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
0.475839	42d933dac55b425ebd821dfad2052302		2012.11149	marmousi_model_true.jpg
0.452352	0cfd4d5d8c1547e6b7573ec67041908d		2012.11149	profile_BP.jpg
0.42259	05a054b361b54712977bef83a5a782ef		2012.11149	profile_marmousi.jpg
0.435321	018f2762328b4d828a98b9f4fc3a90d8		2012.11149	profile_marmousi_noise05.jpg
0.444678	cfcef5f55f2f4d0c97e2beaaf5c4494a	The architectures of the token-level model (a) and sentence-level model (b).	2012.11157	arch.jpg
0.421576	a3a81c6604f24e369b52e150498733d3	Illustration of our Narrative Incoherence Detection tasks. Note that a narrative contains more sentences in real cases.	2012.11157	example.jpg
0.418672	99aac855ba02459cae372a998e27d846	FCR Activation Ratios	2012.11158	FCR.jpg
0.442527	c6176bf9499f41c7986ac8eefd854d98	Used Prices for Model Testing	2012.11158	Inputs.jpg
0.378049	add35754fef946e4857dcd80caed6e16	Deterministic Model – Day 1	2012.11158	Out_avgA_day-1.jpg
0.449298	f20bee7247a24917b230e2b85df9a08e	Robust Model – Day 1	2012.11158	Out_robM_day-1.jpg
0.373244	f1c10ac6e178483fb214fc553a95f84d	Stochastic Model – Day 1	2012.11158	Out_scen_day-1.jpg
0.448659	45484dc3cbd145ec93e02b52635447ea	aFRR Activation Ratios	2012.11158	aFRR.jpg
0.409521	ec71039947dd41d089c34224b6437f6a	Cases of inconsistency errors. Foreign key constraints are annotated by arrows; transformations are highlighted in red; inconsistencies are highlighted in green; denotes the prediction on the transformed inputs and denotes the prediction on the original inputs.	2012.11163	case_new.jpg
0.492584	3074b6e17f754f69ba732c843eb163b2	Cumulative fluency scores distributions.	2012.11163	cdf.jpg
0.457883	d64725d1c34947748282068a8f7ca8dd	Inconsistent predictions on semantically equivalent inputs.	2012.11163	motivating_example.jpg
0.480691	67c45be25cd9400b8bdc9a36bfc94434	Illustrative examples of our designed MRs on schemas.	2012.11163	schema_mutation.jpg
0.420821	4f5d22d8e9bc4c96bff52a5ca33823a2	The FAST photo taken with a wide-angle lens (a) and its dish structure(b)	2012.11167	fig1.jpg
0.544673	603cf87fad6f4f6ca6495a89780daec3	Schematic diagram of FAST observation principle	2012.11167	fig2.jpg
0.424344	60db2f9a996b4075972f4b6a33e6e6a6	Diagram of the paraboloid and the spherical surface in polar coordinates. The solid line represents the spherical surface, the dashed line represents the paraboloid shaped by distorting the spherical surface; the vertex of the paraboloid is also shown. The aperture of the paraboloid is denoted by D(m), the radius of the spherical surface is R(m) , and the opening aperture of the spherical surface is given by D'(m).ρ_p is the radial length from the pole to the paraboloid in meters,which is a function of polar angle(θ).	2012.11167	fig3.jpg
0.427203	c8a17144bd7c4c1194d56195593de722	Radial deviations of the paraboloids from the spherical surface. The abscissa represents the polar angle range corresponding to a paraboloid of 315 m aperture in radians. The ordinate shows the radial length from the pole to the spherical surface in meters, and it also shows the radial length from the pole to the paraboloids under the two calculation conditions, in meters. The dashed line indicates the radial length from the pole to the spherical surface. The dash-dotted line indicates the radial length from the pole to the paraboloid under condition (1) and the solid line indicates the radial length from the pole to the paraboloid under condition (2). As mentioned in section 2,condition(1) requires the radial distance of the paraboloid from the spherical surface at the parabolic vertex to have a minimum value; condition (2) requires that the maximum radial distance of the paraboloid from the spherical surface is a minimum.	2012.11167	fig4.jpg
0.377529	f1b5be36846d4fef9815bf18b6b80912	The maximum zenith angle corresponding to the apertures of the paraboloid. The abscissa represents the aperture of the paraboloid in meters. The ordinate shows the maximum zenith angle corresponding to the apertures of the paraboloid in degrees. The solid line is for the spherical surface of FAST, whose radius is 300 m and opening aperture is 500 m. The dotted line is for the spherical surface with a radius of 400 m and an opening aperture of 670 m. A star indicates the maximum zenith angle of the 300 m aperture of FAST, which is about 26.4 deg.	2012.11167	fig5.jpg
0.388419	d3c30f292d8b4fbcbc9af0442d9d935d	The calculated maximum radial deviation of the paraboloid from the spherical surface. The abscissa represents the aperture of the paraboloid in meters. The ordinate shows the calculated maximum radial deviation of the paraboloid from the spherical surface in meters. The solid line indicates the calculated maximum radial deviation when the radius of the main spherical surface is 300 m. The dotted line is for the calculated maximum radial deviation when the radius of the main spherical surface is 400 m. A star indicates that the maximum radial deviation of the 300 m aperture of FAST is about 0.47 m.	2012.11167	fig6.jpg
0.461796	a9ca987433714fd995af1f972df17bfb	Optimality gap between the proposed approach and the ideal case for different antenna parameters.	2012.11168	fig_GT_opt.jpg
0.453782	59e9e7b851024ddab86515b1c2c2f8d0	The effect of feedback overhead on the performance of the proposed approach.	2012.11168	fig_effect_overhead.jpg
0.449435	361580e1549e45ea8e634571579764ed	Solution flow of the proposed distributed beam scheduling framework. The original network utility maximization problem which involves maximizing a function of the time averaged UE throughput is transformed into a new problem whose objective is to maximize the time averaged value of the utility function. The transformed problem is then decomposed into two sub-problems by introducing the auxiliary variables γ_j,i(k) in each epoch. The first sub-problem is convex and thus can be easily solved. The second sub-problem is non-convex and aims to determine the BS transmit powers p_j,i(k,n) in each block. We formulate the distributed beam scheduling problem as a non-cooperative game where each BS attempts selfishly to maximize its own payoff which is a function of the Lyapunov virtual queue parameters Z_i(k),H_j,i(k) and the transmit powers of other BSs. We then propose a parallel update algorithm to find the Nash Equilibrium of the game which serves as a distributed solution to the second sub-problem.	2012.11168	fig_solution_flow.jpg
