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
|
|---|---|---|---|---|
0.457011 | b8d872e7b26f4a44b5900067ac77f4c1 | Distribution of accuracy for different orders of features. Accuracies obtained with the CMI ordering are shown with a cross. Results using two learning algorithms for datasets and . | 2012.13798 | plot_accuracy_order.jpg |
0.477097 | 37a195a2902242fda9975fc1ef7796a5 | Schematic illustration of the physics of hyperheavy nuclei. Solid black line shows the deformation energy curve of the ^466156 nucleus obtained in axial RHB calculations with CEDF DD-PC1 in Ref. <cit.>. Open red circles indicate selected points on this curve for which neutron density distributions ρ_n are shown. The density distributions in the minima A and B (C, E, and F) are shown in the plane which is perpendicular (along) the axis of symmetry. The density colormap starts at ρ_n=0.005 fm^-3 and shows the densities in fm^-3. The density profiles reflect their relative sizes with respect of the spherical shape in the minimum D. See Fig. <ref> in the present paper and Fig. 2 in Ref. <cit.> for these density profiles in their actual sizes. | 2012.13799 | fig-1.jpg |
0.453802 | 1b9940d5a663429d92d544692297afcf | The distribution of ellipsoidal and toroidal shapes in the nuclear landscape obtained in the RHB calculations with CEDF DD-PC1. The nuclei with ellipsoidal shapes are shown by the squares the color of which indicates the equilibrium quadrupole deformation β_2 (see colormap). Note that ellipsoidal shapes with the heights of fission barriers smaller than 2.0 MeV are considered as unstable (see the discussion in Sect. III of Ref. <cit.> and in Sect. XI in Ref. <cit.>). Two-proton and two-neutron drip lines for toroidal nuclei are shown by solid black lines. White region between them (as well as the islands inside this region shown in gray) corresponds to the nuclei which have toroidal shapes in the lowest in energy minimum for axial symmetry (LEMAS). The islands of relatively stable spherical hyperheavy nuclei in the Z>130 nuclei, shown in light grey color, correspond to the solutions which are excited in energy with respect of the LEMAS corresponding to toroidal shapes. Note that in the same nucleus two-neutron drip lines for spherical and toroidal shapes are somewhat different. This is the reason why some islands of stability of spherical hyperheavy nuclei extend beyond the two-neutron drip line for toroidal shapes. The extrapolation of the two-proton drip for ellipsoidal shapes, defined from its general trends seen in the Z<120 nuclei, is displayed by thick orange dashed lines. Similar extrapolation for two-neutron drip line of ellipsoidal shapes is close to the two-neutron drip line of toroidal shapes (see Fig. 1 in Ref. <cit.>); thus it is not shown. Partially based on Fig. 24 of Ref. <cit.>. | 2012.13799 | fig-2.jpg |
0.392748 | 90dc7c3acff1413496c626416a1e0753 | The fission barrier heights E_B [in MeV] as a function of proton and neutron numbers. Only the nuclei with fission barriers higher than 2 MeV are shown. Partially based on the results presented in Fig. 6a of Ref. <cit.>. | 2012.13799 | fig-4.jpg |
0.446347 | af777987c3b947c493ca4cfda7fc3521 | Proton β_2 values (see colormap on right) of the lowest in energy solutions of selected set of nuclei (see text for details). Solid black lines indicate two-proton and two-neutron drip lines. Neutron density distributions of some nuclei are shown: orange arrows are used to indicate these nuclei. The same colormap [shown in the upper left corner] as the one used in Figs. <ref> and <ref> is employed here for the densities. The density colormap starts at ρ_n=0.005 fm^-3 and shows the densities in fm^-3. The density of the ^348138 nucleus is used here as a reference [see Fig. <ref> for its actual geometrical size] with respect of which the geometrical sizes of the density distributions in other nuclei are normalized. | 2012.13799 | fig-8.jpg |
0.459143 | 467074600ea14f23adad33908aaa1d4c | Unified network enhancement and pruning search framework. | 2012.13801 | framework.jpg |
