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
|
|---|---|---|---|---|
0.454454 | c9e34f624b8f4dba9f95f02a09546119 | The pipeline of the proposed algorithm framework. | 2012.13320 | framework.jpg |
0.434431 | 8579c721d8af467f8f66bc65e22b4edb | Illustration of the mantis hexapod robot model. (a) top view, (b) close view of a single leg, (c) schematic diagram of the hexapod robot model, (d) the rotation angles of a single leg. | 2012.13320 | robot.jpg |
0.417604 | 5fd5d5793fba4623a5232faf8a71d604 | Representation of control trajectories for an individual robot. | 2012.13320 | trajectoy.jpg |
0.424382 | 9c470cd6ae4c48afab40f35caafe451b | Bar plot comparison between testing set accuracy (measured as Dice score) for unsupervised deep clustering and reinforcement learning map selection (UDC + RL) and supervised deep learning/U-net (SDL). The average value is given by the height of the bars, and the standard deviation is represented by the error bars. | 2012.13321 | acc_comp_abbrev.jpg |
0.484747 | a6480c3a89f74c19bdcb330d313f38d7 | Schematic of the unsupervised clustering convolutional neural network. The raw image is fed as input into the network with convolutional layers via 3 × 3 kernals and zero padding to keep subsequent layers the same H × W size. The three convolutional layers have 100 channels and each pixel has 100 features. Then argmax is obtained over the 100 features for each pixel, selecting the most important feature over the image. | 2012.13321 | clustering_cnn_architecture.jpg |
0.401041 | d02dd5c621204edabd103227f1489b77 | Overview of the training scheme. | 2012.13321 | flow_chart_unsupervised_segmentation.jpg |
0.396631 | 25f2d0e33276458e8438d3b92f09c21a | Sample mask predictions. (A) shows one of the testing set images, (B) displays the hand annotated mask overlaid in red, (C) shows the predicted mask from the unsupervised deep clustering reinforcement learning approach overlaid in green, and (D) shows both masks overlaid, noting that their region of overlap becomes yellow. | 2012.13321 | mask_example.jpg |
0.410895 | 7ff0d3474ccb44038d77aad0a4e5c5fb | Set of clusters produced by the unsupervised deep clustering CNN. These serve as lesion mask candidates. The user can select the best image or mask number; here the best mask is clearly the first one, in the top left image. That cluster ℳ and its complement ℳ^C are then used to train the reinforcement learning network. The reinforcement learning network learns to select which is the best cluster (mask or its complement) to serve as lesion mask. | 2012.13321 | mask_picks.jpg |
0.449759 | 8175714101404edca0f7600f7fccddd2 | Architecture of deep Q learning CNN to select which region should be the lesion mask. | 2012.13321 | mask_selection_unsupervised_CNN_architecture.jpg |
0.454916 | 9fbce6be53a5494ab47a7786f2ff227c | Illustration of target vs. output prediction from the network illustrated in Figure <ref>. On the right is a tiling of the input image. The individual cells are the so-called superpixels, generated by the SLIC method. For the purposes of illustration, a small subset of the image is analyzed on the left panel. The CNN from Figure <ref> predicts most important features of the 100-feature embeddings in the penultimate layer of said figure. The goal is for each pixel in a given superpixel to be the one type of feature. Here we illustrate three features as three shapes. The algorithm to generate the target is to make all pixels in a given superpixel equal to the most numerous or common feature type predicted in the CNN output. | 2012.13321 | target_from_CNN_output.jpg |
0.369334 | b3e663db2e4d4e2a91047b557d56cab1 | Test set accuracy of reinforcement learning map selection as a function of training time. All 10 test set images have the correct mask selected by the fourth episode and there is no subsequent dip in accuracy to suggest over-fitting. Best fit sigmoid function curve is shown in red. | 2012.13321 | test_set_accuracy_learning_sigmoid_fit.jpg |
