant needs to press the two blue buttons on the steering wheel simultaneously. The participant wears a pair of eye-tracking glasses, and a wearable device with GSR and PPG sensors for the biometrics acquisition. | 2012.15441 | Study_Setup.jpg |
0.424615 | 0d0e758f7ac9452a90beef644e6fb04d | Average trajectories when drivers took over control from automated system after receiving TORs. Top graph shows the lateral position of the vehicle with respect to no alarm (silent failure) and true alarm (explicit alarm). Bottom graph shows the lateral position of the vehicle for three categories of takeover time (low, mid, and high). The light shaded area representing standard deviation at each time point. | 2012.15441 | Trajectory4final_2.jpg |
0.440981 | ec1107ba53d4465380fb0a54a0de3952 | Serial pipeline that consists of two commands f_1, f_2 | 2012.15443 | pipeline1.jpg |
0.408712 | 91fe704a00c4466792c3377bbb2445db | Unoptimized parallel pipeline that executes a combiner after each parallel command | 2012.15443 | pipeline2.jpg |
0.453544 | b8b0a99cd7f948e79010a9a3b5d7997d | Optimized parallel pipeline that executes multiple commands in parallel after eliminating intermediate combiners | 2012.15443 | pipeline3.jpg |
0.488562 | c11b40c7b2444aed861f2f9abcada49f | system workflow includes splitting an original serial pipeline into commands, synthesizing combiners for each of these commands, and reassembling the commands and synthesized combiners into a new parallel pipeline. | 2012.15443 | system.jpg |
0.457188 | c718bcf3e038497e9292770c8ccf5d18 | Instantaneous flow | 2012.15444 | 20_inst_meandering.jpg |
0.409624 | 8a827ba911ec48969f7ea970acdde54b | isolines Reynolds stresses at X/L=2.0 in yz-plane. (a): ⟨ u'u' ⟩; (b): ⟨ v'v' ⟩; (c): ⟨ w'w' ⟩; (d): ⟨ u'v' ⟩; (e): ⟨ v'w' ⟩; (f): ⟨ u'w' ⟩. | 2012.15444 | 20_stresses_xbyL2.jpg |
0.496151 | 5da9909e49824a4995bd98c42239f532 | Meandering of the L-vortex using reconstructed flow fields. Black and red circles indicate the beginning and the end of the meandering trajectory. | 2012.15444 | L_vort_recon2.jpg |
0.465785 | fc3ebf960b0a485c9f172512a537788f | Meandering of R-vortex using reconstructed flow fields. | 2012.15444 | R_vort_recon2.jpg |
0.397275 | 77b39009236f47c3a1ec070a96c07a19 | Top: R-vortex from the simulation. Bottom: Theoretical Model (Batchelor vortex). U, V, W velocity components are shown from left to right. Note the R-vortex in simulation is centered to (0,0) to compare with the fitted Batchelor vortex. | 2012.15444 | Simulation_Batchelor_xbyL2.jpg |
0.486819 | 70d79d3644bc425ab83af737c162f557 | 2012.15444 | U_azimuthal_avg_xbyL2.jpg |
|
0.450154 | cdc9829eefec430a87b32099cc8b344c | Computational domain and grid employed for current afterbody simulations. Grids on the center xy-plane of the base, and at downstream farfield location are shown to illustrate the grid topology. Cylindrical grid with origin at the center of the base and periodicity in the azimuthal direction is used. Domain extends 12D and 14D in downstream (green) and radial farfield (red) directions respectively. Bottom half of the computational domain is not shown. | 2012.15444 | cutcyl20_mesh_annotated.jpg |
0.435333 | 4b34078f452b4e5d9fc2b039004d38a3 | 2012.15444 | energy_modes_compare.jpg |
|
0.461473 | 5dad552ab467449bbbbcf85ff1625a4c | 2012.15444 | losses_compare.jpg |
|
0.437809 | ca0697eb56cf4ee49700fab9daf3e048 | Canonical configuration employed to reproduce the streamwise-oriented wake vortex pair. Cylinder diameter is D and upsweep angle is ϕ=20. The coordinate axis origin is located at the center of the upswept surface, (0,0,0), (red cross: ×) and the axial slant length is L = D cot(ϕ). The shaded colored region encompasses the near-computational domain. The upstream apex (UA) and the downstream apex (DA) are located at x/D=-cot(ϕ)/2,y/D=-0.5,z/D=0 and x/D=cot(ϕ)/2,y/D=0.5,z/D=0 respectively. To describe the flow around the upswept base, the transformation X=x+D/2cot(ϕ) is used such that UA and DA are at X/L=0 and X/L=1 respectively. At the inlet, simulated boundary layer data matching the experiments <cit.> are used. | 2012.15444 | meadering_sketch.jpg |