0.509333 | 3201ff98935841dfac1a6d3f99be58fe | WL kernel illustration. At initialization, there are two pruning proposals with features at m=0. At step 1, WL kernel collects the next labels of each node. At step 2, it re-encodes or re-labels the new nodes incorporating their neighbour information. At step 3, it obtains a new graph with features at m=1. At step 4, WL kernel compares the histogram on both m=0 and m=1 features. The iteration repeats until m=M. | 2012.13801 | kl2.jpg |
0.479048 | caa689c6470b4614a7e0f38e20f96b7d | Unified scheme generation. | 2012.13801 | rnn.jpg |
0.447659 | 9e8afa61908541c89c5a680626168e35 | Convergence of NNL for Setting 1. | 2012.13805 | class_simu1.jpg |
0.449618 | 16af2d9e12bc4d9995d2395e6e616242 | Convergence of NNL for Setting 2. | 2012.13805 | class_simu2.jpg |
0.469441 | 4fb4345412684b2ebde8201c3e2220f4 | Convergence of NNL for Setting 3. | 2012.13805 | class_simu3.jpg |
0.463416 | b8987b45e294470e878f1cf937c5cb29 | Convergence of NNL. | 2012.13805 | convergence_simu.jpg |
0.532731 | 1727d79abf4f4f01b691ce1ba8411b89 | Convergence of NNL in Twins data. | 2012.13805 | convergence_twins.jpg |
0.417833 | 5e4436f421bf404085785ea596fa0d04 | The impact of parameters on the performance of ATT estimation. | 2012.13805 | performance.jpg |
0.472124 | be872dc8543f4f59980c5fcec7209e28 | Overall framework of ATT estimation via distribution learning. | 2012.13805 | weighting.jpg |
0.4364 | e8e96cad52914ea4ba396241104f224a | The crowdsourcing framework | 2012.13807 | Framework.jpg |
0.510494 | 32c27fd08d4e40f2a8b8f17f1220929b | Example of the secret sharing | 2012.13807 | Presentation1.jpg |
0.390771 | 4ceeb358400a4e2d9fb87f9c49ba2e3f | Example of the statistic based on the trie | 2012.13807 | Tree.jpg |
0.528135 | bfea7d5c8d2f424abd142df1a5b10487 | Bottom: Kohn-Sham potential V_ KS(x) (scaled by 0.1) and ground-state density n_0(x) of a model 1D insulator with four electrons per unit cell. Top: associated band structure (occupied bands in blue, empty bands in red). | 2012.13815 | fig1.jpg |
0.400801 | 4dcef9fe41644fa9b609a7855e714475 | Time dependent exciton wave functions, following resonant excitation with pulses of strength E_0, subject to static electric fields E_ stat. The exciton wave functions are plotted as a function of electronic coordinate x' for the hole fixed at x=0; the peak height is capped at 0.05. The wave function develops an asymmetry under sufficiently strong static fields, indicating dissociation. | 2012.13815 | fig10.jpg |
0.524524 | 192fded2331a44f6a4b5ad82cc0766f4 | Imaginary part of the macroscopic dielectric function of the 1D solid, for various values of the LRC parameter α, as indicated. Inset: exciton binding energy E_b versus α. | 2012.13815 | fig2.jpg |
0.456549 | 29f91cdd6d5a49dca8afcd61f25ad169 | Top: frequency-dependent TDM |Γ_s(x,x')| of the bound exciton for α=3. Bottom: electron distribution |Γ_s(x_ h,x')| for various reference positions x_ h of the hole. | 2012.13815 | fig3.jpg |
0.4038 | 73a176a1b7b542e3b03f0388689deb8b | Top: frequency-dependent TDM |Γ_s(X,X_ r)| of the bound exciton for α=3. Bottom: associated exciton wave function |Γ_s(X_ cm,X_ r))| for various reference positions X_ cm of the center of mass of the exciton. | 2012.13815 | fig4.jpg |
0.516187 | 29a85febb858497e8572577900e21adc | Bottom: ground-state density n_0(x) in a 1D supercell consisting of 7 primitive unit cells of the cosine potential (<ref>); defects are simulated by reducing the depth of selected potential wells. Top and middle: TDM |Γ_s(x,x')| and PHM |Ξ(x,x')| of the exciton (α=2) within one supercell. | 2012.13815 | fig5.jpg |
0.455839 | e70dac39442943e995a1b2f091fd34ee | Comparison of the two representations of the TDM, |Γ_s(x,x')| (top) and |Γ_s(X,X_ r))| (bottom), for the exciton in the 1D solid with defects (see Fig. <ref>). | 2012.13815 | fig6.jpg |
0.401811 | 1b6a67db525e41de898e7aa808196445 | Time-dependent TDM for the 1D solid of Fig. <ref>, subject to 5-cycle pulsed fields of the form given in Eq. (<ref>), with E_0=0.0001, for three different values of ω_d: above resonance (top), on resonance (middle) and below resonance (bottom) with the 1D exciton (here, α=2). | 2012.13815 | fig7.jpg |