0.398587 | d7806c1ae0f84339ae732647f5340e07 | Training set accuracy as a function of training time. A steady and monotonic increase is manifest. | 2012.13321 | training_score_learning.jpg |
0.455958 | cf87b277716b49848a04578cc977f4c8 | The architecture of the proposed low-light enhancement network. The proposed LEUGAN consists of two branches, one is to process the original image, and the other is to enhance the original image after edge detection. Specific details are described in Sec. <ref>. | 2012.13322 | moxing.jpg |
0.45319 | ce4acc63b76940c896f1e2e4710443b6 | Sketch of a section of constant t and φ of the black ring (figure based on Ref. <cit.>). The disc at x=1 and infinite annulus at x=-1 are the axes (fixed points) of ∂_φ, while the axis of the ring is at y=-1 (fixed points of ∂_ψ). Surfaces of constant y have topology S^1× S^2. y=-1/ν corresponds to the horizon (shaded surface) while surfaces of constant y∈(-1/ν,-1) are fatter rings containing the horizon in their interior. | 2012.13323 | BRPic.jpg |
0.480357 | d196886bd74d41db93bf953578ff3b34 | ROC curves for all classes which are showing area under the curve (AUC) approximately equal to one. | 2012.13325 | Roc_MC.jpg |
0.434186 | 89eb41f417914243bbc5ebe6ac808047 | An illustration of the proposed hybrid model that has two components, i.e., Bayesian CNN module as a feature descriptor and active learning module. Image X_i is the queried image, Z_i is the Bayesian feature vector, T is the threshold value of uncertainty, and y_i is the predicted label. | 2012.13325 | Upd_dia_meth_FV3_1.PNG.jpg |
0.456101 | a5ae6daacf7843a0b062c7f61fdcbad5 | Images of different dimensions in original data set (from ID=1 to ID=3) are centre-cropped to produce the dimensions of 224 × 224. | 2012.13325 | images_updated_last.jpg |
0.424277 | ba4be76683354e81bd989db344447fed | Diagram of CNN. Dropout layer is added after each convolutional layer, whereas feature map shows the output of each block after applying max-pooling where dropout is applied after each convolutional layer in the block.Also, FC1, FC2, FC3, and Sigmoid layer has 1024, 256, 64, and 1 neurons respectively. | 2012.13325 | mydiagarch2.jpg |
0.523808 | 66493dbe7fab4cc694f79dd8496f1c55 | (Color online) The usual MD simulation gives structure (e.g. pair correlation function g(r)) as an output given a pair potential V(r) at a fixed temperature and density. The above shows the conventional route to go to structure given the potential through MD and below shows the inverse problem of asking which pair potential lead to the observed pair correlation function using neural network (NN). | 2012.13330 | Fig1.jpg |
0.448102 | 0434868b3640460d8ec26104b462c52d | (Color online) Learning curve for the NN based ML model studied: (top) Accuracy (r^2) and (bottom) relative error (RE) of the test data set (see text for details), as a function of number of data (N) in the training set. Both r^2 and RE saturate to a value of ∼ 0.975 and ∼ 0.09 respectively indicating a fairly accurate ML model within few hundred data points in the training set. | 2012.13330 | Fig2.jpg |
0.385449 | c9bc9f6ac2974d91a3d8c19553edf6e8 | (Color online) Comparison between the true potential (red solid line) and the predicted potential using the ML algorithm for (a) crystalline state (b) liquid like state (c) gaseous state, where attractive part in the potential is relatively strong, moderate and identically equal to zero correspondingly. Inset shows typical configurations (for this potential) from the simulation for each cases, i.e. insets in (a), (b) and (c) shows crystalline state, liquid like state and gaseous state respectively. | 2012.13330 | Fig3.jpg |
0.442482 | c652773eecbf4107a860a12df96d76c9 | (Color online) (a) Segregation order parameter f_ρ, that characterises mesoscopic fluctuations in the system shows crossover from a homogeneous steady state (see (c) for the snapshot) to a motility induced phase separated state (see (d) for the snapshot). (b) Strength of the attractive interaction or V_a (measured from the predicted potentials from ML) also show a similar crossover at Pé∼ 100. | 2012.13330 | Fig4.jpg |