0.5181 | c589403e300d484e9b8b8cd210206805 | Phase-portrait of instantaneous vortices at X/L=2.0. The location of the mean vortex is shown as the large filled square. Black and red circles respectively indicate the beginning and end of the trajectory. | 2012.15444 | meandering1.jpg |
0.4754 | b88831eb034c4685ada94e091157d763 | 2012.15444 | meanflow_xbyL2.jpg |
|
0.487323 | a7b59fbae04b49f29f7d70631f129d9c | q̃_1,2(x,t) | 2012.15444 | recon12.jpg |
0.450929 | d58567e3a9c8419390e6cdeb4a6c3f3d | q̃_2,2(x,t) | 2012.15444 | recon2.jpg |
0.4734 | 09db486a991a45e69ab0e59d4cfdd9cb | Isolated vortex (72× 72, c=0.975). | 2012.15444 | tangrid_1vort_cp975.jpg |
0.517354 | a0f430ec337b42dfbe0bc6dde8928fc8 | Vortex pair (100× 100, c=0.75). | 2012.15444 | tangrid_2vort_cp9.jpg |
0.410595 | 3e14fc4b02c5458f917e77321a50ee6b | Unstable mode at k=2. | 2012.15444 | temporal_mode1_k2.jpg |
0.504353 | fa33dbf8e45e4f1fbc7a8a5bc82779b7 | 2012.15444 | u_vortex.jpg |
|
0.395789 | fb7ae45373b34799ba94912984aad18c | Isolated vortex. | 2012.15444 | vort1_temporal1_tanh_k2.jpg |
0.410221 | 68e0d220422a4a23996da624581baa2e | Spatial (ω=2). | 2012.15444 | vort2_spatial_om2.jpg |
0.520403 | 4235cf14d90846708634b8485ceba71d | Vortex pair (u=0). | 2012.15444 | vort2_temporal_U0_k2.jpg |
0.477505 | 91ac5e4504614c39970eb61fc29635e8 | Temporal (k=2). | 2012.15444 | vort2_temporal_convergence_tanh_k2.jpg |
0.410508 | 89572872a38c4255a04b4280b019e6a8 | Pair of counter-rotating vortices at a streamwise location. δ-radius and |Γ|-circulation of vortex. b-separation between vortices. | 2012.15444 | vortices_sketch3.jpg |
0.40521 | a9325bc082174b7db7157fa243157e35 | The general pipeline of the decomposition methods. These methods distribute the prediction score to input space to indicate the input importance. The target score is decomposed layer by layer in a back-propagation manner. Note that the main difference among these methods is the employed score decomposition rule. | 2012.15445 | decomposition.jpg |
0.548579 | 2a6b52a80dc64633ae80dfa64690b156 | A overview of our open-source library, which includes unified implementations of multiple common baseline methods, commonly used datasets, and three evaluation metrics. | 2012.15445 | dig_subx2.jpg |
0.432074 | 9f29e3cbc1054557b63139fd773e7250 | A data example of our graph sentiment datasets. Each node represents a word while edges indicates the relationship between the words. Note that edges are directed but their labels are ignored. | 2012.15445 | graph_sentence_structure.jpg |
0.432003 | 9b73e2f7d71c4357b73b45ea78f258a1 | The general pipeline of the perturbation-based methods. They employ different mask generation algorithms to obtain different types of masks. Note that the mask can correspond to nodes, edges, or node features. In this example, we show a soft mask for node features, a discrete mask for edges, and an approximated discrete mask for nodes. Then the mask is combined with the input graph to capture important input information. Finally, the trained GNNs evaluate whether the new prediction is similar to the original prediction and can provide guidance for improving the mask generation algorithms. | 2012.15445 | perturbation.jpg |
0.464147 | cbd60ff0a0844ae0adfe9a1e20da5000 | The general pipeline of the surrogate methods. Given an input graph and its prediction, they first sample a local dataset to represent the relationships around the target data. Then different surrogate methods are applied to fit the local dataset. Note that surrogate models are generally simple and interpretable ML models. Finally, the explanations from the surrogate model can be regarded as the explanations of the original prediction. | 2012.15445 | surrogate2.jpg |