0.449438 | ad6f9b9f97464c839a7d5ae361cbc502 | Time average of the time-dependent TDM for resonant excitation (see middle panel of Fig <ref>), calculated after the end of the pulse. | 2012.13815 | fig8.jpg |
0.491962 | 845562b5d99d47429e434a2e669c166e | A scenario of 6G-supported cooperative connected vehicles. | 2012.13817 | system.jpg |
0.40934 | 6626dff1b7a6439ebf0b57ddf9aef203 | Allowed region for f_DM versus M_PIMBH with only ironclad constraints included. This may be compared and contrasted with Fig. 10 in <cit.> | 2012.13821 | CMBfigure.jpg |
0.445305 | 1a674555309644438661035844a0abac | Illustrative example of our method. Prior to training a metric on the initial data, no class association could be formed given a skeleton sequence. After training our one-shot action recognition model, skeleton sequences can be encoded. A euclidean distance on the encoded sequence allows class association by finding the nearest neighbour in embedding space from a set of reference samples. The colors are encoding the following classes: throw, falling, grab other person's stuff. Brighter arrow colors denote higher distance in embedding space. | 2012.13823 | motivation2.jpg |
0.390106 | 972efd32478f44789993687be341ff60 | A possible intermediate state of the embeddings during the training process of two classes (left). During training, pairs, that are difficult to push apart in embedding space, are mined (middle). Given the blue anchor sample, the most difficult positive pair is the blue sample with the highest distance in embedding space. Similar, the closest red sample in embedding space is the corresponding negative sample. The overall goal is to separate the samples in embedding space (right) by minimizing the inter-class scatter and maximize the intra-class distance to the class centers in embedding space. | 2012.13823 | ms_example.jpg |
0.421505 | 98adf10d397849a1b97a7a868a48ab3a | skeleton representation. x and z denote the skeleton joint component in joint space, the number of joints is reflected by , which relates to the height of the image H, the sequence length relates with the width of the image W. Note, instead of projecting the temporal information throughout the width of the image, we project the joint space locally for each dimension and assemble the joint axis blocks over the width. | 2012.13823 | reindex_representation_overview.jpg |
0.465567 | d69589a76af04b99a9bedd63f96aa548 | Exemplary representation for a throwing activity of the NTU-RGB+D 120 dataset. A skeleton-sequence serves an input and can be represented as an image directly <cit.>. Our representation groups x-, y-, z joint values locally in /3 blocks per axis and assembles them into the final image representation. All axis blocks are laid out aside. | 2012.13823 | skeleton-dml_representation.jpg |
0.480646 | 85161e2be88e45e18af162142b67c543 | NTU RGB+D 120 skeleton joint positions. | 2012.13823 | skeleton_kinect2.jpg |
0.452694 | a24ee71b088b4b968d48ba3bcb0fbc30 | Schematic diagram for the crossing of two identical nuclei, where partons can be produced anywhere inside the rhombus, for (a) the first and (b) the second piecewise solution in Table I. The solid diagonal lines represent the light cone boundaries for partons that can reach z ≈ 0 at time t, while the hyperbola represents the boundary of these partons after considering the formation time t_F=τ_ F cosh y. | 2012.13825 | crossing-diagram.jpg |
0.44986 | bff8bb3a1d114058acb018d5407c770f | Net-proton dN/dy data (circles) for central Au+Au (Pb+Pb) at = 2.4, 5, (17.3), and 200 GeV in comparison with the scaled net-baryon parametrization (curves). Filled circles represent actual data and open circles are reflected data across y=0. | 2012.13825 | dnetbaryondy-vs-dnetpdydata.jpg |
0.476382 | f979569e74d04868b6dfd91c90d6a68a | Maximum energy density using the parton dm_ T/dy, the hadron dm_ T/dy, or modified parton dm_ T/dy (see text for details) for τ_ F= 0.3 fm/c as functions of energy. The inset shows ratios of ϵ^ max from various dm_ T/dy profiles to ϵ^ max from the default parton dm_ T/dy, while the solid straight line shows the energy density if two boosted nuclei simply overlap. | 2012.13825 | emax-parton-vs-hadron.jpg |