0.412742 | b3dacb67b91f4d2c91f5a548412388f1 | (Color online) (a) The bare interaction potential (solid red line) used for ABP simulation and the predicted effective potential (blue dashed line) (b) comparison between pair correlation function g(r) from MIPS (red solid line) and passive simulation with NN predicted potential (blue dashed line). Section of the snapshots taken from MIPS (c) and passive simulation (d) with NN predicted potential for a visual comparison between the states. | 2012.13330 | Fig5.jpg |
0.479865 | 095a9d31e69d4a0da158177dc6bb1744 | Learning curve for the NN based ML model: (left) Accuracy (r^2) and (right) relative error (RE) of the test data set as a function of number of data (N) in the training set where the dimensionality of the output vector n_0=40. | 2012.13330 | S1.jpg |
0.420219 | 050fa1976f4e4465894b1a284e2ad44e | Learning curve for the RF model : (left) Accuracy (r^2) and (right) relative error (RE) of the test data set as a function of number of data (N) in the training set. | 2012.13330 | S2.jpg |
0.416319 | 6ceb9c0a2daa48f0bfed5cf06ffa4239 | (left) Accuracy (r^2) and (right) relative error (RE) of the test data set, as a function of epoch for Case A (see text). | 2012.13330 | S3.jpg |
0.390979 | 606b1e6bfa064fe291c7a7d3e4e734b0 | (left) Accuracy (r^2) and (right) relative error (RE) of the test data set, as a function of epoch for Case B (see text). | 2012.13330 | S4.jpg |
0.417546 | 17460b4b52134653845230cc5f864659 | Evaluation of q(λ), for λ = (10, 10, λ_3), 90 ≤λ_3 ≤ 1200. | 2012.13331 | Figure1.jpg |
0.45435 | f18cd7fdb9b34a05a8eef6bb6ad441b7 | CG iterations for flat and warm start. Left axis shows the RMP objective function value as the difference of g^(k) from the optimal LD bound g^* = $32,169,154.7. Right axis shows total columns at each iteration. UC MILP solved with 0.01% optimality gap. For comparison purposes, note that the objective function value of IR is $32,149,006.2 (less tight bound). | 2012.13331 | Figure2.jpg |
0.436003 | 260a84bdaa9e4bb38ffe84ec50a382d4 | Raman maps in BRS configuration for the reference a) and stressed sample c). The spectrum acquired on top of the waveguide is also displayed on the right. Only one Voigts function fits the reference spectrum in b), and two functions are needed for the stressed waveguide in d). A schematic view of the device cross-section is also included. | 2012.13336 | BRS.jpg |
0.413274 | 7b510eeb7c2e4b398b488fbb919a171e | Raman spectra recorded by IRS with a x'-polarised laser pump (e_i=x', quasi-TE optical mode) at λ=1335 nm and free scattered polarization (e_s). The Raman peak is fitted with one or two Voigt functions for the reference a) and stressed sample b), respectively. | 2012.13336 | IRS2peaksfit2.jpg |
0.487118 | 32ba3510adf84a7a8b4abf58db02417d | Experimental areas of the two Voigt components used to fit the IRS spectra in function of the scattered polarization e_s. The incident laser pump polarization is kept along the e_i=x' direction. When e_s=0^∘, the scattered polarization is along the x' direction. | 2012.13336 | Polar.jpg |
0.428149 | 8a41e2a5624a499d9b903bca5468b680 | Raman spectroscopy set-up combining IRS and BRS configurations. BRS experiments are performed with a λ=473 nm laser pump incident along the z'=[001] direction of silicon and only one phonon is typically probed. This technique is coupled to a mobile stage allowing the mapping of the device. IRS experiments exploit the confinement of a 1335 nm laser pump into a y'=[1-10] oriented monomode waveguide. The incident and scattered polarization allow to probe up to three phonons. | 2012.13336 | Ramansetup2.jpg |
