nopperl/arxiv-image-text · Datasets at Hugging Face

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
0.430815	b33fe57819794079b6fb36c446b65d49	Transmittance through (air/silver/air)^n SLs as a function of the frequency ω and the conductor-layer width d_c. In a) and c), the number of unit cells n is odd, while in b) and d), it is even. The difference in the behavior of T_n below and above ω_p is clear in the upper panels. The line shapes of resonant plasmons with ω>ω_p are thinner, therefore more localized, than the plasmon frequencies at low frequencies. We assume that the latter result from a complex coupling of interface plasmons. A parity effect is also clear for d_c larger than ∼50nm. At larger conductor widths, the low frequency transmission either vanishes, for n odd, or tends asymptotically to one, for n even. In the lower-panel graphs, we plot, together with the transmission coefficient, the dispersion relation of Equation (<ref>), and we see that, if n is odd, the number of plasmon resonances is the same for SLs with n and with n+1 unit cells.	2012.10546	Fig6e.jpg
0.44484	feb76a36c9eb4bffbbc09508ec7a7d65	A metamaterial superlattice air(L R_2)^nair with n unit cells, where left- and right-handed media alternate. We assume the normal incidence of the electromagnetic field with parallel polarization.	2012.10546	Fig7e.jpg
0.423578	90686c690db44107b69e8f1f47391d0c	The transmission coefficients for two superlattices air(RR)^n air (upper panel) and air(LR)^n air (lower panel), with the same parameters except for the signs of the refraction indices ϵ_1 and μ_1. For the upper panel, we have n=6, n_1=2.22, n_2=1.41, d_1=79nm, and d_2=497nm. For the lower panel, we have n=6, n_1=-2.22, n_2=1.41, d_1=79nm, and d_2=497nm.	2012.10546	Fig8e.jpg
0.439495	11887e1d65c647ebb91c054304c6a308	Transmission coefficients (TCs) for the superlattices air(RR)^6 air and air(L R_2)^6air, as functions of the frequency ω, when d_L=d_1=d_2 n_2/n_L=316.5nm and d_2 =497nm. The TCs of the superlattices air(LR)^6 air become a sequence of isolated and equidistant peaks, while the TCs of the superlattices air(RR)^6 air become a periodic sequence of resonant bands. This characteristic is independent of the number of unit cells.	2012.10546	Fig9e.jpg
0.488613	741626baa54a45cd98866021793b22a8	In this figure, we plot the transmission coefficient for the quarter lambda in the manuscript and for other relations, indicated in the graphs. The resonant transmission coefficient (blue curves) are plotted together with the dispersion relations for an infinite periodic structure (black curves) and with the dispersion relation of the theory of finite periodic systems (red lines). In (a), the narrow bands in the quarter lambda limit agree with the predictions of the Kramer condition and with those of the resonant dispersion relation. In the other panels, we find again that the resonant dispersion relation predicts the resonances of the transmission coefficients.	2012.10546	QuarterLambdae.jpg
0.404514	50f38c0557e14c8eb3d6e331f908d1e1	Framework overview of our proposed efficient approach for training deep neural networks from multiple data sources.	2012.10547	overview_framework.jpg
0.445543	a0228ee53c2248439cfa38f450e1d4f1	Illustration of secure two-party computation approaches between the client pool and the server.	2012.10547	overview_s2pc.jpg
0.448676	2aa584ad15c1486db3ffb8d4527fe5cf	ROC curves showing the trade-off between MMPMR and FRR.	2012.10548	MMPMR_FRR_ROC.jpg
0.470078	811c489c1e104a6cbe9f958b7e3e7e23	Sketch of the resonance overlap criterion <cit.>.	2012.10552	fig11a.jpg
0.522872	94601e66d5ac48e08ba320a0895a7dbe	Distribution of the minimum DA for 60 different machine configurations at the end of leveling for both baseline and adaptive (i.e. lower crossing angle compared to the baseline in order to meet the DA requirement) scenarios and the two leveled luminosity targets. The worst configuration is found at less than 1σ_beam from the statistical average, with the overall spread in the range of 0.2-0.3σ_beam <cit.>.	2012.10552	fig13.jpg
0.434988	b8877b76a268434192c4a956f955c760	Optimization of the ALS lattice using MOGA with the solution front in the objectives space, composed by the horizontal emittance versus DA area, colour-coded with the horizontal beta function <cit.>.	2012.10552	fig14.jpg
0.444205	c86cbeca35be4dbbacf392fc4441ba7c	LHC run2 working point scan color-coded with the DA <cit.>.	2012.10552	fig15.jpg
0.53371	76b07b8c4e1b49259e1aaeb20095470a	Intensity of Beam1 (blue) and Beam 2 (red) for consecutive fills of the LHC <cit.>.	2012.10552	fig16.jpg
0.473689	88da26d8a3954866b1d0e80acc7b2181	Sketch of the phase space in the vicinity of a third order resonance.	2012.10552	fig6.jpg
0.420388	3cfdf934ac794a138fb2c0c16f7f4aac	Distributions of matching scores for mated image sets within the real or synthetic datasets indicated in the legend.	2012.10553	05_SD_IVI_FRR_3.4_degrees.jpg
0.386717	72fc1c6ef564452b9c2557d948ea429b	ROC curves for the identity-disentangled datasets.	2012.10553	ROCs.jpg
0.451609	087d350ed97c4a6db0afbfd7765f7bc7	The SD-GAN architecture used in our evaluation. 𝐳_ID, 𝐳_IV^1 and 𝐳_IV^2 are random vectors; the dotted line indicates shared network weights and the plus indicates channel-wise concatenation of generated images. (IV refers to “Intra-class Variation”.)	2012.10553	SD-GAN_diagram.jpg