0.434334	f6646b8a2d754076a163d5467d673e94	Distribution of the BSs and the UEs over a planar grid of dimension 800× 800 meter squared. Each BS is associated with ten UEs and these UEs are uniformly distributed in the disk coverage area with a radius of 150 meters. This BS-UE association is fixed over all epochs.	2012.11168	fig_sys_new.jpg
0.444441	719bbcfdfd4d421b84177ce220332bde	Domain(language) loss, domain UAR, and emotion development set UAR from the âRec2Iem arousal" training	2012.11174	loss.jpg
0.427854	95db83ea9f8a4319a86c444b0b4e12f8	The proposed model structure. BN and GRL are Batch Normalization, and Gradient Reversal Layer respectively.	2012.11174	model_final.jpg
0.407972	75b1acc4f3064aaf9998f01171513a27	The graph decomposition sample. The left sub-figure is the adjacency matrix of the graph in the right sub-figure, where the green and blue represent the decomposed two subgraphs.	2012.11175	3_1.jpg
0.4366	b1f3deda835d493e956dc0545fbb9725	MPG outperforms other methods in the antibacterial properties prediction open task hosted by MIT J-Clinic.	2012.11175	AI_cure_top10_new.jpg
0.425925	4afad659f36440e6a603b824d79882ce	MPG provides explainability for DDI prediction. The left part of the figure illustrates nitrate-based drugs with each atom colored by the attention weight. Note that our MPG always pays more attention to the nitrate group.	2012.11175	DDI.jpg
0.434085	8ebf15e1b16e4b53b10589dfbfb47d73	Classification performance for DTI prediction on Human (a) and C.elegans (b) datasets.	2012.11175	cpi.jpg
0.451607	f0b29170faf540aeb29409287a94e874	Two example molecules with attention weights coincided with HOMO and LUMO. The attention weights represent the importance of atoms to a molecule's global characteristics, extracted from the last layer of and normalized. The larger shading size and deeper color both denote larger attention weight. HOMO/LUMO orbitals are calculated by Materials Studio DMol3.	2012.11175	homo.jpg
0.51385	8a2b2ecfd20441f1b334fedfc7a12a12	The input representation of graph data is constructed by summing two parts: feature embedding and segment embedding. (a) Feature embedding: a set of node and edge features go through the embedding transformation to describe a graph. (b) Segment embedding: a learned segmentation embedding to every node and every edge indicating which subgraph it belongs to, different colors represent different segmentation.	2012.11175	input.jpg
0.482419	d590ca54dae54b0686cf5773c0e94080	The overview of MPG framework. The bottom left sub-figure (b) illustrates the workflow of MPG framework. The MPG framework includes two key components— and self-supervised strategies. The architecture of is shown in the bottom right sub-figure (c). The top sub-figure (a) illustrates the schemes of our self-supervised strategy, including PSD we proposed and AttrMasking for pre-training GNN model.	2012.11175	overview.jpg
0.482667	b9c774ee956e4800a127eb32c3c6710b	5-fold cross-validation classification performance on TWOSIDES dataset.	2012.11175	twosides.jpg
0.412626	55aa9219722344858020c66584303145	Visualization of the molecular representation by UMAP. The molecular representation is the collection node's embedding extracted from the last layer of non-pre-trained or pre-trained . In (a) and (b), pre-trained is capable of distinguishing valid and invalid molecules. In (c) and (d), The different colors represent different scaffolds the molecules belong to. DB index is Davies Bouldin index <cit.>, a lower DB index means that the clustering has a more appropriate separation.	2012.11175	vis_mol.jpg
0.443583	c9875cc7a5bb43999a6598b0f7526067	Situation of the proof of Claim <ref>. The dark gray disks are the points at distance at most '+(1+_n)τ from _n-1(i_n-1/p) and . The light gray hippodrome consists of the points at distance at most '+2(1+_n)τ from the segment _n-1(I_i).	2012.11181	ipeFigs.jpg
0.469978	3ef8f8a783864a14b8e3e53a3ba7a3c4	Monte Carlo histories of x^2 and x^4 for a random walk in the interval [-1,1]. Clearly the two histories are correlated.	2012.11183	hist.jpg
0.400229	5fb0251f280a43d5932797d483653bf2	The proposed NE-SGD framework for optimizing a basic ResNet18 network. The rounded rectangle shows the decoding implementation of the ResNet18 network. The solid colored region maps k_i^j=1 in the binary code. The gray area maps k_i^j=0 in the binary code. Each ResNet18 network is represented an individual in the evolutionary algorithm. The test accuracy of the model is used as the fitness value of an individual. The parental individual is selected using the tournament method for further crossover and mutation operations. The above process is iterated until the stop criteria are met.	2012.11184	Figure01new2.jpg
0.50639	4f8ff124790843929ff1d97933bb00e9	DarknetConv2D structure	2012.11185	1.jpg
0.429969	20d887b8dd9c4005bd70e77e2f761aed	Diou-Yolov3 detection results	2012.11185	10.jpg
0.470776	1778fcdac03f4c68985a38be0690e5a5	Yolov3 Loss curve	2012.11185	11.jpg
0.478891	368d5ee98e5b4ef9abf3cafd466fe81a	Diou-Yolov3 Loss curve	2012.11185	12.jpg
0.461936	62f145b0f7164112a270beb1ec64d4d2	Comparison of loss curves between two models	2012.11185	13.jpg
0.473904	fd1f38d4d6ba44e0954bb9412ae870a4	Diou-Yolov3 Accuracy curve of the model	2012.11185	14.jpg
0.548521	c059d4f652de4168962a55867d55446a	Yolov3 Accuracy curve of the model	2012.11185	15.jpg
0.488911	e822d1660d8646d9a246f25fade5e8d1	Residual block structure	2012.11185	2.jpg
0.401915	03c9aa1ba3aa4c31a7bfef3a9b2782da	I/O block diagram of yolov3	2012.11185	3.jpg
0.43637	29267922c8f146bd981676e02e5bf71e	iou loss	2012.11185	4.jpg
0.446372	2f278ceb1e2b4bf497acb95aaa2fcbeb	Diou loss	2012.11185	5.jpg
0.43137	0ebb3ce7d3634974b714f33005991194	Part of the training set	2012.11185	6.jpg
0.397723	11d467735d4241269efad68d80f064cb	Partial test set sample	2012.11185	7.jpg
0.488688	6d038929970341349ac4447fecfe5a90	Experimental flow diagram	2012.11185	8.jpg
0.394466	c8fc893531c443a88f8fbf2c1b8606c2	Yolov3 detection results	2012.11185	9.jpg