0.418072 | 4a148a0e81984e4abda987be8ce95017 | Stress and optical mode confinement simulation. a) Dimensions of the waveguide and simulation window. b) Optical quasi-TE mode confined in the waveguide at λ=1335 nm. c) to f) Stress field distribution of the four non-zero components of the stress tensor. The effective stress values, considering the overlap with the electric field distribution of the optical mode, are indicated on the top of each respective window. | 2012.13336 | Simu.jpg |
0.490488 | 4f6a382b644d462dbc5c3ec4ba5fc4f9 | PSD of transmitted signal from the antenna with the highest spectral regrowth with DAB under per-antenna power constraint in (<ref>); LoS channel, B = 16 antennas, U = 1 users, equal PAs as in (<ref>). By controlling ρ_max,b, we can bound the spectral regrowth within a given mask limit. | 2012.13337 | PSD_resub.jpg |
0.466103 | 96bc1d34b06746ad93842cc2a91d6f40 | CDF of the sum rate with ZF and DAB precoders with perfect and imperfect CSI at the transmitter; geometric channel model, B = 32 antennas, U ∈{2, 8} users, equal PAs with β_1^(b) = 1, and β_3^(b) = -0.049-0.023j for b = 1, …, B, ρ_tot = 43 dBm, N_0 = -85 dBm (corresponding to SNR_avg = 18 dB). | 2012.13337 | resub_fig6_v6.jpg |
0.469006 | ce45e53ea7974c61b131d8b491142c15 | Achievable rates averaged over 500 realizations of LoS channels in (<ref>) with uniformly randomly distributed AoDs ψ∈ [0, 180), B=10 antennas at base station and U=1 user. The nonlinearity coefficients are as in Table <ref>. With distortion-aware precoding, the achievable rate becomes a strictly increasing function of the SNR. | 2012.13337 | rev_awesome_plot_example8.jpg |
0.412914 | b3ad905403ae4252a8a85a80eb57a01a | CDF of sum rates with different linear precoders; geometric channel model, B = 32 antennas, U = 4 users, equal PAs with β_1^(b) = 1, and β_3^(b) = -0.049-0.023j for b = 1, …, B, ρ_tot = 43 dBm, and SNR_avg∈{-15, 25} dB. | 2012.13337 | rev_fig5_v2.jpg |
0.431728 | ed6e48c810c64f65862f14b8270d849b | Average sum rate with different linear precoders with and without power control; geometric channel model, B = 32 antennas, U = 4 users, equal PAs with β_1^(b) = 1, and β_3^(b) = -0.049-0.023j for b = 1, …, B, N_0 = -85 dBm. | 2012.13337 | rev_fig7_v10.jpg |
0.46767 | 8f8cc92657cb40c9b4a32800c7ca565e | CDF of total base station side consumed power with ZF, DAB and EE-DAB precoders for R_0 = 35 (bits/c.u.); geometric channel model, B = 64 antennas, U = 4 users, equal PAs with β_1^(b) = 1, and β_3^(b) = -0.049-0.023j for b = 1, …, B, N_0 = -85 dBm, power consumption model parameters η_max,b = 0.55 and ρ_max,b= 25 dBm. | 2012.13337 | rev_fig8_v3.jpg |
0.407049 | 7dc08c47eec84c499963efa21a3d3486 | Massive MIMO downlink with linear precoding and nonlinear PAs at the transmitter. | 2012.13337 | system_model_v9.jpg |
0.396084 | 756aa8e1a7694ad5a57032f1a159f599 | Diagrammatic representation of the 1+→ 2 all-orders equation defined in Eq. (<ref>), where the open circle representing _0 contains all two-particle irreducible diagrams in the s_f = P_f^2 channel. | 2012.13338 | H_amp.jpg |
0.477628 | 39541ae379b74636942f08a2881d95de | On-shell projection of , where the left hand side is Eq. (<ref>) and the right hand side is Eq. (<ref>). The white box indicates the real function 𝐇. | 2012.13338 | H_ampOnShell.jpg |
0.396084 | 597f79e9ef244f2e9991463bfc8d575f | (a) Self-consistent integral equation for the hadronic scattering amplitude as given in Eq. (<ref>). (b) Examples of diagrams contributing to the Bethe-Salpeter kernel, _0, which contains all s-channel two-particle irreducible diagrams. | 2012.13338 | M_amp.jpg |
0.497666 | f9807a73700743feac093d982fb9cf1b | Singularities of the scalar and S wave case of the loop integral as reproduced by the function Sing _00;00. The value of the spatial momentum |P⃗_f,i^⋆|=(2π/6)m. | 2012.13338 | Sing.jpg |
0.357343 | cf981f174fc84137b794cf9475637c4b | Diagrammatic representation of the expansion shown in Eq. (<ref>). The dotted line in the first term represents the pole piece of the propagator, D_2, and the second term is a new smooth kernel we define as _L|0. | 2012.13338 | W_L.jpg |