0.42801	174f06184f72481db5b25aab2dd33341	Effective potential for p=3, θ=0, m>0, and M>0. The bounce solution renders the false vacuum unstable. The blue and green curves coincide with the branches in Figure <ref>.	2012.10555	bounceplot.jpg
0.468863	db0e6e8fd04d40a6842dce559838dea8	The θ-dependence of the chiral condensate ψ̅ψ∼cosφ in massless Schwinger models with modified instanton sums characterized by an integer p. The unmodified theory has p=1.	2012.10555	condensate.jpg
0.435345	a2b7b849de6147e98369527ea1153084	Effective potential for bosonized scalars φ,η with p=3, m =0, M>0. The four curves coincide with the four branches in Figure <ref>.	2012.10555	plottop.jpg
0.452024	8090b6b8c0504fe8b14724e15b7500f6	Domain wall configuration between two chiral symmetry breaking vacua for p=3, θ=0. Rather than pass through the cusp, the domain wall trajectory must go above the barrier, whose height scales as M^2. The curves coincide with the four branches in Figure <ref>.	2012.10555	tunnelplot.jpg
0.470303	e62607812fce4d85a16723d39b9f0355	A factor analysis of that shows the impact of removing dynamic resource allocation (Ekya-FixedRes) or removing retraining configuration adaptation (Ekya-FixedConfig).	2012.10557	ablation_cityscapes_2.jpg
0.495195	83efcb8f5e794818a1c9bc29ca1b1aa4	factor analysis by removing dynamic resource allocation and reducing the hyperparameter space.	2012.10557	ablation_cityscapes_golden.jpg
0.453459	8c7c09a39fff47019d91f3ebc36b923a	Urban Traffic	2012.10557	bellevue_10cam_scheduler_comparison_across_resources.jpg
0.499544	036f223dd63a43888bc1fdc864844077	's profiler can predict the model accuracy after retraining with less error than the baseline of moving average over recent history windows.	2012.10557	cdf-profiler-error.jpg
0.500591	28e4e00328d74de6afc5f8fc0c0c7d93	Cityscapes	2012.10557	cityscapes_across_resources.jpg
0.445701	b323cbb638e14b858c7f484a597741b5	Cityscapes	2012.10557	cityscapes_scheduler_comparison_across_resources.jpg
0.518086	3661b943ea04473b9f2e50d9dcf5ff53	Resource-accuracy	2012.10557	jena.jpg
0.442078	20bcb57f53bf49b9951b79c3c5730758	Effect of Hyperparameters	2012.10557	jena_hypparams.jpg
0.465334	ccdce70385f4432db4cd41beab298978	Urban Building	2012.10557	las_vegas_scheduler_comparison_across_resources.jpg
0.500124	bdd442655efe421eb06ab3b447bfcc5c	Video stream #1 (Inference accuracy = 0.82)	2012.10557	long_video_gpu_allocation_vs_time_cam1.jpg
0.504088	f40c1f1b6ce9417f93542a7887a83782	Video stream #2 (Inference accuracy = 0.83)	2012.10557	long_video_gpu_allocation_vs_time_cam5.jpg
0.461507	b0359cf565e24bc1989285e459290c37	Distribution of accuracy estimation errors.	2012.10557	microprofiling_accerror.jpg
0.508068	16b7c837c4f0455aa9fdd1aed7c6d41c	Accuracy	2012.10557	motivation_cityscapes_zurich_accuracy.jpg
0.418152	02b5ed293e674c95aa50d7e1bc38dba3	Class Distribution	2012.10557	motivation_cityscapes_zurich_classdist.jpg
0.433389	12161fb33b5c48e9930d74e0f7ca8802	Window 1 snapshot	2012.10557	motivation_cityscapes_zurich_win_1_5.jpg
0.446631	1ca409181cd44cf5823efc2e9539d54a	Window 9 snapshot	2012.10557	motivation_cityscapes_zurich_win_6_10.jpg
0.477568	67a43a332dc945e8a1ddb9b2fce1433a	Cityscapes	2012.10557	multicam_acc_vs_res_cityscapes.jpg
0.455402	f5c7b7365a0649d29938dfc8a3186720	Cityscapes	2012.10557	multicam_acc_vs_res_sysimpl_cityscapes.jpg
0.417253	c542ce22b4c146729e0b8e140570dafb	Waymo	2012.10557	multicam_acc_vs_res_waymo.jpg
0.515842	5aeac0c7fa1248d48aecad4849a30126	Cityscapes	2012.10557	multicam_individual_stream_acc_cityscapes.jpg
0.521518	466c36de7e164e6b80f6b9b50ca7945b	Waymo	2012.10557	multicam_individual_stream_acc_waymo.jpg
0.522924	c22ebc999e9f4444a406be9cfeca226c	Timeseries of our performance vs. fair sharing. Again, adaptive resource sharing constantly outperforms fair sharing. Accuracy drops over time.	2012.10557	multicam_taskwise_acc_zurich_4_cityscapes.jpg
0.445298	d90a353a53c04a2ab28d8326e09f7da5	's components and their interactions.	2012.10557	overview_cropped.jpg
0.42609	068398d1838f403e9b79de9b69bb794f	Throughput measured as number of video streams that can be concurrently supported for different resource sizes subject to accuracy thresholds.	2012.10557	scalability_GPUs_cams_cityscapes.jpg
0.463213	eb2b63c2c6444f5d8a21358943b1dcd1	Using golden model. Effect of adding more video streams on accuracy for different schedulers. When insufficient resources are provisioned, gracefully degrades to the No-retrain baseline, while the retraining decisions made by the fair scheduler reduce the mean inference accuracy.	2012.10557	scalability_cams_fixedGPUs_cityscapes_golden_model.jpg
0.487483	6aac994f176e4ee6af65f60eb6e2e9bc	Uniform scheduler	2012.10557	schedmot_multiwindow_eventual_best_cfgs.jpg
0.438354	069c1d10deaf46d0a7f5b1c2178e026f	Accuracy-optimal sched.	2012.10557	schedmot_multiwindow_prioritization_and_optimal_cfgs.jpg
0.401534	078ca6a1424f403c8e04e86962738ef7	Naive, eventual best cfgs	2012.10557	schedmot_res_eventual_best_cfgs.jpg
0.397626	75526972238843d69e26c71577c9a7d8	Optimal cfgs with prioritization	2012.10557	schedmot_res_prioritization_and_optimal_cfgs.jpg