0.512359	3aa44148440c44a8b1eadad498bd219c	Core order for the graph from <ref>. Each entry of is displayed horizontally. We omit [1] since we assume μ≥ 2. The number beside each vertex in a row [μ] is the core threshold for that vertex for that μ. For example, in [2], the number .75 beside vertex 6 means that when μ = 2 and ≤ .75, vertex 6 is a core vertex. Like in <ref>, we consider the specific case where (μ, ) = (3, .6); we highlight [μ] in gray and color the core vertices (i.e., vertices with core threshold at least ) blue.	2012.11188	core-order-example.jpg
0.407288	cb333d985ace49159d9da2e7e78eee7d	Neighbor order for the graph from <ref>. In this figure, for each v ∈ V, we display [v] as a column. The numbers beside each vertex are similarity scores. For example, in [3], the .87 label beside vertex 2 represents the cosine similarity of .87 between vertices 3 and 2. Like in <ref>, we consider the specific case where =0.6 and color all -similar neighbors green and all other neighbors red.	2012.11188	neighbor-order-example.jpg
0.434094	1ce2162de7e646028c7a87d56c78352e	Example SCAN clustering with μ = 3 and =.6. The labels on the edges are cosine similarities. Core vertices are blue, whereas non-core vertices are white. Edges with similarity greater than are green, whereas other edges are red. There are two clusters (vertices 1, 2, 3, 4 and vertices 6,7,8,11) as well as three unclustered vertices (hub vertex 5 and outlier vertices 9 and 10).	2012.11188	scan-clustering-example.jpg
0.53004	7e86a2adf3804b49b4359fde138b657b	The log of the RMSE of entropy estimations versus the log of the sample size for NJEE and benchmark methods in different simulated studies. The results are the average of 100 measurements per each sample size and distribution type.	2012.11197	univariate_entropy.jpg
0.494418	dfada66572e94342b62e1a0057074970	MAE in equation (<ref>) versus the inverse temperature β when (a) p = 0.2, (b) p = 0.4, and (c) p = 0.8. These plots present the averages over 1000 experiments.	2012.11198	AIS_vs_SMCI.jpg
0.440487	a6d2f21c663247a6a8b37fcc953447c8	Results on the Hopfield-type Ising models; MAE in equation (<ref>) versus β when (a) α = 0.2 and (b) α = 0.5. These plots present the averages over 2000 experiments.	2012.11198	HFM.jpg
0.401155	b4c024ab801b4418aed7f51ae23bc694	MAE in equation (<ref>) versus K when (a) p = 0.2 and β = 0.5, (b) p = 0.2 and β = 2, (c) p = 0.8 and β = 0.5, and (d) p = 0.8 and β = 2. These plots present the averages over 1000 experiments.	2012.11198	K_dependency.jpg
0.44811	1b4028dfc6154ffba9e4f87a0b8d2306	MAE in equation (<ref>) versus N when (a) p = 0.2 and β = 0.5, (b) p = 0.2 and β = 2, (c) p = 0.8 and β = 0.5, and (d) p = 0.8 and β = 2. These plots present the averages over 1000 experiments.	2012.11198	N_dependency.jpg
0.575675	690a2b05fac14bc8bf27c254bea54ff6	MAE in equation (<ref>) versus β when (a) p = 0.2, (b) p = 0.4, and (c) p = 0.8. The results of SMCI and AIS are identical to those in figure <ref>. These plots present the averages over 1000 experiments.	2012.11198	PM_AIS_vs_SMCI.jpg
0.450878	0fabace20b5a44fbb289488011d5c0c3	Results on the bipartite Ising models; MAE in equation (<ref>) versus β when (a) p = 0.5 and (b) p = 1. These plots present the averages over 3000 experiments.	2012.11198	RBM.jpg
0.510972	bbed062e4740436cb8fc51107b9de51c	Qualitative illustration of the expected performance of the proposed method.	2012.11198	accuracy_image.jpg
0.50837	7171f9873af44cc0bb44ec063e03131a	Illustration of the target, sum, and sample regions of SMCI.	2012.11198	region_sheme.jpg
0.459331	27c9f7b11f0e4fafbf95e472fec71ba0	Comparison with PT-based method; MAE in equation (<ref>) versus β on (a) the random graph with p = 0.5 and (b) the Hopfield-type Ising models with α = 0.2. The results of SMCI and the proposed method in panel (b) are identical to those in figure <ref>(a). These plots present the averages over 2000 experiments.	2012.11198	vs_PT.jpg
0.430086	36aecf662d034ad492ec4e52c26f9408	Targeted transferability of three simple transferable attacks (CE-TMDI, Po+Trip-TMDI and Logit-TMDI) with varied step size, S.	2012.11207	curve_stepsize.jpg
0.401809	b02a7dbdbed34b8698c9478f5803c0e3	Targeted transfer success rates (%) of the logit loss compared to the C&W loss varied settings of the confidence controlling parameter K <cit.> in the single-model transfer scenario.	2012.11207	cw_vs_logit.jpg
0.437921	9455b8f4c30f45c2b5a38fc2e4b2108a	Targeted transfer success rates (%) in our challenging ensemble transfer scenario, where each hold-out target model shares no similar architecture with the source models used for ensemble.	2012.11207	ens.jpg
0.412418	a35a176a37a0413894f4ef66e7c7e6e8	Adversarial images for attacking Google Cloud Vision.	2012.11207	goo.jpg
0.413926	b1b81a13fc214e95b647693d251c3b21	Targeted universal adversarial perturbations (ϵ=16) for different classes.	2012.11207	tUAPs.jpg
0.423966	131592d738e7447480327c13f63c94fe	Unbounded adversarial images targeting different classes.	2012.11207	unbounded.jpg
0.439597	e301df78d3d54c368420e51ab6b51139	Schematic interactions within the proposed HPS framework.	2012.11208	fig.jpg
0.450433	77728515c63b47b799ffbdf492ce3c59	The diffusion constant in the J_2 channel (solid line), being the sum of the Drude approximation D_2^D plus the weak-localization correction Δ D_2 (k) in Eq. (<ref>) from two dipoles, as a function of the detuning δ = (ω-ω_0)/γ. It is normalized by the diffusion constant in the J_0 channel. The 4 weak localization corrections discussed in this section are shown separately as dashed lines.	2012.11210	DALL2norm.jpg
0.514366	2c65805a33394b869ea52bd629aff4e7	The contribution of dipole-dipole coupling to the DOS (in units of N_0 × Q_0 × (4π n/k_0^3)^2) as a function of detuning δ = (ω-ω_0)/γ. The dashed lines show the separate contributions of the modes with electric field perpendicular (P) and parallel (Q) to the separation vector 𝐫.	2012.11210	DOSPQ5.jpg
0.542205	e058fb71389d48fe888871b45a6158b4	Diagrammatic presentation of the ladder series (without external lines) involving two different electric dipoles. Dashed lines connect identical dipoles, solid lines denote the Green's tensor 𝐆_0(𝐫), crosses denote transition matrix t(k), bottom line denotes Hermitian conjugation. The first diagram on the left is reducible (a simple product) and is not part of the collision operator U_𝐩𝐩'.	2012.11210	IRRLADD.jpg