0.396084 | cbcc8be6dc2c463386d884ed9d61c929 | Diagrammatic representations of the all-orders equations for _1B given in Eq. (<ref>) and _1B in Eq. (<ref>), where the open circle is the kernel i_0|0. | 2012.13338 | W_amp.jpg |
0.432389 | 0ef824d15d2849259884c9cf86073115 | (a) The relation between the full 2+→ 2 amplitude and _ as defined in Eq. <ref>. (b) Depicted are all allowed current insertions over the external legs. The symbol “Σ" is used as a shorthand to express this sum. The dotted lines represent the pole piece D of the propagator, the other objects are defined in Fig. <ref>. | 2012.13338 | W_and_Sigma.jpg |
0.540057 | 6709a72817144cd9bd0c8d962df5790b | Diagrammatic representation of the amplitudes considered in this work. Shown are the (a) 2→ 2, (b) 1+→ 1, (c) 1+→ 2, and (d) 2+→ 2 amplitudes along with momentum assignments. | 2012.13338 | amplitudes.jpg |
0.477628 | 4883ed3092584200adcc0087155af470 | Iterated loop identity as given in Eq. (<ref>), where the open circle is the Bethe-Salpeter kernel _0, the box with label j is the j-th kernel, and the dashed line indicates the two-particle phase space cut which places intermediate state particles on-shell. | 2012.13338 | loop_id.jpg |
0.471325 | a67d888862874074b1e93bf2dedf853e | (a) 2→ 2, (b) 1+→ 2, and (c) 2+→ 2 amplitudes near the resonance pole, s_R. The residues at the pole define the coupling, c, transition form-factor, f_1→ R, and elastic resonance form-factor, f_R→ R. | 2012.13338 | resonance.jpg |
0.471214 | 71f46cadacc044d6b51919619da1eb21 | Decomposition of triangle diagram into on-shell pieces involving new kernels _0|C|0, _L|0, _R|0. | 2012.13338 | triangle_decomp.jpg |
0.483904 | d3615e1afd5944339c914002478b4d1f | Triangle-function contribution to _, Eq. (<ref>), written in terms of the single-hadron form-factors (f_j) and the triangle loops (_j) defined in Eq. (<ref>). The gray circle and dashed lines were defined in Figs. <ref> and <ref>, respectively. The open semi-circles represent the modified spherical harmonics, defined in Eq. (<ref>), which have the angular momenta associated with the initial and final two-particle states. | 2012.13338 | triangle_function.jpg |
0.40628 | 95f7e29baf03440596969c8438672bfe | Cycle constraint. We link audio and video latent spaces with a cycle constraint that ensures that the audio signal content is preserved through video decoding and encoding. | 2012.13341 | cycle.jpg |
0.458373 | a181f33e52b64777afe4ed0ee7aa8d35 | Disentanglement, by mixing content encodings from two speakers saying the same phones (_c_i with '_c_i) and style encodings from the same speaker saying different words (_s_a with '_s_a). | 2012.13341 | disentanglement_compact.jpg |
0.456598 | a330fc650c9e4fa8a4ba3548cd945a91 | Word analysis. Instances (different rows) of the same words are encoded with a similar video sequence (single row), yet the show pronunciation variations if spoken by different speakers. | 2012.13341 | flattened_video_face_content.jpg |
0.422351 | 808886c6ce8847cbaab96b99110baaf0 | Environment sounds visualization. Although not trained for it, our approach also visualizes natural sounds. Here showing a clear difference between chirp and alarm clock. | 2012.13341 | flattened_video_face_content_esc.jpg |
0.407215 | 30b7b6e1bc6c4680bdabdd725e6035af | Style analysis. Opposed to the content encoding, the style encoding enhances different speakers influence (rows) while having similar encodings across different words (columns). | 2012.13341 | new_words_style_face_cycle_log_flipped.jpg |
0.433573 | e6c8983919ac4e38a6cae18a08b12dd6 | Overview. A joint latent encoding is learned with audio and video VAEs that are linked with a cycle constraint and augmented with a smoothness constraint (Δ v annotation). | 2012.13341 | overview.jpg |
0.439151 | e34f5e70bc8b4a0e85ec0d88073d339d | Smoothness evaluation by plotting the latent embedding of an audio snipped in 3D via dimensionality reduction. Left: Without smoothness constraint, consecutive audio sample embeddings are scattered. Right: Our full model leads to smoother trajectories. | 2012.13341 | smoothness_MDS4.jpg |