0.38549	7ac5d18ed42d4a128185b34c96f5729d	We add a controlled amount of error to the accuracy prediction and our performance does degrade but only marginally. Pick 5 cities, horizontal graph	2012.10557	sensitivity-accuracy-error.jpg
0.488915	a99269e10b664211bba0dd9a3f686f76	Adding a controlled error ϵ to the accuracy prediction - 's performance degrades, but only marginally.	2012.10557	sensitivity_profileerrors_cityscapes.jpg
0.435815	9355db4bb11f449bb89f9e2d44093f86	Single-video performance: Ours (with continuous retraining) vs. various baseline DNN configurations (without retraining). Retraining allows ours to achieve much better accuracy-cost tradeoffs.	2012.10557	singlecam_acc_vs_cost_cityscapes.jpg
0.43237	7f57d2c77ed24b76a743a5fbd4cc2419	Cityscapes	2012.10557	singlecam_acc_vs_cost_golden_cityscapes.jpg
0.407254	f9591bccdce7459289770cabec841e9b	Waymo	2012.10557	singlecam_acc_vs_cost_golden_waymo.jpg
0.432279	0be4e037abb9407886e69d3fa086cded	Cityscapes	2012.10557	sysimpl_varyingcities_streams_cityscapes.jpg
0.487651	8f9a776a4bee4a5b95db73bff21f0c73	Waymo	2012.10557	sysimpl_varyingcities_streams_waymo.jpg
0.371886	efebc7e0b56d4e308f471a523c7ded2a	Waymo	2012.10557	sysimpl_varyingcities_waymo.jpg
0.513173	2a9c9b283af44da0953ca5f40ca77d7b	Waymo	2012.10557	waymo_across_resources.jpg
0.445798	3d1919937f2f4f07968a7a80e7772928	Waymo	2012.10557	waymo_scheduler_comparison_across_resources.jpg
0.406301	a155493c1ae14ac0875d0f7d5239450c	Cameras connect to the edge server (e.g., Azure Stack Edge <cit.>) through local networks (wireless, typically). The edge server is equipped with consumer-grade GPUs and executes DNN inference containers and retraining containers.	2012.10557	xxx_cropped.jpg
0.445986	abf1934439ca43fdbbaf4b22cc5a6654	Number of cliques of size ℓ∈{4, 5,6 } for the Indian village data set. The median values for ℓ = 4, 5, and 6 are 132, 13, 0, respectively.	2012.10559	Number_of_Cliques-eps-converted-to.jpg
0.464953	19ff601114bd46ed87d42b6bdb313a53	We generate a graph using a 3-dimensional Euclidean latent space with K = 10 cliques. We plot the scree function ϕ and the bootstrap variability function f_n defined in Algorithm <ref>. We also plot their sum, defined as the objective function. The horizontal axis represents the possible ranks of the matrix. We see the objective function has a minimum at 3, so we estimate the rank of the matrix to be 3, which is the true dimension of the latent space.	2012.10559	Rank_Estimate-eps-converted-to.jpg
0.445603	f53e9d9bedf049039d534118e19e3b83	Predicted dimensions for the predicted geometries of the Indian village networks. Most predictions are low-dimensional, usually between 1 and 3. Some of the predictions are higher. In these cases, our method is predicting that W_κ is full rank.	2012.10559	Rank_Estimates-eps-converted-to.jpg
0.440232	410d921d931a491b87f9c887afc6d7b2	Regression coefficients showing the relationship between loans and geometry. Each line in the plot corresponds to the coefficient in a multivariate linear regression where the outcome is the average amount of loans (in thousands of INR) and the predictors are geometry types (with spherical as the reference). The wide bars correspond to one standard error and the narrow bars represent two standard errors. The reference value for spherical is 16.71 (again in thousands of INR).	2012.10559	loans_coef.jpg
0.383797	db7c0a705a254e938a8097561a8177e7	Client-server architecture for remote visualisation.	2012.10560	O1-22_f1.jpg
0.467524	ef574b8644884586816d4402faeac3b3	IOI Model in Layer-1 of Anonymous Cryptocurrency	2012.10563	anonymous_cryptocurrency_ioi_model.jpg
0.47348	50a7317b39fa4ca3acad6cce0778d13d	Architecture Model for Anonymous Cryptocurrency	2012.10563	architecture_of_anonymous_cryptocurrency_02.jpg
0.487384	5870918c23d34c9786e78be3d3549ded	IOI Model in Layer-1+0 of Anonymous Cryptocurrency	2012.10563	cryptocurrency_ioi_model_ipaddress3.jpg
0.489403	717ca7c8b3b740d68f120a4fa6ecc639	Pfitzmann's Definition of Anonymity	2012.10563	pfitzmann_anonymity.jpg
0.4483	e2ae7477373c420ab270f377644e9ba3	Pfitzmann's Definition of Unlinkability	2012.10563	pfitzmann_unlinkability.jpg
0.416721	a6a5c37140c64e07a9177ec8049e3f63	SCCLA collusion	2012.10563	sccla_collusion.jpg
0.434043	48d764aebdf84849a20d0e58a0f85a82	Zerocash Coin Structure (This figure is cited from <cit.>)	2012.10563	zerocash_coin_structure.jpg
0.423599	1420206bf8564973a78e39451d21d03d	Wavelet scattering coefficients	2012.10564	chestxray-scattering-comparison.jpg
0.46405	e7ed0d7d9fbc41ac929ccd9666266455	Empirical MMD distributions under null (ℋ_0) and alternative (ℋ_1) hypothesis between the ChestXray14 and MIMIC-CXR data sources. We used scattering transform in composition with the radial basis function (RBF) as kernel function in order to find the B-test statistics. We set to compute up to the second-order of the scattering coefficients and set other filter parameters and . We summed the scattering coefficients over the image domain to obtain a translational invariant 417 features. We set the RBF scale parameter γ=1.	2012.10564	cxr-btest-sc-noadap.jpg
0.503843	ed31889e11774ec3bfd6fd103d362119	ChestXray14	2012.10564	cxr_chestxray_sc_hist2d-cyclegan.jpg