0.456324 | faff3d3b682842af898417ccf275723d | Maximum energy density for central Au+Au collisions as a function of collision energy at τ_ F=0, 0.1, 0.3 and 0.9 fm/c in comparison with the upper bound of ϵ^ max of Eq.(<ref>), where the analytical low- and high-energy bounds are also shown. Dashed curves represent the ϵ^ max results when using a constant formation time t_ F=0.1, 0.3, and 0.9 fm/c. | 2012.13825 | emax-proper-vs-constant-tauf.jpg |
0.423377 | 1fa2750da7c74cf59af4deb451bb4b91 | Energy density of produced partons in central Au+Au collisions at = 3, 7.7, 19.6, and 39 GeV for τ_ F=0.1, 0.3 and 0.9 fm/c; the triangular solution for τ_ F=0.3 fm/c is also shown for comparison. | 2012.13825 | epsilon-vs-t.jpg |
0.490705 | e306f049bacc4d1fbe5c69608d899e6f | Parametrized initial rapidity density of transverse mass of produced partons (solid curves) for central Au+Au collisions at = 2, 3, 5, 50, and 200 GeV. Symbols represent the results of initial partons from the AMPT model, while dashed curves represent the parametrized hadron dm_ T/dy at these energies. | 2012.13825 | parton-vs-hadron-dmtdy.jpg |
0.442965 | dfde23cdb5824ea5a81d878ad621b119 | Dependence of free energy of tail (solid) and loop (dashed) on the number of segments they contain | 2012.13827 | F-tail-loop.jpg |
0.491142 | 0c26ea9aa3aa41f78ff0d573fc2265fb | The effect of modification of ligands on location of superselectivity peak for 4-valent MV-particles. Blue curve is unmodified 4-valent particle, red is modification of one ligand, green modification of two ligands. Dashed curves correspond to modification with weaker stickers, solid curves correspond to modification with stronger stickers. Concentration of the polymer solution is cR_g^3=0.001. | 2012.13827 | MV-copolymer-pic.jpg |
0.625407 | 2e520df65f7e4419b406103b99056ac3 | Visual description of Langmuir adsorption model | 2012.13827 | langmuir.jpg |
0.531651 | f0973f8303154238b51102639ccca7fa | Selectivity curves for 10-valent polymers with linker lengths: blue curve n=1, red n=10, green n=100 and orange n=1000. Concentration of the polymer solution is cR_g^3=0.001. | 2012.13827 | linker.jpg |
0.422874 | e2ef03f7db984c5ead28d34d8c4c6ca2 | Visual description of multivalent particles adsorption problem | 2012.13827 | multivalent.jpg |
0.459358 | ad0353039d734900ba7dff2d055f14d7 | Schematic illustrating of the division of adsorbed polymer into tails and loops by adsorption segments | 2012.13827 | polymer.jpg |
0.46826 | d900054af55947e79d85a8037e5015d4 | Comparison of selectivity for multivalent polymers (solid lines) with linker length n=1 with selectivity of Martinez-Veracoechae's multivalent particles (dashed lines). Length of linkers n=1. Blue curves valency κ=10, red to κ=6, green to κ=4 and orange to κ=2. Concentration of the polymer solution is cR_g^3=0.001. | 2012.13827 | selectivity_n1.jpg |
0.457578 | fd1ee2608a9f4785b824dc1c1e399824 | Comparison of selectivity for multivalent polymers (solid lines) with linker length n=10 with selectivity of Martinez-Veracoechae's multivalent particles (dashed lines). Length of linkers n=1. Blue curves valency κ=10, red to κ=6, green to κ=4 and orange to κ=2.Concentration of the polymer solution is cR_g^3=0.001. | 2012.13827 | selectivity_n10.jpg |
0.496235 | b34cee7b44b3420889ea3a3b02ffa017 | Comparison of selectivity for multivalent polymers (solid line) and Martinez-Veracoechae's multivalent particles (dashed lines). Total length of polymer is fixed for all chains N=90, so length of linker depends on valency n=N/(κ-1). Valencies are κ=10 for blue curve, κ=6 red, κ=4 green, κ=2 orange. Concentration of the polymer solution is cR_g^3=0.001. | 2012.13827 | selectivity_nf.jpg |
0.499426 | c77ebd0babe84f16831c717b9669767c | The effect of modification of stickers on a polymeric 4-valent particle. Dark blue curve corresponds to unmodified polymer with valency κ=4 and linkers n=10. Other curves represent modified polymers. The colour represents modification pattern. Solid curves correspond to modification with stronger ligands and dashed curves correspond to modification with weaker ligands. Concentration of the polymer solution is cR_g^3=0.001. | 2012.13827 | sequence-n10.jpg |