0.421103	bd0807dd30484c88bd0471ae9f56b58d	The self-consistent solution for the electromagnetic transport mean free path ℓ^* defined by σ = k_e^2 ℓ^/6π. Shown are the values for k_eℓ^ for the full solution in this section, the conventional picture described by Eq. (<ref>) with only the transverse mode J_0 considered, with a mobility edge predicted around k_eℓ≈ 1, the lower threshold imposed by the existence of the diffusion modes J_2 and J_3, as well as k_eℓ^* associated with the fictitious conductivity and J_3. We used a cut-off q_m = 1/ℓ.	2012.11210	MEabs.jpg
0.607628	3a0842b1dc214979b4b07e04d2dc29b7	The diagrammatic convention associated with the reducible vertex R_ij\|kl, 𝐩𝐩'(𝐪), with external lines. Top line denotes retarded Green's function 𝐆(k + i0), bottom line 𝐆^†(k + i0) =𝐆(k - i0) is the advanced Green's function and travel in the opposite direction. The polarization labels are ij on the left hand side (“observer”) and kl on the right hand side (“source”).The sum of incoming and outgoing wave numbers is conserved.	2012.11210	Rdia.jpg
0.48707	03603904e71b48b8a88b2daddaf5713b	Average, on-resonance, electric field in the most dense sample. Red and blue lines show the real and imaginary parts of the field, respectively. The black solid line is the fit of Eq. (<ref>) to the numerical data; the dashed line is the result for the imaginary part obtained from the fit to the real part.	2012.11210	fig_average_field.jpg
0.604535	6ddfd543e03d4b5993860b2b0363fb84	The ratio of transport and scattering mean free paths ℓ^/ℓ as a function of k_e ℓ compared to the self-consistent theory for 4π n/k_0^3 = 3.77 and q = 0.5 (black solid line, for the two branches, see text for explanation). Points of different colors correspond to different scatterer number densities n for detunings δ = (ω-ω_0)/γ∈ [-3, 6.5] from the resonance. The dashed line is the lower bound for ℓ^/ℓ described by Eqs. (<ref>) and (<ref>), again with q=0.5.	2012.11210	fig_comparison.jpg
0.403274	d9aa48aa47b644c88d986bffa2954c88	The average energy density, consisting of a coherent and a diffuse part. The solid straight line is a linear fit to the diffuse energy density for k_0 z = 4–9.	2012.11210	fig_energy_density.jpg
0.472575	b8247c0833804de786fccf0d2244a6ca	The effective wave number obtained from the fit to the coherent field (see Fig. <ref>), the real part of K_L(∞) calculated using Eq. (<ref>), and the ISA result.	2012.11210	fig_keff.jpg
0.445968	43d56a4285034cf5b3155687a6e8bf24	Scattering and transport mean free paths for the most dense sample. We also show the scattering length 1/2 Im K_L(∞) associated with longitudinal waves.	2012.11210	fig_tr.jpg
0.449839	f0944746e16242d3bcecaac73e6f644e	Transport mean free paths for different densities n. The dashed line shows the ISA results for the lowest density 4 π n/k_0^3 = 0.25.	2012.11210	fig_tr_all.jpg
0.468834	0730c5277813413d94d52d1cac042df4	Observing the brain tumor from different views, the FLAIR, T1, T1c, T2 indicate four modalities of MRI images. (a): The image in Axial View, (b): The image in Coronal View, (c): The image in Sagittal View.	2012.11211	Multi-view_brain_tumor.jpg
0.406308	f6529be8a2da4467ad7a50d72c06d07e	The FCN network structure and UNET network structure for the multi-view network. The network operation layer includes the initialization layer, resblock layer, conv layer, upconv layer, resize layer, logit out layer and softmax, where BN represents the batch normalization layer, Relu represents the activation function, Conv represents the downsampling operation, DeConv represents the upsampling operation, and Bilinear represents the bilinear interpolation.	2012.11211	Network_Architecture.jpg
0.474213	502bc7d948244f38a26390955be0482d	The overall architecture of the Multi-View Dynamic Fusion Framework for segmenting the 3D brain tumor from the axial view, coronal view and sagittal view.	2012.11211	Overall_Architecture.jpg
0.447402	116ecdaf0f1a4ac19643ceab7cd06d6d	Visually comparing the segmentation results of the two decision fusion methods based on the FCN segmentation network in the validation dataset. From left to right are: Input: Input images connected by four modalities data; GT: the Ground-Truth; Segmentation: Original outputs of the network; Voting: The multi-view fusion result based on the voting method of referring the axial view; WA: The multi-view fusion results based on the weighted averaging method of fusing three views.	2012.11211	Segmentation.jpg
0.481027	5d9140402d6a40dca9d75ec5342e74b1	An example of adopting the voting method with reference to the view 1 to perform the data fusion.	2012.11211	Voting.jpg
0.515806	dfc55c7ec22544ee99e81f1b3c70cdca	An example of adopting the weighted averaging method to perform the data fusion.	2012.11211	Weighted_averaging.jpg
0.580969	dc5b613ac79844dcbc18571a3e93c869	Examples of three types of backdoor attack triggers	2012.11212	3-attack.jpg
0.467364	76e8e21b7214445eb4b92004b3a4c4c7	CycleGAN structure	2012.11212	CycleGAN.jpg
0.4156	122e452111194d8f9be36527846c66e5	Style transfer by CycleGAN	2012.11212	CycleGAN_overview.jpg
0.558917	e78dbf62c4294f2493d1a8b4cab6fbe2	Examples to illustrate ABS. From left to right, the original image, the image with our trigger applied, and the image with the trigger by ABS applied. Note that here ABS is applied before detoxification. The trigger by ABS hence induces misclassification.	2012.11212	abs-result.jpg
0.459877	5cc5df88dfc14930ba577dbd204b87de	Model internals after individual rounds of detoxification for VGG on GTSRB.	2012.11212	detoxification_V_G.jpg
0.442259	812f42e777aa405fb6539957b723fb0b	Overview of detoxification	2012.11212	dex.jpg
0.449773	71c32ba0d2174ca6a9a724fe272fc2c1	Deep Feature Space Trojaning	2012.11212	featureattack.jpg
0.446447	852c4496593f4ee2bf3128001d91e514	Trigger pattern examples by Neural Cleanse. The first is the target label trigger pattern and the other two are for other labels. Note that there is not much size difference.	2012.11212	nc-result.jpg