0.467388	491fcc5ae72443679bd94c7a5a39b05a	ChestXray14	2012.10564	cxr_chestxray_sc_hist2d-noadap.jpg
0.417871	03d8a63f5f96449a8babc00700f3ad0e	Six training loss of , trained between the samples of and , are displayed for 14 epochs. We denote two discriminator losses by and , two generator losses by and , and two cycle-consistency losses by and .	2012.10564	cxr_cyclegan_training_loss.jpg
0.421638	26bf11dc2a504eab8ff63f8605e7bc69	MIMIC-CXR	2012.10564	cxr_mimic_sc_hist2d-cyclegan.jpg
0.449009	b6fc41a1347a4d58abab0d6865e6d0f3	MIMIC-CXR	2012.10564	cxr_mimic_sc_hist2d-noadap.jpg
0.435761	5cf1f14325cb44ff8ce69466ee422030	Differences in the abnormality prediction scores obtained by on and style-adjusted by is binned into non-overlapping intervals, and the counts in every interval are displayed.	2012.10564	cxr_preds_compare_hist.jpg
0.499476	1912f8606bcf4e01b7f790bc9ab5e13e	CycleGAN contains mappings between two domains G : X → Y and F : Y → X, and one discriminator for each domain, D_X and D_Y. The purpose of including discriminators is to encourage the generators G and F to generate samples that can not be indistinguished with the available real samples from the two domains. Additionally, CycleGAN introduced cycle consistency losses to enforce forward and backward cyclic consistency between the generators, i.e., x → G(x) → F(G(x)) ≈ x, and y → F(y) → G(F(y)) ≈ y. We have included this figure into this manuscript from the CycleGAN paper <cit.>.	2012.10564	cyclegan.jpg
0.478027	ebd07631a67c4fc3b7e42b08b125bb13	Clinical workflow for computer-aided abnormality detection in chest radiographs	2012.10564	domain-adaptation-workflow.jpg
0.405248	8a81b57ad6de4ea685414b556b54c920	Overview of our approach. Our system takes an input image, a mask of the object to remove, and renderings of an approximate scene model with and without the object. We first perform an intrinsic decomposition of the input image into texture and lighting. We then remove the object's shadows from the lighting image. We inpaint the object mask region separately in both texture and lighting before recompositing to get our final result.	2012.10565	architecture_horizontal_3.jpg
0.532693	28201e83b63645c8af0b6272e6517698	Comparison of the dynamic temperature profiler sensor offsets across the four calibration campaigns. The offsets are consistent and the small differences are attributed to local variations during movement of the profiler. The one outlier (October 2018 calibration of the 6th sensor from bottom) eventually converged to give consistent readings in subsequent calibration campaigns.	2012.10567	plot2A.jpg
0.383955	b63e58684bb24dda9cc24508935b72be	The dynamic temperature profiler sensor values compared to the CFD simulation prediction, as a function of height from the detector bottom.	2012.10567	plot3A.jpg
0.444508	d53e7432bd7d44b48cbba71b0be072fc	Dynamic temperature profiler sensor values as a function of time. The y-axis values represent the differences in temperature of each sensor with respect to the bottom-most reference sensor.	2012.10567	plot4A.jpg
0.481229	144c30d1461e4a91b3900b613d9d5fba	The dynamic temperature profiler sensor values as a function of height from the bottom of the detector. Most of the values are within a 3 mK band for sensors immersed in liquid-argon. Beyond 7 m the sensors are in the gas-liquid interface.	2012.10567	post_1A.jpg
0.429246	582936d69a804844a14be212c83e7f68	Example of a science observation reduction page in the JSA processing tracking system. A JCMT calibration observation of N207IR taken by the instrument HARP in 2016 is shown.	2012.10568	Data-QA-pages-HARP.jpg
0.471871	31e8f3a6b34d4653b4c4d85a1f78f1f6	Observatory hours broken down by category. Science, Calibration, Engineering and Commissioning(E&C), Fault, Weather (time when observatory was closed due to weather), and Shutdown (time when observatory was closed). The months displayed are 1) four months (March - June 2019) prior to shutdown due to the restricted access on Maunakea when the JCMT was operated from the summit. 2) the five months after the return to operations (June - October 2020) from global pandemic shutdown when the JCMT was operated remotely. The time gap and switch to Remote Operations are indicated by the dashed vertical line.	2012.10568	Hours-EO-night-since-March2019.jpg
0.479009	17054fab6bf0478bbd819c5973fe6499	First night of Remote Operations on November 1st 2019 from the JCMT Remote Control room (JROC) by JCMT TSS Mimi Fuchs.	2012.10568	first-night-remote-ops.jpg
0.526928	1111bdbb89144f95839fe64bc5433745	Injection of flow from a slot into a cross flow	2012.10572	F01.jpg
0.477801	560cdd3dfc4a475d9b161a22b5af8d51	Two-phase fluid	2012.10572	F010.jpg
0.498633	ffd070c9912f4db39eece85b0af39dcc	The critical case Q=0	2012.10572	F011.jpg
0.521604	1a23a63008af41869dd41b8130dd10cc	Case 1	2012.10572	F012.jpg
0.492657	2898fb12761840fc8d0e3dd2d36f1726	Case 2	2012.10572	F013.jpg
0.470916	b02bc994ad214ecbb2786c27782d4fea	Horizontal blade surfaces	2012.10572	F02.jpg
0.478203	c8b98c37850e4afdaf557f913b275b3a	Figure 9.3 in <cit.>	2012.10572	F03.jpg
0.488139	737650f68e6c4cd38807525e288e9dfc	Oblique injection of flow into a free stream	2012.10572	F04.jpg
0.516258	d8187058ce8f43869dfc732c14df7ef9	The nozzle walls and the ground	2012.10572	F05.jpg