0.426129 | 6d211b507f6f4dee93f53914bdad1475 | : A tool developed towards the long-term goal of helping hearing impaired persons to see what they can not hear. We map an audio stream to video and thereby use faces or numbers for visualizing the high-dimensional audio features intuitively. Different from the principle of lip reading, this can encode general sound and pass on information on the style of spoken language. | 2012.13341 | teaser.jpg |
0.414728 | 8bc94056a8fa46e69a325b7b5e9c39ed | Throughput. The information throughput can be measured by mapping from audio to video and back. Playing the depicted Mel Spectrogram shows that the speech is still recognizable. | 2012.13341 | throughput.jpg |
0.509229 | c894044b0b494b06af1f369eeb075f19 | Model Comparison on content questions. The mean user accuracy and standard deviation for distinguishing phoneme sequences is compared between the models, annotated with the significance level († for 0.05<p<0.1, * for p<0.05, ** for p<0.01). The mean accuracy for random guessing is 0.384. | 2012.13341 | user_study_phoneme.jpg |
0.439649 | 860cfc28f5f94510aff4f6141fb3b724 | `A schematic illustration of an unconstrained plate exhibiting residual stress. (a) The two elements composing the plate are shows side by side. (b) As the red trapezoid is too large to fit into the square opening, it is compressed. (c) For a plate sufficiently thin, the induced compression exceeds the buckling threshold, and the trapezoid buckles out of plane. Note that there are many shapes that preserve all lengths along the faces of the plate, yet they cannot be planar' <cit.>. | 2012.13345 | EfratiPic1.jpg |
0.415509 | 844fbbc08ed84a9c991b09742f4025fa | An example vertical volume slice is given on the left with high cone beam angle artefacts visible on top of the reconstruction. The graph showing the total of grey-values per slice is given on the right. The red and orange bars highlight the slices that were discarded (50 slices from top and bottom). The corresponding graph of grey-values for the first 50 slices also show the spike of high cone-beam artefact on top, and the bottom 50 the lack of any apple information. The middle 668 slices were used for parallel beam simulations (Dataset A) and those with added Gaussian noise (Dataset B). | 2012.13346 | 3Drecon_highconebeam.jpg |
0.422063 | 6a5e1b37de7245fa8150c6b0473805e9 | Baseline filtered back-projection (FBP) reconstructions for full angular range and the two limited-view data setup. The grey-values of the Dataset B and C reconstructions are set to the same range of Dataset A. | 2012.13346 | baseline_recons.jpg |
0.412905 | d91cf28763f547e7bd66d5577e129561 | A horizontal cross-section of a healthy apple is shown on the left, and another cross-section with browning defects of the same apple shown on the right. The red arrows point to larger browning defects across the slice. | 2012.13346 | defect_browning.jpg |
0.364452 | 37d81dde572b47998c708fb804d7808e | A bar chart that shows the defect percentage distribution for all defects within the Test subset. The percentage distribution is determined using the resulting data split sequence obtained via the simple heuristic method (referred to as "Empirical") and mixed-integer quadratic programming solvers MOSEK via MATLAB ("MOSEK") and ECOS_BB via Python, ("ECOS_BB"). | 2012.13346 | defect_percentage_full.jpg |
0.397039 | 69e684cc49b04a7cb89b8b5af8b5881a | A bar chart that shows the defect percentage distribution for all defects within the Test subset. The percentage distribution is determined using the resulting data split sequence obtained via the simple heuristic method (referred to as "Empirical") and mixed-integer quadratic programming solvers MOSEK via MATLAB ("MOSEK") and ECOS_BB via Python, ("ECOS_BB"). | 2012.13346 | defect_percentage_partial.jpg |