0.534944 | 2c815693a5b74a70870c95be8f20202b | The effect of modification of stickers on a polymeric 4-valent particle with long linkers. Dark blue curve corresponds to unmodified polymer with valency κ=4 and linkers n=1000. Other curves represent modified polymers. The colour represents modification pattern. Solid curves correspond to modification with stronger ligands and dashed curves correspond to modification with weaker ligands. Concentration of the polymer solution is cR_g^3=0.001. | 2012.13827 | sequence-n1000.jpg |
0.431344 | 6c901716487d4f4b936e70bf8fddc385 | Schematic illustrating the "tail" and "loop" conformations | 2012.13827 | tail-loop.jpg |
0.493309 | 78444f26342d44daa1fc1ad811f88c1d | Err_ℛ,k as a function of for k=1, 3, 7, 10. | 2012.13833 | U_eps_neww.jpg |
0.431345 | 41906251f512443cb24a74916116fd0b | MNLI Matched | 2012.13838 | MNLI_degrad.jpg |
0.463745 | edcf96624ccf480c954059e6db61bfb3 | MNLI Mismatched | 2012.13838 | MNLI_degrad_mm.jpg |
0.465885 | 22faf3a1b4f24af49059c6d573b76bcc | AG News | 2012.13838 | agnews_degrad.jpg |
0.452388 | 9c647a22426945b2b2dc49a26f0f8fbb | IB with different β. | 2012.13838 | cross_beta.jpg |
0.466102 | 0b4dac3aa2bd4c41bfc3c21e9a3f20d4 | IB after different layers. | 2012.13838 | cross_layer.jpg |
0.397317 | caab8b28e87a45a1bca46ba13cdb0ba7 | Degradation test results across all layers. | 2012.13838 | cross_layer_all.jpg |
0.402371 | bfcaa8ed3d5340d7b76c0a68687ccb1f | IMDB | 2012.13838 | imdb_degrad.jpg |
0.474744 | 760ec33b57a84f6784eb707b2e5f3a5c | RTE | 2012.13838 | rte_degrad.jpg |
0.39787 | c94d8c599dc241869e332425e6180cce | β=10^-3 | 2012.13838 | vary_beta_1e-3.jpg |
0.4263 | b62a6efff3e04105b706150d978171e1 | β=10^-7 | 2012.13838 | vary_beta_1e-7.jpg |
0.431193 | 40d576291a58467da421c1b9af258fee | Nanoparticle-seeded GLAD. Top: GLAD of a source material (blue arrows) onto well-dispersed seeds (red spheres). Middle: nanoparticle-nanowire array on substrate. Bottom: individual heterostructure, detached via micromanipulator. | 2012.13840 | Fig1.jpg |
0.457631 | a0e72293db9f42868de15907ea9b4251 | Scanning electron micrographs of glancing angle-deposited tip-handle heterostructures. Nanoparticle dispersion on Si substrates: (a) Er_2O_3, (b) Fe@C, (c) Fe. Heterostructure arrays: (d) Er_2O_3 NP-Ge NW, (e) Fe@C NP-SiO NW, (f) Fe NP-FeCo NW. Single heterostructures: (g) Er_2O_3 NP-Ge NW, (h) Fe@C NP-SiO NW, (i,j) Fe NP-Ti NW, (k,l) Single Fe NP-FeCo NW. Scale bars (nm): 1000 (a,b,d,e), 200 (h), 100 (c,f,g,k,l), 50 (i,j). Insets are magnified (2x) and colorized, with red tips and blue handles. See Tab. S3 for quantification. | 2012.13840 | Fig2.jpg |
0.412327 | 64914ad0365b484299ba8e083daec3cd | Norms of weights of DQN for a couple of games from the Atari suite, all trained with identical hyperparameters. Norm growth differs not only for shuffled the labels, but also across more realistic datasets. One sees significant differences in the learning dynamics compared to image classification, the non-stationary nature of the dataset in RL might be partly responsible for this difference. | 2012.13841 | atari_norms.jpg |
0.388721 | 878a5ad6811f42a0841bc15793e6352f | Learning curves for various Atari games for λ = 0.001. | 2012.13841 | atari_scores_all.jpg |
0.482751 | 47a45b70381842c199b65c4e5e3a20ce | The gradient norm during training for a standard network, and a network with shuffled labels and no weight decay. The gradients have roughly the same order of magnitude, and later in the training shuffled labels typically leads to larger gradients. | 2012.13841 | batch_grad.jpg |
0.391055 | 7ac8017f8b324042a03a1f3f92ec5bbf | The leftmost figure illustrates the quantiles of log |m_i/w_i| during training of a transformer <cit.>. They vary roughly two orders of magnitude, suggesting that the gradients for different parameters differ substantially. The three following figures illustrate the translation quality, measured in bleu, for three different values of λ and three different weigh decay schemes. Standard WD underperforms unless λ is taken small whereas separating the buffers matches decoupled WD. | 2012.13841 | bleu_plus_q.jpg |