0.430815

b33fe57819794079b6fb36c446b65d49

Transmittance through (air/silver/air)^n SLs as a function of the frequency ω and the conductor-layer width d_c. In a) and c), the number of unit cells n is odd, while in b) and d), it is even. The difference in the behavior of T_n below and above ω_p is clear in the upper panels. The line shapes of resonant plasmons with ω>ω_p are thinner, therefore more localized, than the plasmon frequencies at low frequencies. We assume that the latter result from a complex coupling of interface plasmons. A parity effect is also clear for d_c larger than ∼50nm. At larger conductor widths, the low frequency transmission either vanishes, for n odd, or tends asymptotically to one, for n even. In the lower-panel graphs, we plot, together with the transmission coefficient, the dispersion relation of Equation (<ref>), and we see that, if n is odd, the number of plasmon resonances is the same for SLs with n and with n+1 unit cells.

2012.10546

Fig6e.jpg

0.44484

feb76a36c9eb4bffbbc09508ec7a7d65

A metamaterial superlattice air(L R_2)^nair with n unit cells, where left- and right-handed media alternate. We assume the normal incidence of the electromagnetic field with parallel polarization.

2012.10546

Fig7e.jpg

0.423578

90686c690db44107b69e8f1f47391d0c

The transmission coefficients for two superlattices air(RR)^n air (upper panel) and air(LR)^n air (lower panel), with the same parameters except for the signs of the refraction indices ϵ_1 and μ_1. For the upper panel, we have n=6, n_1=2.22, n_2=1.41, d_1=79nm, and d_2=497nm. For the lower panel, we have n=6, n_1=-2.22, n_2=1.41, d_1=79nm, and d_2=497nm.

2012.10546

Fig8e.jpg

0.439495

11887e1d65c647ebb91c054304c6a308

Transmission coefficients (TCs) for the superlattices air(RR)^6 air and air(L R_2)^6air, as functions of the frequency ω, when d_L=d_1=d_2 n_2/n_L=316.5nm and d_2 =497nm. The TCs of the superlattices air(LR)^6 air become a sequence of isolated and equidistant peaks, while the TCs of the superlattices air(RR)^6 air become a periodic sequence of resonant bands. This characteristic is independent of the number of unit cells.

2012.10546

Fig9e.jpg

0.488613

741626baa54a45cd98866021793b22a8

In this figure, we plot the transmission coefficient for the quarter lambda in the manuscript and for other relations, indicated in the graphs. The resonant transmission coefficient (blue curves) are plotted together with the dispersion relations for an infinite periodic structure (black curves) and with the dispersion relation of the theory of finite periodic systems (red lines). In (a), the narrow bands in the quarter lambda limit agree with the predictions of the Kramer condition and with those of the resonant dispersion relation. In the other panels, we find again that the resonant dispersion relation predicts the resonances of the transmission coefficients.

2012.10546

QuarterLambdae.jpg

0.404514

50f38c0557e14c8eb3d6e331f908d1e1

Framework overview of our proposed efficient approach for training deep neural networks from multiple data sources.

2012.10547

overview_framework.jpg

0.445543

a0228ee53c2248439cfa38f450e1d4f1

Illustration of secure two-party computation approaches between the client pool and the server.

2012.10547

overview_s2pc.jpg

0.448676

2aa584ad15c1486db3ffb8d4527fe5cf

ROC curves showing the trade-off between MMPMR and FRR.

2012.10548

MMPMR_FRR_ROC.jpg

0.470078

811c489c1e104a6cbe9f958b7e3e7e23

Sketch of the resonance overlap criterion <cit.>.

2012.10552

fig11a.jpg

0.522872

94601e66d5ac48e08ba320a0895a7dbe

Distribution of the minimum DA for 60 different machine configurations at the end of leveling for both baseline and adaptive (i.e. lower crossing angle compared to the baseline in order to meet the DA requirement) scenarios and the two leveled luminosity targets. The worst configuration is found at less than 1σ_beam from the statistical average, with the overall spread in the range of 0.2-0.3σ_beam <cit.>.

2012.10552

fig13.jpg

0.434988

b8877b76a268434192c4a956f955c760

Optimization of the ALS lattice using MOGA with the solution front in the objectives space, composed by the horizontal emittance versus DA area, colour-coded with the horizontal beta function <cit.>.

2012.10552

fig14.jpg

0.444205

c86cbeca35be4dbbacf392fc4441ba7c

LHC run2 working point scan color-coded with the DA <cit.>.

2012.10552

fig15.jpg

0.53371

76b07b8c4e1b49259e1aaeb20095470a

Intensity of Beam1 (blue) and Beam 2 (red) for consecutive fills of the LHC <cit.>.

2012.10552

fig16.jpg

0.473689

88da26d8a3954866b1d0e80acc7b2181

Sketch of the phase space in the vicinity of a third order resonance.

2012.10552

fig6.jpg

0.420388

3cfdf934ac794a138fb2c0c16f7f4aac

Distributions of matching scores for mated image sets within the real or synthetic datasets indicated in the legend.

2012.10553

05_SD_IVI_FRR_3.4_degrees.jpg

0.386717

72fc1c6ef564452b9c2557d948ea429b

ROC curves for the identity-disentangled datasets.

2012.10553

ROCs.jpg

0.451609

087d350ed97c4a6db0afbfd7765f7bc7

The SD-GAN architecture used in our evaluation. 𝐳_ID, 𝐳_IV^1 and 𝐳_IV^2 are random vectors; the dotted line indicates shared network weights and the plus indicates channel-wise concatenation of generated images. (IV refers to “Intra-class Variation”.)

2012.10553

SD-GAN_diagram.jpg

0.42801

174f06184f72481db5b25aab2dd33341

Effective potential for p=3, θ=0, m>0, and M>0. The bounce solution renders the false vacuum unstable. The blue and green curves coincide with the branches in Figure <ref>.

