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
0.397471	82a1c5d0510d4be799be8516bf781184	Triggers in detoxification rounds.	2012.11212	poisonous_samples_2.jpg
0.436963	d760ec802fd74542a8219d6e86366991	Samples on GTSRB, VGG-Face and ImageNet before (the first row) and after injecting the DFST triggers (the second row), and after injecting triggers by existing attacks, including patch, Instagram filter and reflection (the third row). We use their default settings for existing attacks.	2012.11212	style_transfer.jpg
0.415758	3d2fc063475a433db703a5cd975714ea	Examples of attentions from SciBERT which attends to semantically similar tokens (training data is from the CS domain).	2012.11213	attention.jpg
0.456452	3a02a25757b64626ac4af12482ab742c	Creation of paragraph-figure pairs used as training instances for our models.	2012.11213	data.jpg
0.454051	e08167739aa64178a2c2f8949244d6e3	Our model for abstract-caption pair scoring. Paragraphs explicitly mentioning figures are paired with the figure captions during training.	2012.11213	model.jpg
0.486995	e838e092bf324f328fb0aa69522a0f6b	(Color online) Normalized (ζ≡ξ/1+ξ) non-Markovianity measure ζ (of the SSS measure ξ given in Eqs. (<ref>) and (<ref>)) with respect to the parameter p, which provides a unified description of the CP-divisible region corresponding to different waiting time distributions i.e., (λ_1 λ_2), and CP-indivisible region corresponding to same waiting time distributions (λ_1 = λ_2) regimes of the QSM process. Here s=1, with a fixed time-interval of T=1 and for the parameter p, we consider the range {0,1/2}. We can show that the semigroup form corresponds to the point where p = 0. That is, when p is this value, then γ = 0 (a constant, for a given value of s). The figure illustrates the effects of deviation from QDS; when p > s^2/8 (for s=1), we get the regime where its non-Markovian by the standard measures. The vertical line at p= 1/8 divides the CP-divisible and CP-indivisible regions of the quantum semi-Markov processes.	2012.11219	semi-Markov-plot.jpg
0.425199	6d4737cbd7f1410f824d1ba7d9fdca60	(Color online) The Holevo information χ as a function of time for various values of the parameter p, which determines the appropriate regions. Here, s=1 in all the considered cases. The bold, orange curve corresponds to CP-indivisible region with p=2; the blue, dashed curve to semi-Markov with p=1/10; and yellow dot-dashed curve to the semigroup evolution with p=1/100.	2012.11219	semi-Markov_von.jpg
0.43226	d998bc30ba3f4849913c7ee8af198cb8	(Color online) Plot of the decay rate of the semi-Markov process (<ref>), in the CP-indivisible regime.	2012.11219	semi-rate.jpg
0.457037	a8148b533f9f4389aa53893d94c8e400	Co-evolutionary interplay of opinions and links. Red (blue) links denote positive (negative) relationships. Among the three triads in the graph, the chosen one, (i, j , k) is circled. We consider the case Q = 1. As agent i flips one of its attributes, A_i^1, this triad becomes balanced and i decreases its individual stress from -5/3 to -3. The opinion vectors of l and m are A_l = A_i and A_m = - A_i, respectively (not shown in the figure). For Q = 2, the same flip of A_i^1 leads to an increase of Δ H^(i)_Q = 2 = 2/3, if one of the two chosen triads contains k, or Δ H^(i)_Q = 2 = 14/3, if none of them contains k. For Q = 3, Δ H^(i)_Q = 3 = 8/3. The higher Q, the more important becomes the role of social balance.	2012.11221	Fig1.jpg
0.418141	1af8ffb322a6405f96eaa70947a71817	(a) Clusters of positive links (red) are identified in the fragmented and balanced phase, characterized by G = 3 opinions, while in (b) a single cluster appears in the cohesive phase for G = 99. N= 100, Q = 2. Note that N^Δ = 2, 3, or 6, depending on the position of the agents on the lattice (free boundary conditions). Clusters are highlighted with colors.	2012.11221	Fig2.jpg
0.469058	627617f6c15c4e7d9276ea53dd6d0f57	: a white-box fuzzer framework for C Programs.	2012.11223	FuSeBMC_Fig.jpg