0.538887 | 36b805664c8f4a8ea35ad4bfb8ce8dc4 | An overview of our proposed taxonomy. We categorize existing GNN explanation approaches into two branches: instance-level methods and model-level methods. For the instance-level methods, the gradients/features-based methods include SA <cit.>, Guided BP <cit.>, CAM <cit.>, and Grad-CAM <cit.>; the perturbation-based methods are GNNExplainer <cit.>, PGExplainer <cit.>, ZORRO <cit.>, GraphMask <cit.>, Causal Screening <cit.>, and SubgraphX <cit.>; the decomposition methods contains LRP <cit.>, Excitation BP <cit.> and GNN-LRP <cit.>; the surrogate methods include GraphLime <cit.>, RelEx <cit.>, and PGM-Explainer <cit.>. For the model-level methods, the only existing approach is XGNN <cit.>. | 2012.15445 | taxonomy.jpg |
0.474726 | bfdcb8343252483f8367e181d15cc63d | (a) Energy levels of ^85Rb relevant to our experiment. δ_780nm and δ_515nm are the detunings of the two pump lasers from transitions 1 and 2, respectively. Numeric labels have been assigned to each transition for readers' convenience. (b) Schematics of the experimental setup. The two pump laser beams are denoted by the red and green lines, and selected optical elements are shown along beam paths. The vacuum vapor cell is sketched as a square with 4 labelled windows, while the THz beam is denoted by grey dotted line and the UV beam by purple dotted line. Details on beam paths and optical components along them are given in the text. | 2012.15449 | Fig1.jpg |
0.443468 | ba16360696ce40d38495526fb92d84f1 | Knife-edge measurements of the generated THz beam at 3 cm from the exiting Si window. The solid line is a fit to the data (black squares) with the error function, which is the integration of a Gaussian beam profile along the radial direction. The knife edge is placed after the Si window but before any other optics. | 2012.15449 | Fig2.jpg |
0.455834 | 5c5eedd249a24aa6b5cbdd0e0a42d3f3 | Generated THz and UV powers plotted versus the two detunings δ_780nm and δ_515nm for two different beam configurations. The power of the 515 nm pump is ∼15 mW for both configurations; while the power of the 780 nm laser is ∼120 mW for the collimated configuration, but ∼160 mW for the focused configuration. The plotted THz powers are calibrated to be that just before the Si window inside the chamber by taking into account the efficiencies of various THz components listed in Table <ref>. The two spectra in each row share the same false-color bar right to them. | 2012.15449 | Fig3.jpg |
0.451517 | 786ebc8779aa4fc78df74367b9706385 | Generated powers of 3.3 THz (open symbol) and 7.8 THz (filled symbol) as a function of the 780 nm pump power for the 515 nm laser powers of 10 mW (green triangle), 12 mW (red circle), and 15 mW (black square). The input pump beams are in the collimated configuration. The detunings of both pump lasers are kept at the values that produce maximum THz power. The solid lines are a guide for the eye. | 2012.15449 | Fig4.jpg |
0.534821 | 242e917ac72d4a559d348afa2448afc1 | Comparison between an attractive Coulomb potential (solid), and the two model nuclear potentials, Eq. (<ref>) (short dashed) and Eq. (<ref>) (dashed). | 2012.15451 | PotentialFig.jpg |
0.498873 | 6fafb4d08f6d47ff87d9ea9c63fc61ea | The energy per particle ϵ for 5 alike spin particles on a ring of varying radii with an interaction potential Z/|r_ij| for Z∈{-1,0,+1}. The two model nuclear interactions in Eqs. (<ref>) and (<ref>) are depicted with triangles, right and inverted respectively. Each set of points is given alongside its corresponding HF analytic expression from Eqs. (<ref>), (<ref>), and (<ref>). The corresponding Seitz radii r_s are given in the top axis. | 2012.15451 | ZConstantNExtendedFig.jpg |
0.398116 | b80382a6dd3c411785f084f1f3d66b72 | QMC (symbols) and HF (lines) energy per particle ϵ for n alike spin particles on a r_s=1 ring with an interaction potential Z/r for Z∈{-1,0,+1}. | 2012.15451 | ZConstantRhoFig.jpg |