0.475839

42d933dac55b425ebd821dfad2052302

2012.11149

marmousi_model_true.jpg

0.452352

0cfd4d5d8c1547e6b7573ec67041908d

2012.11149

profile_BP.jpg

0.42259

05a054b361b54712977bef83a5a782ef

2012.11149

profile_marmousi.jpg

0.435321

018f2762328b4d828a98b9f4fc3a90d8

2012.11149

profile_marmousi_noise05.jpg

0.444678

cfcef5f55f2f4d0c97e2beaaf5c4494a

The architectures of the token-level model (a) and sentence-level model (b).

2012.11157

arch.jpg

0.421576

a3a81c6604f24e369b52e150498733d3

Illustration of our Narrative Incoherence Detection tasks. Note that a narrative contains more sentences in real cases.

2012.11157

example.jpg

0.418672

99aac855ba02459cae372a998e27d846

FCR Activation Ratios

2012.11158

FCR.jpg

0.442527

c6176bf9499f41c7986ac8eefd854d98

Used Prices for Model Testing

2012.11158

Inputs.jpg

0.378049

add35754fef946e4857dcd80caed6e16

Deterministic Model – Day 1

2012.11158

Out_avgA_day-1.jpg

0.449298

f20bee7247a24917b230e2b85df9a08e

Robust Model – Day 1

2012.11158

Out_robM_day-1.jpg

0.373244

f1c10ac6e178483fb214fc553a95f84d

Stochastic Model – Day 1

2012.11158

Out_scen_day-1.jpg

0.448659

45484dc3cbd145ec93e02b52635447ea

aFRR Activation Ratios

2012.11158

aFRR.jpg

0.409521

ec71039947dd41d089c34224b6437f6a

Cases of inconsistency errors. Foreign key constraints are annotated by arrows; transformations are highlighted in red; inconsistencies are highlighted in green; denotes the prediction on the transformed inputs and denotes the prediction on the original inputs.

2012.11163

case_new.jpg

0.492584

3074b6e17f754f69ba732c843eb163b2

Cumulative fluency scores distributions.

2012.11163

cdf.jpg

0.457883

d64725d1c34947748282068a8f7ca8dd

Inconsistent predictions on semantically equivalent inputs.

2012.11163

motivating_example.jpg

0.480691

67c45be25cd9400b8bdc9a36bfc94434

Illustrative examples of our designed MRs on schemas.

2012.11163

schema_mutation.jpg

0.420821

4f5d22d8e9bc4c96bff52a5ca33823a2

The FAST photo taken with a wide-angle lens (a) and its dish structure(b)

2012.11167

fig1.jpg

0.544673

603cf87fad6f4f6ca6495a89780daec3

Schematic diagram of FAST observation principle

2012.11167

fig2.jpg

0.424344

60db2f9a996b4075972f4b6a33e6e6a6

Diagram of the paraboloid and the spherical surface in polar coordinates. The solid line represents the spherical surface, the dashed line represents the paraboloid shaped by distorting the spherical surface; the vertex of the paraboloid is also shown. The aperture of the paraboloid is denoted by D(m), the radius of the spherical surface is R(m) , and the opening aperture of the spherical surface is given by D'(m).ρ_p is the radial length from the pole to the paraboloid in meters,which is a function of polar angle(θ).

2012.11167

fig3.jpg

0.427203

c8a17144bd7c4c1194d56195593de722

Radial deviations of the paraboloids from the spherical surface. The abscissa represents the polar angle range corresponding to a paraboloid of 315 m aperture in radians. The ordinate shows the radial length from the pole to the spherical surface in meters, and it also shows the radial length from the pole to the paraboloids under the two calculation conditions, in meters. The dashed line indicates the radial length from the pole to the spherical surface. The dash-dotted line indicates the radial length from the pole to the paraboloid under condition (1) and the solid line indicates the radial length from the pole to the paraboloid under condition (2). As mentioned in section 2,condition(1) requires the radial distance of the paraboloid from the spherical surface at the parabolic vertex to have a minimum value; condition (2) requires that the maximum radial distance of the paraboloid from the spherical surface is a minimum.

2012.11167

fig4.jpg

0.377529

f1b5be36846d4fef9815bf18b6b80912

The maximum zenith angle corresponding to the apertures of the paraboloid. The abscissa represents the aperture of the paraboloid in meters. The ordinate shows the maximum zenith angle corresponding to the apertures of the paraboloid in degrees. The solid line is for the spherical surface of FAST, whose radius is 300 m and opening aperture is 500 m. The dotted line is for the spherical surface with a radius of 400 m and an opening aperture of 670 m. A star indicates the maximum zenith angle of the 300 m aperture of FAST, which is about 26.4 deg.

2012.11167

fig5.jpg

0.388419

d3c30f292d8b4fbcbc9af0442d9d935d

The calculated maximum radial deviation of the paraboloid from the spherical surface. The abscissa represents the aperture of the paraboloid in meters. The ordinate shows the calculated maximum radial deviation of the paraboloid from the spherical surface in meters. The solid line indicates the calculated maximum radial deviation when the radius of the main spherical surface is 300 m. The dotted line is for the calculated maximum radial deviation when the radius of the main spherical surface is 400 m. A star indicates that the maximum radial deviation of the 300 m aperture of FAST is about 0.47 m.

2012.11167

fig6.jpg

0.461796

a9ca987433714fd995af1f972df17bfb

Optimality gap between the proposed approach and the ideal case for different antenna parameters.

2012.11168

fig_GT_opt.jpg

0.453782

59e9e7b851024ddab86515b1c2c2f8d0

The effect of feedback overhead on the performance of the proposed approach.

2012.11168

fig_effect_overhead.jpg

0.449435

361580e1549e45ea8e634571579764ed

Solution flow of the proposed distributed beam scheduling framework. The original network utility maximization problem which involves maximizing a function of the time averaged UE throughput is transformed into a new problem whose objective is to maximize the time averaged value of the utility function. The transformed problem is then decomposed into two sub-problems by introducing the auxiliary variables γ_j,i(k) in each epoch. The first sub-problem is convex and thus can be easily solved. The second sub-problem is non-convex and aims to determine the BS transmit powers p_j,i(k,n) in each block. We formulate the distributed beam scheduling problem as a non-cooperative game where each BS attempts selfishly to maximize its own payoff which is a function of the Lyapunov virtual queue parameters Z_i(k),H_j,i(k) and the transmit powers of other BSs. We then propose a parallel update algorithm to find the Nash Equilibrium of the game which serves as a distributed solution to the second sub-problem.