2012.10555

bounceplot.jpg

0.468863

db0e6e8fd04d40a6842dce559838dea8

The θ-dependence of the chiral condensate ψ̅ψ∼cosφ in massless Schwinger models with modified instanton sums characterized by an integer p. The unmodified theory has p=1.

2012.10555

condensate.jpg

0.435345

a2b7b849de6147e98369527ea1153084

Effective potential for bosonized scalars φ,η with p=3, m =0, M>0. The four curves coincide with the four branches in Figure <ref>.

2012.10555

plottop.jpg

0.452024

8090b6b8c0504fe8b14724e15b7500f6

Domain wall configuration between two chiral symmetry breaking vacua for p=3, θ=0. Rather than pass through the cusp, the domain wall trajectory must go above the barrier, whose height scales as M^2. The curves coincide with the four branches in Figure <ref>.

2012.10555

tunnelplot.jpg

0.470303

e62607812fce4d85a16723d39b9f0355

A factor analysis of that shows the impact of removing dynamic resource allocation (Ekya-FixedRes) or removing retraining configuration adaptation (Ekya-FixedConfig).

2012.10557

ablation_cityscapes_2.jpg

0.495195

83efcb8f5e794818a1c9bc29ca1b1aa4

factor analysis by removing dynamic resource allocation and reducing the hyperparameter space.

2012.10557

ablation_cityscapes_golden.jpg

0.453459

8c7c09a39fff47019d91f3ebc36b923a

Urban Traffic

2012.10557

bellevue_10cam_scheduler_comparison_across_resources.jpg

0.499544

036f223dd63a43888bc1fdc864844077

's profiler can predict the model accuracy after retraining with less error than the baseline of moving average over recent history windows.

2012.10557

cdf-profiler-error.jpg

0.500591

28e4e00328d74de6afc5f8fc0c0c7d93

Cityscapes

2012.10557

cityscapes_across_resources.jpg

0.445701

b323cbb638e14b858c7f484a597741b5

Cityscapes

2012.10557

cityscapes_scheduler_comparison_across_resources.jpg

0.518086

3661b943ea04473b9f2e50d9dcf5ff53

Resource-accuracy

2012.10557

jena.jpg

0.442078

20bcb57f53bf49b9951b79c3c5730758

Effect of Hyperparameters

2012.10557

jena_hypparams.jpg

0.465334

ccdce70385f4432db4cd41beab298978

Urban Building

2012.10557

las_vegas_scheduler_comparison_across_resources.jpg

0.500124

bdd442655efe421eb06ab3b447bfcc5c

Video stream #1 (Inference accuracy = 0.82)

2012.10557

long_video_gpu_allocation_vs_time_cam1.jpg

0.504088

f40c1f1b6ce9417f93542a7887a83782

Video stream #2 (Inference accuracy = 0.83)

2012.10557

long_video_gpu_allocation_vs_time_cam5.jpg

0.461507

b0359cf565e24bc1989285e459290c37

Distribution of accuracy estimation errors.

2012.10557

microprofiling_accerror.jpg

0.508068

16b7c837c4f0455aa9fdd1aed7c6d41c

Accuracy

2012.10557

motivation_cityscapes_zurich_accuracy.jpg

0.418152

02b5ed293e674c95aa50d7e1bc38dba3

Class Distribution

2012.10557

motivation_cityscapes_zurich_classdist.jpg

0.433389

12161fb33b5c48e9930d74e0f7ca8802

Window 1 snapshot

2012.10557

motivation_cityscapes_zurich_win_1_5.jpg

0.446631

1ca409181cd44cf5823efc2e9539d54a

Window 9 snapshot

2012.10557

motivation_cityscapes_zurich_win_6_10.jpg

0.477568

67a43a332dc945e8a1ddb9b2fce1433a

Cityscapes

2012.10557

multicam_acc_vs_res_cityscapes.jpg

0.455402

f5c7b7365a0649d29938dfc8a3186720

Cityscapes

2012.10557

multicam_acc_vs_res_sysimpl_cityscapes.jpg

0.417253

c542ce22b4c146729e0b8e140570dafb

Waymo

2012.10557

multicam_acc_vs_res_waymo.jpg

0.515842

5aeac0c7fa1248d48aecad4849a30126

Cityscapes

2012.10557

multicam_individual_stream_acc_cityscapes.jpg

0.521518

466c36de7e164e6b80f6b9b50ca7945b

Waymo

2012.10557

multicam_individual_stream_acc_waymo.jpg

0.522924

c22ebc999e9f4444a406be9cfeca226c

Timeseries of our performance vs. fair sharing. Again, adaptive resource sharing constantly outperforms fair sharing. Accuracy drops over time.

2012.10557

multicam_taskwise_acc_zurich_4_cityscapes.jpg

0.445298

d90a353a53c04a2ab28d8326e09f7da5

's components and their interactions.

2012.10557

overview_cropped.jpg

0.42609

068398d1838f403e9b79de9b69bb794f

Throughput measured as number of video streams that can be concurrently supported for different resource sizes subject to accuracy thresholds.

2012.10557

scalability_GPUs_cams_cityscapes.jpg

0.463213

eb2b63c2c6444f5d8a21358943b1dcd1

Using golden model. Effect of adding more video streams on accuracy for different schedulers. When insufficient resources are provisioned, gracefully degrades to the No-retrain baseline, while the retraining decisions made by the fair scheduler reduce the mean inference accuracy.

2012.10557

scalability_cams_fixedGPUs_cityscapes_golden_model.jpg

0.487483

6aac994f176e4ee6af65f60eb6e2e9bc

Uniform scheduler

2012.10557

schedmot_multiwindow_eventual_best_cfgs.jpg

0.438354

069c1d10deaf46d0a7f5b1c2178e026f

Accuracy-optimal sched.