0.482822 | 3daceb2457504a55b644f22f56712103 | Quantiles of log| m_i/w_i| for weight w_i with buffer m_i as a function of time for DQN <cit.> trained on various Atari games. The gradient signal is small compared to the weights, but the ratio differs dramatically between parameters. | 2012.13841 | buf_over_weight.jpg |
0.493708 | f49e2f508b764618a7d34d13ee141667 | Learning curves for various Atari games with λ = 0.00000001. | 2012.13841 | cropped_scores_wd0d00000001.jpg |
0.430397 | 4d3947cfdbe44eaca3aa621b7649ebae | Learning curves for various Atari games with λ = 0.0000001. | 2012.13841 | cropped_scores_wd0d0000001.jpg |
0.470966 | 6d644c0e15aa4c789c8c4730eafa6bba | Learning curves for various Atari games with λ = 0.000001. | 2012.13841 | cropped_scores_wd0d000001.jpg |
0.459881 | ec38bd6bc2084f04bf34b1beacabd39d | Learning curves for various Atari games with λ = 0.00001. | 2012.13841 | cropped_scores_wd0d00001.jpg |
0.410535 | 2e8cff85c0b54b5b8a42ba28172cd415 | Learning curves for various Atari games with WD (λ = 0.0001). We compare decoupled WD <cit.>, original WD, no WD and an Adam variant with separate buffers for the WD and gradient signal. Original WD underperforms, whereas separating the buffers performs on par with decoupled WD. This suggests that the mixing of WD signal with the gradients, rather than the adaptivity itself, is responsible for the poor performance of normal WD in this setting. | 2012.13841 | cropped_scores_wd0d0001.jpg |
0.378034 | 9a7672f8dd0047d9a4683d4abf1a07a4 | The contributions of the cross and square term towards the norm growth for networks trained without batch normalization. The cross term dominates. | 2012.13841 | cross_vs_sq.jpg |
0.349321 | 4bfb9870441a4ff7a5281cb2832fc1bf | The contributions to the change in weight norm for the square and cross terms, defined in (<ref>). The cross term dominates and thus the norm grows primarily in the radial direction, scaling up subsets of the weights that align with the gradient. | 2012.13841 | cross_vs_square.jpg |
0.393057 | 0ac6550677974102881e02cc1ee8d151 | The distance from start for networks without batch normalization trained on the original and shuffled labels. The networks move significantly. | 2012.13841 | dist_from_start.jpg |
0.446262 | 3ce1e2c9cab74570af72b44d3d9d08ba | Cosines between the weights and the gradients for networks without batch normalization. The results are similar to networks with batch normalization. | 2012.13841 | grad_cosines.jpg |
0.404456 | acad099b47b54dd199e38b5ec26c08c8 | Learning curves for DNNs trained without WD with weights scaled to match norms in Fig. <ref>, and the original learning curves that these DNNs are made to match. Scaling the weights roughly matches the performance of various WD schedules, suggesting that WD mediates the observations of <cit.> through a simple scaling mechanism. | 2012.13841 | matching_appendix.jpg |
0.419582 | bf2c6740644e47ef8ad91eea1624f63a | The cosine between w and - ∇ℓ_neg for networks trained on shuffled labels. They are only negative, and the gradients w.r.t ℓ_neg do not contribute to norm growth along the radial direction. | 2012.13841 | neg_grad_cos_shuffled.jpg |
0.442574 | 8adb26a8f2cc4d9fa1d320253c0a1fe3 | The test accuracy and l2 norm of networks trained without batch normalization. Again, we see that using weight decay only at the start vastly outperforms using it after the start. | 2012.13841 | nobn_acc.jpg |
0.429285 | 4076a2fbcd1a4c358d39da2add323e19 | (Top.) <cit.> have shown that for image classification, starting WD only after epoch 50 brings little benefit, whereas stopping it after epoch 50 performs on par with using it throughout the training. (Bottom.) The l_2 norm of the weights increases dramatically at the start. Applying WD only during early parts of training ensures small weights throughout the optimization process. By applying WD late, it takes many epochs for the norm to shrink. We also plot curves for networks trained on datasets with shuffled labels and note that weight norms grow less under such settings. | 2012.13841 | norm_and_acc_3datasets.jpg |