2012.11168

fig_solution_flow.jpg

0.434334

f6646b8a2d754076a163d5467d673e94

Distribution of the BSs and the UEs over a planar grid of dimension 800× 800 meter squared. Each BS is associated with ten UEs and these UEs are uniformly distributed in the disk coverage area with a radius of 150 meters. This BS-UE association is fixed over all epochs.

2012.11168

fig_sys_new.jpg

0.444441

719bbcfdfd4d421b84177ce220332bde

Domain(language) loss, domain UAR, and emotion development set UAR from the âRec2Iem arousal" training

2012.11174

loss.jpg

0.427854

95db83ea9f8a4319a86c444b0b4e12f8

The proposed model structure. BN and GRL are Batch Normalization, and Gradient Reversal Layer respectively.

2012.11174

model_final.jpg

0.407972

75b1acc4f3064aaf9998f01171513a27

The graph decomposition sample. The left sub-figure is the adjacency matrix of the graph in the right sub-figure, where the green and blue represent the decomposed two subgraphs.

2012.11175

3_1.jpg

0.4366

b1f3deda835d493e956dc0545fbb9725

MPG outperforms other methods in the antibacterial properties prediction open task hosted by MIT J-Clinic.

2012.11175

AI_cure_top10_new.jpg

0.425925

4afad659f36440e6a603b824d79882ce

MPG provides explainability for DDI prediction. The left part of the figure illustrates nitrate-based drugs with each atom colored by the attention weight. Note that our MPG always pays more attention to the nitrate group.

2012.11175

DDI.jpg

0.434085

8ebf15e1b16e4b53b10589dfbfb47d73

Classification performance for DTI prediction on Human (a) and C.elegans (b) datasets.

2012.11175

cpi.jpg

0.451607

f0b29170faf540aeb29409287a94e874

Two example molecules with attention weights coincided with HOMO and LUMO. The attention weights represent the importance of atoms to a molecule's global characteristics, extracted from the last layer of and normalized. The larger shading size and deeper color both denote larger attention weight. HOMO/LUMO orbitals are calculated by Materials Studio DMol3.

2012.11175

homo.jpg

0.51385

8a2b2ecfd20441f1b334fedfc7a12a12

The input representation of graph data is constructed by summing two parts: feature embedding and segment embedding. (a) Feature embedding: a set of node and edge features go through the embedding transformation to describe a graph. (b) Segment embedding: a learned segmentation embedding to every node and every edge indicating which subgraph it belongs to, different colors represent different segmentation.

2012.11175

input.jpg

0.482419

d590ca54dae54b0686cf5773c0e94080

The overview of MPG framework. The bottom left sub-figure (b) illustrates the workflow of MPG framework. The MPG framework includes two key components— and self-supervised strategies. The architecture of is shown in the bottom right sub-figure (c). The top sub-figure (a) illustrates the schemes of our self-supervised strategy, including PSD we proposed and AttrMasking for pre-training GNN model.

2012.11175

overview.jpg

0.482667

b9c774ee956e4800a127eb32c3c6710b

5-fold cross-validation classification performance on TWOSIDES dataset.

2012.11175

twosides.jpg

0.412626

55aa9219722344858020c66584303145

Visualization of the molecular representation by UMAP. The molecular representation is the collection node's embedding extracted from the last layer of non-pre-trained or pre-trained . In (a) and (b), pre-trained is capable of distinguishing valid and invalid molecules. In (c) and (d), The different colors represent different scaffolds the molecules belong to. DB index is Davies Bouldin index <cit.>, a lower DB index means that the clustering has a more appropriate separation.

2012.11175

vis_mol.jpg

0.443583

c9875cc7a5bb43999a6598b0f7526067

Situation of the proof of Claim <ref>. The dark gray disks are the points at distance at most '+(1+_n)τ from _n-1(i_n-1/p) and . The light gray hippodrome consists of the points at distance at most '+2(1+_n)τ from the segment _n-1(I_i).

2012.11181

ipeFigs.jpg

0.469978

3ef8f8a783864a14b8e3e53a3ba7a3c4

Monte Carlo histories of x^2 and x^4 for a random walk in the interval [-1,1]. Clearly the two histories are correlated.

2012.11183

hist.jpg

0.400229

5fb0251f280a43d5932797d483653bf2

The proposed NE-SGD framework for optimizing a basic ResNet18 network. The rounded rectangle shows the decoding implementation of the ResNet18 network. The solid colored region maps k_i^j=1 in the binary code. The gray area maps k_i^j=0 in the binary code. Each ResNet18 network is represented an individual in the evolutionary algorithm. The test accuracy of the model is used as the fitness value of an individual. The parental individual is selected using the tournament method for further crossover and mutation operations. The above process is iterated until the stop criteria are met.

2012.11184

Figure01new2.jpg

0.50639

4f8ff124790843929ff1d97933bb00e9

DarknetConv2D structure

2012.11185

1.jpg

0.429969

20d887b8dd9c4005bd70e77e2f761aed

Diou-Yolov3 detection results

2012.11185

10.jpg

0.470776

1778fcdac03f4c68985a38be0690e5a5

Yolov3 Loss curve

2012.11185

11.jpg

0.478891

368d5ee98e5b4ef9abf3cafd466fe81a

Diou-Yolov3 Loss curve

2012.11185

12.jpg

0.461936

62f145b0f7164112a270beb1ec64d4d2

Comparison of loss curves between two models

2012.11185

13.jpg

0.473904

fd1f38d4d6ba44e0954bb9412ae870a4

Diou-Yolov3 Accuracy curve of the model

2012.11185

14.jpg

0.548521

c059d4f652de4168962a55867d55446a

Yolov3 Accuracy curve of the model

2012.11185

15.jpg

0.488911

e822d1660d8646d9a246f25fade5e8d1

Residual block structure

2012.11185

2.jpg

0.401915

03c9aa1ba3aa4c31a7bfef3a9b2782da

I/O block diagram of yolov3

2012.11185

3.jpg

0.43637

29267922c8f146bd981676e02e5bf71e

iou loss

2012.11185

4.jpg

0.446372

2f278ceb1e2b4bf497acb95aaa2fcbeb

Diou loss

2012.11185

5.jpg

0.43137

0ebb3ce7d3634974b714f33005991194

Part of the training set

2012.11185

6.jpg

0.397723

11d467735d4241269efad68d80f064cb

Partial test set sample

2012.11185

7.jpg

0.488688

6d038929970341349ac4447fecfe5a90

Experimental flow diagram

2012.11185

8.jpg

0.394466

c8fc893531c443a88f8fbf2c1b8606c2

Yolov3 detection results

2012.11185

9.jpg

0.512359

3aa44148440c44a8b1eadad498bd219c

Core order for the graph from <ref>. Each entry of is displayed horizontally. We omit [1] since we assume μ≥ 2. The number beside each vertex in a row [μ] is the core threshold for that vertex for that μ. For example, in [2], the number .75 beside vertex 6 means that when μ = 2 and ≤ .75, vertex 6 is a core vertex. Like in <ref>, we consider the specific case where (μ, ) = (3, .6); we highlight [μ] in gray and color the core vertices (i.e., vertices with core threshold at least ) blue.