2012.10557

schedmot_multiwindow_prioritization_and_optimal_cfgs.jpg

0.401534

078ca6a1424f403c8e04e86962738ef7

Naive, eventual best cfgs

2012.10557

schedmot_res_eventual_best_cfgs.jpg

0.397626

75526972238843d69e26c71577c9a7d8

Optimal cfgs with prioritization

2012.10557

schedmot_res_prioritization_and_optimal_cfgs.jpg

0.38549

7ac5d18ed42d4a128185b34c96f5729d

We add a controlled amount of error to the accuracy prediction and our performance does degrade but only marginally. Pick 5 cities, horizontal graph

2012.10557

sensitivity-accuracy-error.jpg

0.488915

a99269e10b664211bba0dd9a3f686f76

Adding a controlled error ϵ to the accuracy prediction - 's performance degrades, but only marginally.

2012.10557

sensitivity_profileerrors_cityscapes.jpg

0.435815

9355db4bb11f449bb89f9e2d44093f86

Single-video performance: Ours (with continuous retraining) vs. various baseline DNN configurations (without retraining). Retraining allows ours to achieve much better accuracy-cost tradeoffs.

2012.10557

singlecam_acc_vs_cost_cityscapes.jpg

0.43237

7f57d2c77ed24b76a743a5fbd4cc2419

Cityscapes

2012.10557

singlecam_acc_vs_cost_golden_cityscapes.jpg

0.407254

f9591bccdce7459289770cabec841e9b

Waymo

2012.10557

singlecam_acc_vs_cost_golden_waymo.jpg

0.432279

0be4e037abb9407886e69d3fa086cded

Cityscapes

2012.10557

sysimpl_varyingcities_streams_cityscapes.jpg

0.487651

8f9a776a4bee4a5b95db73bff21f0c73

Waymo

2012.10557

sysimpl_varyingcities_streams_waymo.jpg

0.371886

efebc7e0b56d4e308f471a523c7ded2a

Waymo

2012.10557

sysimpl_varyingcities_waymo.jpg

0.513173

2a9c9b283af44da0953ca5f40ca77d7b

Waymo

2012.10557

waymo_across_resources.jpg

0.445798

3d1919937f2f4f07968a7a80e7772928

Waymo

2012.10557

waymo_scheduler_comparison_across_resources.jpg

0.406301

a155493c1ae14ac0875d0f7d5239450c

Cameras connect to the edge server (e.g., Azure Stack Edge <cit.>) through local networks (wireless, typically). The edge server is equipped with consumer-grade GPUs and executes DNN inference containers and retraining containers.

2012.10557

xxx_cropped.jpg

0.445986

abf1934439ca43fdbbaf4b22cc5a6654

Number of cliques of size ℓ∈{4, 5,6 } for the Indian village data set. The median values for ℓ = 4, 5, and 6 are 132, 13, 0, respectively.

2012.10559

Number_of_Cliques-eps-converted-to.jpg

0.464953

19ff601114bd46ed87d42b6bdb313a53

We generate a graph using a 3-dimensional Euclidean latent space with K = 10 cliques. We plot the scree function ϕ and the bootstrap variability function f_n defined in Algorithm <ref>. We also plot their sum, defined as the objective function. The horizontal axis represents the possible ranks of the matrix. We see the objective function has a minimum at 3, so we estimate the rank of the matrix to be 3, which is the true dimension of the latent space.

2012.10559

Rank_Estimate-eps-converted-to.jpg

0.445603

f53e9d9bedf049039d534118e19e3b83

Predicted dimensions for the predicted geometries of the Indian village networks. Most predictions are low-dimensional, usually between 1 and 3. Some of the predictions are higher. In these cases, our method is predicting that W_κ is full rank.

2012.10559

Rank_Estimates-eps-converted-to.jpg

0.440232

410d921d931a491b87f9c887afc6d7b2

Regression coefficients showing the relationship between loans and geometry. Each line in the plot corresponds to the coefficient in a multivariate linear regression where the outcome is the average amount of loans (in thousands of INR) and the predictors are geometry types (with spherical as the reference). The wide bars correspond to one standard error and the narrow bars represent two standard errors. The reference value for spherical is 16.71 (again in thousands of INR).

2012.10559

loans_coef.jpg

0.383797

db7c0a705a254e938a8097561a8177e7

Client-server architecture for remote visualisation.

2012.10560

O1-22_f1.jpg

0.467524

ef574b8644884586816d4402faeac3b3

IOI Model in Layer-1 of Anonymous Cryptocurrency

2012.10563

anonymous_cryptocurrency_ioi_model.jpg

0.47348

50a7317b39fa4ca3acad6cce0778d13d

Architecture Model for Anonymous Cryptocurrency

2012.10563

architecture_of_anonymous_cryptocurrency_02.jpg

0.487384

5870918c23d34c9786e78be3d3549ded

IOI Model in Layer-1+0 of Anonymous Cryptocurrency

2012.10563

cryptocurrency_ioi_model_ipaddress3.jpg

0.489403

717ca7c8b3b740d68f120a4fa6ecc639

Pfitzmann's Definition of Anonymity

2012.10563

pfitzmann_anonymity.jpg

0.4483

e2ae7477373c420ab270f377644e9ba3

Pfitzmann's Definition of Unlinkability

2012.10563

pfitzmann_unlinkability.jpg

0.416721

a6a5c37140c64e07a9177ec8049e3f63

SCCLA collusion

2012.10563

sccla_collusion.jpg

0.434043

48d764aebdf84849a20d0e58a0f85a82

Zerocash Coin Structure (This figure is cited from <cit.>)

2012.10563

zerocash_coin_structure.jpg

0.423599

1420206bf8564973a78e39451d21d03d

Wavelet scattering coefficients

2012.10564

chestxray-scattering-comparison.jpg

0.46405

e7ed0d7d9fbc41ac929ccd9666266455

Empirical MMD distributions under null (ℋ_0) and alternative (ℋ_1) hypothesis between the ChestXray14 and MIMIC-CXR data sources. We used scattering transform in composition with the radial basis function (RBF) as kernel function in order to find the B-test statistics. We set to compute up to the second-order of the scattering coefficients and set other filter parameters and . We summed the scattering coefficients over the image domain to obtain a translational invariant 417 features. We set the RBF scale parameter γ=1.