0.402258 | 54678635f1894d48a2a0bfc2adecf7c3 | (Top.) We divide the cross entropy loss into two parts as per (<ref>). The cosine between the weight vector w and -∇ℓ_pos is positive whereas the cosine between w and -∇ℓ_neg is negative. This suggests that network norm increases as subset of weights responsible for correct predictions grow in magnitude. (Bottom.) cos(w, -ℓ_pos) with ℓ_pos defined as per (<ref>). We see that for a network with shuffled labels the gradient barely points in the radial direction, which would lead to less growth as per Figure <ref>. | 2012.13841 | pos_neg_std_shf.jpg |
0.398991 | f919d68f8a5740ccb7ef49822b38c86e | The sharpness of networks, typically used as a proxy for generalization, using different WD schemes. We compare four metrics of sharpness: the largest hessian eigenvalues (compute via <cit.>, measured logarithmically), the sharpness metric in <cit.> and additive/multiplicative perturbations. All metrics except multiplicative perturbations fail to consistently explain the differences in generalization of Figure <ref>. The loss under multiplicative perturbations increases when WD isn't used, suggesting that a sharp minima hypothesis might explain observations of <cit.>. | 2012.13841 | sharpness_metrics.jpg |
0.418232 | 0edb0c2266634003b366e18014722111 | The effect of applying wd only every 128 SGD update, which we call stuttered. The final accuracy is unchanged across architectures and datasets, presenting opportunities for computational savings. | 2012.13841 | stuttered_appendix.jpg |
0.400798 | a93f021007e34778bec62868de5ed429 | Schematic of STEM electrical field measurement setup (a) In this approach, an electrical potential is applied across a 2D material system leading to a deflection in the center of mass of the electron diffraction pattern. (b) This momentum transfer experienced by the incident electrons in the detector plane is used to calculate the electric field vector at each probe position. | 2012.13842 | Figure1.jpg |
0.395665 | ff5d6ecc68244d85be9e21f1ad5aab32 | DPC simulations for interrogating long-range electric fields (a-b) Simulated DPC images taken from hBN structure subjected to an varying electrical potential in the vertical direction detailed in Figure S4. The detector configurations are provided in the insets. (c) An intensity line profile taken from (a). (d-e) Simulated annular DPC images taken from same hBN structure with detector configurations are provided in the insets. (f) An intensity line profile taken from (c). From these line profiles, the simulated data suggests that intensity deviations from a uniform contrast profile resulting from atomic electric fields are significantly minimized in the annular geometry. | 2012.13842 | Figure2.jpg |
0.477945 | 68b099f42b6745d28dcbe3f7b8b995ba | Experimental scanning diffraction data acquired from the sample seen in center dark field TEM image. (a) CBED patterns captured at point 1 marked on the hBN sample. Colorbar intensity is in units of relative detection intensity. CBED patterns taken as a function of indicated positions when (b) 0V and (c) 5V are applied to the hBN sample. The purple dots refer to the center of mass of the CBED pattern taken from regions devoid of sample, while the yellow dots represent the center of mass of the CBED pattern taken from a particular position. The border color represents the degree of probe deflection in the x-direction.(d) CBED patterns captured at point 1 marked on the MoS_2/hBN sample. Colorbar intensity is in units of relative detection intensity. CBED patterns taken as a function of indicated positions when (e) 0V and (f) 5V are applied to the MoS_2/hBN sample. Scale bars represent 2 mrad. A clear shift in center of mass is observed when an external field is applied for both samples. | 2012.13842 | Figure3.jpg |
0.456404 | bbe315b57866469a9eb19c1d784f1c58 | Experimental electric field maps. In the absence of an applied bias, electric field maps of the (a) hBN and (c) MoS_2/hBN samples are provided. Similar field maps in the case where an external bias of 5V is applied are provided in (b) and (d) for the hBN and MoS_2/hBN samples, respectively. In the case of an applied field, the measured electric field vectors align with the external potential gradient. | 2012.13842 | Figure4.jpg |