2012.11188

core-order-example.jpg

0.407288

cb333d985ace49159d9da2e7e78eee7d

Neighbor order for the graph from <ref>. In this figure, for each v ∈ V, we display [v] as a column. The numbers beside each vertex are similarity scores. For example, in [3], the .87 label beside vertex 2 represents the cosine similarity of .87 between vertices 3 and 2. Like in <ref>, we consider the specific case where =0.6 and color all -similar neighbors green and all other neighbors red.

2012.11188

neighbor-order-example.jpg

0.434094

1ce2162de7e646028c7a87d56c78352e

Example SCAN clustering with μ = 3 and =.6. The labels on the edges are cosine similarities. Core vertices are blue, whereas non-core vertices are white. Edges with similarity greater than are green, whereas other edges are red. There are two clusters (vertices 1, 2, 3, 4 and vertices 6,7,8,11) as well as three unclustered vertices (hub vertex 5 and outlier vertices 9 and 10).

2012.11188

scan-clustering-example.jpg

0.53004

7e86a2adf3804b49b4359fde138b657b

The log of the RMSE of entropy estimations versus the log of the sample size for NJEE and benchmark methods in different simulated studies. The results are the average of 100 measurements per each sample size and distribution type.

2012.11197

univariate_entropy.jpg

0.494418

dfada66572e94342b62e1a0057074970

MAE in equation (<ref>) versus the inverse temperature β when (a) p = 0.2, (b) p = 0.4, and (c) p = 0.8. These plots present the averages over 1000 experiments.

2012.11198

AIS_vs_SMCI.jpg

0.440487

a6d2f21c663247a6a8b37fcc953447c8

Results on the Hopfield-type Ising models; MAE in equation (<ref>) versus β when (a) α = 0.2 and (b) α = 0.5. These plots present the averages over 2000 experiments.

2012.11198

HFM.jpg

0.401155

b4c024ab801b4418aed7f51ae23bc694

MAE in equation (<ref>) versus K when (a) p = 0.2 and β = 0.5, (b) p = 0.2 and β = 2, (c) p = 0.8 and β = 0.5, and (d) p = 0.8 and β = 2. These plots present the averages over 1000 experiments.

2012.11198

K_dependency.jpg

0.44811

1b4028dfc6154ffba9e4f87a0b8d2306

MAE in equation (<ref>) versus N when (a) p = 0.2 and β = 0.5, (b) p = 0.2 and β = 2, (c) p = 0.8 and β = 0.5, and (d) p = 0.8 and β = 2. These plots present the averages over 1000 experiments.

2012.11198

N_dependency.jpg

0.575675

690a2b05fac14bc8bf27c254bea54ff6

MAE in equation (<ref>) versus β when (a) p = 0.2, (b) p = 0.4, and (c) p = 0.8. The results of SMCI and AIS are identical to those in figure <ref>. These plots present the averages over 1000 experiments.

2012.11198

PM_AIS_vs_SMCI.jpg

0.450878

0fabace20b5a44fbb289488011d5c0c3

Results on the bipartite Ising models; MAE in equation (<ref>) versus β when (a) p = 0.5 and (b) p = 1. These plots present the averages over 3000 experiments.

2012.11198

RBM.jpg

0.510972

bbed062e4740436cb8fc51107b9de51c

Qualitative illustration of the expected performance of the proposed method.

2012.11198

accuracy_image.jpg

0.50837

7171f9873af44cc0bb44ec063e03131a

Illustration of the target, sum, and sample regions of SMCI.

2012.11198

region_sheme.jpg

0.459331

27c9f7b11f0e4fafbf95e472fec71ba0

Comparison with PT-based method; MAE in equation (<ref>) versus β on (a) the random graph with p = 0.5 and (b) the Hopfield-type Ising models with α = 0.2. The results of SMCI and the proposed method in panel (b) are identical to those in figure <ref>(a). These plots present the averages over 2000 experiments.

2012.11198

vs_PT.jpg

0.430086

36aecf662d034ad492ec4e52c26f9408

Targeted transferability of three simple transferable attacks (CE-TMDI, Po+Trip-TMDI and Logit-TMDI) with varied step size, S.

2012.11207

curve_stepsize.jpg

0.401809

b02a7dbdbed34b8698c9478f5803c0e3

Targeted transfer success rates (%) of the logit loss compared to the C&W loss varied settings of the confidence controlling parameter K <cit.> in the single-model transfer scenario.

2012.11207

cw_vs_logit.jpg

0.437921

9455b8f4c30f45c2b5a38fc2e4b2108a

Targeted transfer success rates (%) in our challenging ensemble transfer scenario, where each hold-out target model shares no similar architecture with the source models used for ensemble.

2012.11207

ens.jpg

0.412418

a35a176a37a0413894f4ef66e7c7e6e8

Adversarial images for attacking Google Cloud Vision.

2012.11207

goo.jpg

0.413926

b1b81a13fc214e95b647693d251c3b21

Targeted universal adversarial perturbations (ϵ=16) for different classes.

2012.11207

tUAPs.jpg

0.423966

131592d738e7447480327c13f63c94fe

Unbounded adversarial images targeting different classes.

2012.11207

unbounded.jpg

0.439597

e301df78d3d54c368420e51ab6b51139

Schematic interactions within the proposed HPS framework.

2012.11208

fig.jpg

0.450433

77728515c63b47b799ffbdf492ce3c59

The diffusion constant in the J_2 channel (solid line), being the sum of the Drude approximation D_2^D plus the weak-localization correction Δ D_2 (k) in Eq. (<ref>) from two dipoles, as a function of the detuning δ = (ω-ω_0)/γ. It is normalized by the diffusion constant in the J_0 channel. The 4 weak localization corrections discussed in this section are shown separately as dashed lines.

2012.11210

DALL2norm.jpg

0.514366

2c65805a33394b869ea52bd629aff4e7

The contribution of dipole-dipole coupling to the DOS (in units of N_0 × Q_0 × (4π n/k_0^3)^2) as a function of detuning δ = (ω-ω_0)/γ. The dashed lines show the separate contributions of the modes with electric field perpendicular (P) and parallel (Q) to the separation vector 𝐫.