0.500286 | 0e7ec4e5ae3743cca8afaf8aa3bb0902 | An illustration of the circular cone-beam CT geometry set up is shown on the left; example slices for the vertical and horizontal directions in the reconstructed volumes on the right. | 2012.13346 | fig1.jpg |
0.470734 | 725de5c83e044e1a874378f419c86a55 | Different angular samplings of generated parallel beam data: (left to right) full angular range, missing wedge sampling 1 and missing wedge sampling 2. Blue lines indicate the projection angles that were selected for the limited-view data setup. | 2012.13346 | missing_wedges.jpg |
0.426404 | 078e4e02c900461cb8a31cf78d1559a1 | MS-D post-processing results for the Target 1 shown in Figure <ref>. The FBP reconstructions used as inputs are given for both missing wedge angular samplings for each data, and the resulting output by the network. The grey-values of the Dataset B and C reconstructions are set to the same range of Dataset A. | 2012.13346 | msd_results_target1.jpg |
0.382547 | a05c1325926346bc9920fdda2e74c582 | MS-D post-processing results for the Target 2 shown in Figure <ref>. The FBP reconstructions used as inputs are given for both missing wedge angular samplings for each data, and the resulting output by the network. The grey-values of the Dataset B and C reconstructions are set to the same range of Dataset A. | 2012.13346 | msd_results_target2.jpg |
0.413315 | 1c7a95bedb5449d6a8cc6af5bb4018ea | The target apple slices for the MS-D network. | 2012.13346 | msd_target_slices.jpg |
0.524143 | a5801e3d878e40ccacb9d60d83a7f30a | Parallel beam projection geometry used to generate Datasets A and B | 2012.13346 | par_geom.jpg |
0.438596 | 340b3b87b4f544569ee4e50d4f5a9781 | The effect of scattering on an apple slice as the scattering intensity scaling factor, α is increased. Here, α = 0 signifies no scattering, and α = 1 the scattering of water with no additional intensity. | 2012.13346 | scatter_comparison.jpg |
0.43589 | 2192452891884358ad7bbab058da5531 | An elliptic plate with semi-major axis a and semi-minor axis b. The principal axes 𝐪̂_i's of the plate are related to 𝐞̂_i's under rotation by an angle θ. | 2012.13347 | PA-fig-02.jpg |
0.409153 | 349ad91963304d30ba6c5faaeebe1433 | CORLAT Test Set | 2012.13349 | CORLAT_Test_NB_vs_Target.jpg |
0.408014 | 1b86c8c15b84426794a713edb9efe9ff | CORLAT Training Set | 2012.13349 | CORLAT_Train_NB_vs_Target.jpg |
0.432201 | dc1a5212aeab4fefa70ca1b90f441ed5 | An integer program can be represented as a bipartite graph such that one set of nodes represent variables, the other set constraints, and edges correspond to the non-zeros of the constraint matrix. {A, b, c} can be encoded as features of the graph's nodes and edges. A graph neural network takes the graph and its node and edge features as input and computes a fixed dimensional continuous vector for each node. These vectors can then be used to make node-specific predictions. | 2012.13349 | MIP_GNN.jpg |
0.403758 | e5b9f470d9e74eb6a862d992bfa02191 | Bipartite graph representation of a MIP used as the input to a neural network. The set of n variables {x_1,…,x_n} and the set of m constraints {δ_1,…,δ_m} form the two sets of nodes of the bipartite graph. The coefficients are encoded as features of the nodes and edges. | 2012.13349 | MIP_bipartite_graph2.jpg |
0.458489 | a70e48ff9ed54b66b692232849f139a0 | Neural Network Verification Test Set | 2012.13349 | NN_Verification_Test_NB_vs_Target.jpg |
0.464574 | b831bbaf4e404cb6819e6999f636ffd5 | Neural Network Verification Training Set | 2012.13349 | NN_Verification_Train_NB_vs_Target.jpg |