0.436456 | be3b76f75b2a4ebe91c7b71bd3b5adba | Atomic force microscopy image taken from a flake of mechanically exfoliated hBN used as the support substrate. The flake thickness was found to be roughly 5 nm. | 2012.13842 | FigureS1.jpg |
0.396654 | 9fa7823c447e4ca1822514cdfbd82321 | (a) Colorized SEM image of suspended 2D heterostructure geometry. The gold region is indicative of gold electrodes. The purple region indicates the presence of h-BN, and the green region represents monolayer MoS_2, (b) Current (I_ds) - Voltage (V_ds) characteristics taken from sample indicate that an electric field can be applied across the sample, (c) Photoluminescence (PL) spectra taken from MoS_2/h-BN regions that are supported on SiO_2, supported on gold, and suspended. Raman spectrum taken from the same (d) MoS_2 and (e) h-BN regions are also provided. Due to the inherent differences in charge doping and strain, the position and magnitude of the PL peaks and the position of Raman modes vary with substrate. | 2012.13842 | FigureS2.jpg |
0.42583 | f60ce25ed0c3499aa5ed2ab9ee32153d | Various challenges arise when trying to construct an in situ TEM sample of a 2D heterostructure when using other methods. In this case, the electrodes were first transferred onto the TEM windows and 2D layers were sequentially transferred on top. Unfortunately, the repeated thermal processing involved in this setup prevents successful material suspension and creates cracking and fracturing of the metal electrodes. | 2012.13842 | FigureS3.jpg |
0.42887 | f8b1d06c2213473d9d7ca64e859da55f | (a) Projected potential of structure used for multislice simulation. (b) External potential applied to simulated structure. Simulated dark field images taken from structure with (c) and without (d) external potential. Average CBED taken from structure with (e) and without (f) external potential. The scale bar is consistent for the real space images in a-d as well as the diffraction space images in e-f | 2012.13842 | FigureS4.jpg |
0.466641 | 601c88914b9d461aae226a2dbf9f083b | (a) Normalized intensity line profiles taken from same region in Figure <ref> for inner collection semi-angles of 0 mrad, 15 mrad, 25 mrad, 28 mrad, and 30 mrad. (b) Normalized intensity lines profiles taken from same region in Figure <ref> replotted for inner collection semi-angles of 25 mrad, 28 mrad, and 30 mrad. Minimal intensity variation from atomic electrostatic fields is observed for both 28 mrad and 30 mrad inner collection angles. 28 mrad was selected to provide sufficient latitude in terms of beam and detector alignment in practical situations. | 2012.13842 | FigureS5.jpg |
0.437096 | d69660939cd94861bb8f9c9474e9da75 | TEM selected area diffraction pattern taken from MoS_2/hBN sample indicates that the two layers are present and offset by 5^∘. | 2012.13842 | FigureS6.jpg |
0.365591 | a777f4b128d840ae86e0c44261a90472 | Difference in experimental CBED patterns taken from MoS_2/hBN region at position 1 on Figure <ref> between V_ds=0V and V_ds=5V conditions. This colormap indicates that when the probe is smaller than the feature, the ICOM signal is measured from the shifts of the unscattered beam as opposed to asymmetries within the disc. | 2012.13842 | FigureS7.jpg |
0.406831 | 63e72cc158944435bb54931d83072d49 | Simulated relative electric field intensities in the sample as well as a 1μm vacuum region directly above and below the sample plane. For this simulation, boundary condition values of 0V at the ground electrode and a positive voltage at the source electrode (Vds) were used. The Laplace equation was solved to determine the electric potential at each finite location. By taking the in-plane spatial gradient of these values, the in-plane electric field distribution was calculated. | 2012.13842 | FigureS8.jpg |
0.410023 | f89c2f1e83d64216805cae9296520d5e | Charge density map across MoS_2/hBN region calculated from electric field map in Figure <ref>d. | 2012.13842 | FigureS9.jpg |
0.449994 | 6ba1622ecbdb4698a2f6b6272c27b99e | The accumulative layer cost ratio (ALCR) of compute time, activation size and parameter size for each layer of VGG16-BN model on Titan XP. | 2012.13846 | compute_time_cdf.jpg |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.