2012.11210

DOSPQ5.jpg

0.542205

e058fb71389d48fe888871b45a6158b4

Diagrammatic presentation of the ladder series (without external lines) involving two different electric dipoles. Dashed lines connect identical dipoles, solid lines denote the Green's tensor 𝐆_0(𝐫), crosses denote transition matrix t(k), bottom line denotes Hermitian conjugation. The first diagram on the left is reducible (a simple product) and is not part of the collision operator U_𝐩𝐩'.

2012.11210

IRRLADD.jpg

0.421103

bd0807dd30484c88bd0471ae9f56b58d

The self-consistent solution for the electromagnetic transport mean free path ℓ^* defined by σ = k_e^2 ℓ^*/6π. Shown are the values for k_eℓ^* for the full solution in this section, the conventional picture described by Eq. (<ref>) with only the transverse mode J_0 considered, with a mobility edge predicted around k_eℓ≈ 1, the lower threshold imposed by the existence of the diffusion modes J_2 and J_3, as well as k_eℓ^* associated with the fictitious conductivity and J_3. We used a cut-off q_m = 1/ℓ.

2012.11210

MEabs.jpg

0.607628

3a0842b1dc214979b4b07e04d2dc29b7

The diagrammatic convention associated with the reducible vertex R_ij|kl, 𝐩𝐩'(𝐪), with external lines. Top line denotes retarded Green's function 𝐆(k + i0), bottom line 𝐆^†(k + i0) =𝐆(k - i0) is the advanced Green's function and travel in the opposite direction. The polarization labels are ij on the left hand side (“observer”) and kl on the right hand side (“source”).The sum of incoming and outgoing wave numbers is conserved.

2012.11210

Rdia.jpg

0.48707

03603904e71b48b8a88b2daddaf5713b

Average, on-resonance, electric field in the most dense sample. Red and blue lines show the real and imaginary parts of the field, respectively. The black solid line is the fit of Eq. (<ref>) to the numerical data; the dashed line is the result for the imaginary part obtained from the fit to the real part.

2012.11210

fig_average_field.jpg

0.604535

6ddfd543e03d4b5993860b2b0363fb84

The ratio of transport and scattering mean free paths ℓ^*/ℓ as a function of k_e ℓ compared to the self-consistent theory for 4π n/k_0^3 = 3.77 and q = 0.5 (black solid line, for the two branches, see text for explanation). Points of different colors correspond to different scatterer number densities n for detunings δ = (ω-ω_0)/γ∈ [-3, 6.5] from the resonance. The dashed line is the lower bound for ℓ^*/ℓ described by Eqs. (<ref>) and (<ref>), again with q=0.5.

2012.11210

fig_comparison.jpg

0.403274

d9aa48aa47b644c88d986bffa2954c88

The average energy density, consisting of a coherent and a diffuse part. The solid straight line is a linear fit to the diffuse energy density for k_0 z = 4–9.

2012.11210

fig_energy_density.jpg

0.472575

b8247c0833804de786fccf0d2244a6ca

The effective wave number obtained from the fit to the coherent field (see Fig. <ref>), the real part of K_L(∞) calculated using Eq. (<ref>), and the ISA result.

2012.11210

fig_keff.jpg

0.445968

43d56a4285034cf5b3155687a6e8bf24

Scattering and transport mean free paths for the most dense sample. We also show the scattering length 1/2 Im K_L(∞) associated with longitudinal waves.

2012.11210

fig_tr.jpg

0.449839

f0944746e16242d3bcecaac73e6f644e

Transport mean free paths for different densities n. The dashed line shows the ISA results for the lowest density 4 π n/k_0^3 = 0.25.

2012.11210

fig_tr_all.jpg

0.468834

0730c5277813413d94d52d1cac042df4

Observing the brain tumor from different views, the FLAIR, T1, T1c, T2 indicate four modalities of MRI images. (a): The image in Axial View, (b): The image in Coronal View, (c): The image in Sagittal View.

2012.11211

Multi-view_brain_tumor.jpg

0.406308

f6529be8a2da4467ad7a50d72c06d07e

The FCN network structure and UNET network structure for the multi-view network. The network operation layer includes the initialization layer, resblock layer, conv layer, upconv layer, resize layer, logit out layer and softmax, where BN represents the batch normalization layer, Relu represents the activation function, Conv represents the downsampling operation, DeConv represents the upsampling operation, and Bilinear represents the bilinear interpolation.

2012.11211

Network_Architecture.jpg

0.474213

502bc7d948244f38a26390955be0482d

The overall architecture of the Multi-View Dynamic Fusion Framework for segmenting the 3D brain tumor from the axial view, coronal view and sagittal view.

2012.11211

Overall_Architecture.jpg

0.447402

116ecdaf0f1a4ac19643ceab7cd06d6d

Visually comparing the segmentation results of the two decision fusion methods based on the FCN segmentation network in the validation dataset. From left to right are: Input: Input images connected by four modalities data; GT: the Ground-Truth; Segmentation: Original outputs of the network; Voting: The multi-view fusion result based on the voting method of referring the axial view; WA: The multi-view fusion results based on the weighted averaging method of fusing three views.

2012.11211

Segmentation.jpg

0.481027

5d9140402d6a40dca9d75ec5342e74b1

An example of adopting the voting method with reference to the view 1 to perform the data fusion.

2012.11211

Voting.jpg

0.515806

dfc55c7ec22544ee99e81f1b3c70cdca

An example of adopting the weighted averaging method to perform the data fusion.

2012.11211

Weighted_averaging.jpg

0.580969

dc5b613ac79844dcbc18571a3e93c869

Examples of three types of backdoor attack triggers

2012.11212

3-attack.jpg

0.467364

76e8e21b7214445eb4b92004b3a4c4c7

CycleGAN structure

2012.11212

CycleGAN.jpg

0.4156

122e452111194d8f9be36527846c66e5

Style transfer by CycleGAN

2012.11212

CycleGAN_overview.jpg

0.558917

e78dbf62c4294f2493d1a8b4cab6fbe2

Examples to illustrate ABS. From left to right, the original image, the image with our trigger applied, and the image with the trigger by ABS applied. Note that here ABS is applied before detoxification. The trigger by ABS hence induces misclassification.

2012.11212

abs-result.jpg

0.459877

5cc5df88dfc14930ba577dbd204b87de

Model internals after individual rounds of detoxification for VGG on GTSRB.

2012.11212

detoxification_V_G.jpg

0.442259

812f42e777aa405fb6539957b723fb0b

Overview of detoxification

2012.11212

dex.jpg

0.449773

71c32ba0d2174ca6a9a724fe272fc2c1

Deep Feature Space Trojaning

2012.11212

featureattack.jpg

0.446447

852c4496593f4ee2bf3128001d91e514

Trigger pattern examples by Neural Cleanse. The first is the target label trigger pattern and the other two are for other labels. Note that there is not much size difference.

2012.11212

nc-result.jpg