0.410285 | c7584be2cf824b8c98aab5fd86e88073 | Overview of our approach. We construct two neural network-based components to be used in a MIP solver, Neural Diving and Neural Branching. Neural Diving uses a generative model which, given the MIP to be solved, generates partial assignments for its integer variables. The resulting “sub-MIPs" for un-assigned variables are then solved in parallel to construct high-quality complete assignments for the original input MIP. The generative model is trained on pairs of MIPs and feasible assignments (not necessarily optimal) precomputed with an off-the-shelf solver. Neural Branching is a learned policy that conditions on a leaf node of the current search tree to pick the next integer variable to branch on. It guides the tree search to quickly produce a short proof for how far the solution produced by Neural Diving is from optimum. The policy is trained to imitate an expensive “expert” policy, Full Strong Branching, that is designed to produce small trees. Combining the two produces a Neural MIP solver which is customized to the distribution of MIPs for the application of interest. Results show that such a solver outperforms the solver SCIP, even after tuning SCIP's parameters for each application and even when the parallel computation of Neural Diving is performed sequentially. | 2012.13349 | intro_overview.jpg |
0.475972 | 82beff02f6d349e6a54fdeec47e40ee2 | Main result of the paper: our approach, Neural Branching + Neural Diving, matches or outperforms SCIP with respect to the primal-dual gap <cit.>, averaged on held-out instances, as a function of running time (referred to as calibrated time, see section <ref>) for the five datasets in our evaluation with the largest MIPs. Note the log scale on both axes. | 2012.13349 | intro_plot_23Jul2021.jpg |
0.499276 | 129313a0563c411681653507dba0d163 | Histogram of speedup factors for all MIPs in MIPLIB at root. | 2012.13349 | miplib_hist.jpg |
0.498192 | 7a8203d357ba4375b76d6756518934f4 | Solve time vs number of LPs in a batch. | 2012.13349 | num_lps_v_time.jpg |
0.450124 | 5df38b3da21a4361a81cef91e791dd81 | Network statistics for the presolved datasets. | 2012.13349 | presolved_network_stats.jpg |
0.432817 | 36bef66552784a2ebf3b4e1b7ab869b2 | Average dual gap as a function of calibrated time on neural network verification instances. | 2012.13349 | supercuts_avg_dual_gap_by_time_logy_no_root_time_hline_old.jpg |
0.474283 | f7eb9fb672a34d69938eb8ecf162b993 | Number of instances as a function of the difference between the number of selected cuts at the root and the number of cuts in the baseline. | 2012.13349 | supercuts_cut_diff_hist.jpg |
0.409378 | 5016cd9633974e50a887bed38aa83ee3 | Survival plot with respect to calibrated running time. This plot only considers time after the root node. | 2012.13349 | supercuts_surv_no_root_time.jpg |
0.457473 | 6e47090856144d73b32682a0c47122ef | Conditionally independent model (Section <ref>) infers node embeddings {v_1,…,v_n} for the variable nodes of the bipartite graph representation of an input MIP M using a graph convolutional network and applies a multi-layer perceptron to the node embeddings {v_d}_d ∈ℐ of the set of binary variables ℐ in M to compute their Bernoulli distributions. (See Section <ref> for the case of general integers.) These distributions are independent of each other conditioned on the input MIP. | 2012.13349 | supermip_models2.jpg |
0.682992 | 51117d93e16e4bd9bdfc1830f27be44c | `A beam (a) before bending and (b) after bending; (c) a close-up of the displacement field' <cit.>, where h is the thickness of the beam and R is the radius of the curvature of deformation. | 2012.13351 | HowellWrong.jpg |
0.45959 | bd4f72a654a54c9eae48786d59e032d5 | Slip for δ b. | 2012.13353 | Oden1Cur.jpg |
0.417156 | b28fa50544254350a0e71419fdf5cde6 | Displacement field of the modified Kikuchi and Oden's model. | 2012.13353 | Oden1Eg01.jpg |
0.409897 | 575fc2650634460db50ac06ddac75bd0 | Slip and stick regions of the modified Kikuchi and Oden's model. | 2012.13353 | Oden1Eg02.jpg |
0.533336 | 3a77d909cd4345acb5bacdb3f1c38f16 | Slip for h. | 2012.13353 | Oden1H.jpg |
0.50554 | 46d76a09105b4c09b59aaba76737759c | Slip for ν. | 2012.13353 | Oden1Poi.jpg |
0.590354 | e2731ab8fa694a46a2440d605c3db6b8 | Slip for δτ. | 2012.13353 | Oden1T.jpg |
0.500647 | e77f022c661b4d4ead8bcb3240deaf69 | Tension ratio against δ b. | 2012.13353 | PrismEqnPlot2.jpg |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.