2012.10564

cxr-btest-sc-noadap.jpg

0.503843

ed31889e11774ec3bfd6fd103d362119

ChestXray14

2012.10564

cxr_chestxray_sc_hist2d-cyclegan.jpg

0.467388

491fcc5ae72443679bd94c7a5a39b05a

ChestXray14

2012.10564

cxr_chestxray_sc_hist2d-noadap.jpg

0.417871

03d8a63f5f96449a8babc00700f3ad0e

Six training loss of , trained between the samples of and , are displayed for 14 epochs. We denote two discriminator losses by and , two generator losses by and , and two cycle-consistency losses by and .

2012.10564

cxr_cyclegan_training_loss.jpg

0.421638

26bf11dc2a504eab8ff63f8605e7bc69

MIMIC-CXR

2012.10564

cxr_mimic_sc_hist2d-cyclegan.jpg

0.449009

b6fc41a1347a4d58abab0d6865e6d0f3

MIMIC-CXR

2012.10564

cxr_mimic_sc_hist2d-noadap.jpg

0.435761

5cf1f14325cb44ff8ce69466ee422030

Differences in the abnormality prediction scores obtained by on and style-adjusted by is binned into non-overlapping intervals, and the counts in every interval are displayed.

2012.10564

cxr_preds_compare_hist.jpg

0.499476

1912f8606bcf4e01b7f790bc9ab5e13e

CycleGAN contains mappings between two domains G : X → Y and F : Y → X, and one discriminator for each domain, D_X and D_Y. The purpose of including discriminators is to encourage the generators G and F to generate samples that can not be indistinguished with the available real samples from the two domains. Additionally, CycleGAN introduced cycle consistency losses to enforce forward and backward cyclic consistency between the generators, i.e., x → G(x) → F(G(x)) ≈ x, and y → F(y) → G(F(y)) ≈ y. We have included this figure into this manuscript from the CycleGAN paper <cit.>.

2012.10564

cyclegan.jpg

0.478027

ebd07631a67c4fc3b7e42b08b125bb13

Clinical workflow for computer-aided abnormality detection in chest radiographs

2012.10564

domain-adaptation-workflow.jpg

0.405248

8a81b57ad6de4ea685414b556b54c920

Overview of our approach. Our system takes an input image, a mask of the object to remove, and renderings of an approximate scene model with and without the object. We first perform an intrinsic decomposition of the input image into texture and lighting. We then remove the object's shadows from the lighting image. We inpaint the object mask region separately in both texture and lighting before recompositing to get our final result.

2012.10565

architecture_horizontal_3.jpg

0.532693

28201e83b63645c8af0b6272e6517698

Comparison of the dynamic temperature profiler sensor offsets across the four calibration campaigns. The offsets are consistent and the small differences are attributed to local variations during movement of the profiler. The one outlier (October 2018 calibration of the 6th sensor from bottom) eventually converged to give consistent readings in subsequent calibration campaigns.

2012.10567

plot2A.jpg

0.383955

b63e58684bb24dda9cc24508935b72be

The dynamic temperature profiler sensor values compared to the CFD simulation prediction, as a function of height from the detector bottom.

2012.10567

plot3A.jpg

0.444508

d53e7432bd7d44b48cbba71b0be072fc

Dynamic temperature profiler sensor values as a function of time. The y-axis values represent the differences in temperature of each sensor with respect to the bottom-most reference sensor.

2012.10567

plot4A.jpg

0.481229

144c30d1461e4a91b3900b613d9d5fba

The dynamic temperature profiler sensor values as a function of height from the bottom of the detector. Most of the values are within a 3 mK band for sensors immersed in liquid-argon. Beyond 7 m the sensors are in the gas-liquid interface.

2012.10567

post_1A.jpg

0.429246

582936d69a804844a14be212c83e7f68

Example of a science observation reduction page in the JSA processing tracking system. A JCMT calibration observation of N207IR taken by the instrument HARP in 2016 is shown.

2012.10568

Data-QA-pages-HARP.jpg

0.471871

31e8f3a6b34d4653b4c4d85a1f78f1f6

Observatory hours broken down by category. Science, Calibration, Engineering and Commissioning(E&C), Fault, Weather (time when observatory was closed due to weather), and Shutdown (time when observatory was closed). The months displayed are 1) four months (March - June 2019) prior to shutdown due to the restricted access on Maunakea when the JCMT was operated from the summit. 2) the five months after the return to operations (June - October 2020) from global pandemic shutdown when the JCMT was operated remotely. The time gap and switch to Remote Operations are indicated by the dashed vertical line.

2012.10568

Hours-EO-night-since-March2019.jpg

0.479009

17054fab6bf0478bbd819c5973fe6499

First night of Remote Operations on November 1st 2019 from the JCMT Remote Control room (JROC) by JCMT TSS Mimi Fuchs.

2012.10568

first-night-remote-ops.jpg

0.526928

1111bdbb89144f95839fe64bc5433745

Injection of flow from a slot into a cross flow

2012.10572

F01.jpg

0.477801

560cdd3dfc4a475d9b161a22b5af8d51

Two-phase fluid

2012.10572

F010.jpg

0.498633

ffd070c9912f4db39eece85b0af39dcc

The critical case Q=0

2012.10572

F011.jpg

0.521604

1a23a63008af41869dd41b8130dd10cc

Case 1

2012.10572

F012.jpg

0.492657

2898fb12761840fc8d0e3dd2d36f1726

Case 2

2012.10572

F013.jpg

0.470916

b02bc994ad214ecbb2786c27782d4fea

Horizontal blade surfaces

2012.10572

F02.jpg

0.478203

c8b98c37850e4afdaf557f913b275b3a

Figure 9.3 in <cit.>

2012.10572

F03.jpg

0.488139

737650f68e6c4cd38807525e288e9dfc

Oblique injection of flow into a free stream

2012.10572

F04.jpg

0.516258

d8187058ce8f43869dfc732c14df7ef9

The nozzle walls and the ground

2012.10572

F05.jpg