0.428126	d2452041b517489ba84ea5c552289a3d	(a) Normal self-energy (red line) and electronic spectral function (green dotted lines) obtained by CDMFT for momenta on the line k = (0,π) - (π/2,π/2). The self-energy shows two isolated poles in the anti-nodal region at ω = ±ω_0 but with a clear asymmetry in weight. In the nodal region the peak at positive energy splits. The pole at negative energy is damped is difficult to follow.(b) Anomalous self-energy obtained by CDMFT for the same momenta. It shows pairs of poles at the same position as in the normal part but with an anti-symmetric weight. The same splitting occurs when going from the anti-nodal to the nodal region. The weight close to k=(π/2,π/2) is close to zero, indicating that the coupling to the hidden-fermion is specific to the anti-nodal region.	2012.11226	Figure1.jpg
0.455874	52ecad716db74308ada14c3650f82710	Fermi surface for free electrons with dispersion given by ϵ_k = -2t ( cos(k_x)+cos(k_y) ) -4t^'cos(k_x)cos(k_y)-μ where t=1, t^'=-0.2 and μ = -0.56. The black arrow indicate the charge order wave-vector taken to link different part of the Fermi surface while the red line shows the path between k=(0,π) and k=(π/2,π/2) where the self energy is computed.	2012.11226	Figure2.jpg
0.477613	63a8b1debe0b49258e205f30df1497bf	a Normal self-energy in the hidden-fermion model of Ref.Sakai16. The asymmetry in the anti-nodal region is well reproduced. As there is only one hidden-fermion coupled to electrons it is not possible to recover the splitting in the nodal region. The weight of the positive energy pole is also maximum for k=(π/2,π/2) in contrast to the numerical results shown in Fig.<ref>. (b) Normal self-energy obtained while considering hidden-fermions due to CDW order with a wave-vector along the x axis. Due to the symmetry between ϵ_k+Q = ϵ_k-Q at k=(0,π) we only have two visible poles with an asymmetry consistent with the CDMFT results. This symmetry is lost when going to the nodal region and we observe a splitting of the poles. The change of sign of ϵ_k-Q close to k=(π/2,π/2) explains the significant weight for the pole at negative energy. We choose here Q_0 = 0.27 π, z = 0.22, Δ_0 = 0.55, V_0 = 0.7 and a numerical broadening factor iη = 0.03i.	2012.11226	Figure3.jpg
0.527603	5089714769de4a7890c3609f6df7e033	(a) Electronic spectral function (full line) and self-energy (dotted line) in the pseudogap (red) and superconducting phase (blue) in the hidden-fermion model. Above T_c, the self-energy is reduced to a single pole at ω = ϵ_k± Q > 0 and the resulting electronic spectral function shows 2 poles. There is no spectral weight at ω=0 but there will be gapless states away from k_x=0 in contrast to experimental observations. (b) Electronic spectral function (full line) and self-energy (dotted line) in the pseudogap phase from a fractionalized PDW at k=(0,π). Green lines show effects of CDW fluctuations (Γ_1) while the red lines show the effect of superconducting fluctuations (Γ_2). Due to the superposition of SC and CDW in the fractionalized PDW idea, the electronic spectral function retains the 4 poles structure seen in the superconducting phase. Finite lifetime for the particle-particle or particle-hole pairs have a much stronger dampening effect in the two poles closest to ω=0 which could explain the observed two peaks structure in CDMFT or the single occupied band in ARPES above T_c. We choose here Q_0 = 0.27 π, z = 0.22, \|Ψ_0\| = V_0 = 0.15, Γ_1 = Γ_2 = 0.1 and a numerical broadening factor iη = 0.03i.	2012.11226	Figure4.jpg
0.424722	655a5bb9b9104f7a980f03963fd0f814	(a) Electronic spectral function at ω = 0 for the fractionalized PDW model. The white dotted line indicates the non-interacting Fermi surface while the black dotted line indicates the surface of zeros of the Green's function. We see that the Fermi surface is washed out in the antinodal region. We choose here Q_0 = 0.27 π, z = 0.22, \|Ψ_0\| = 0.15 and a numerical broadening factor iη = 0.03i.	2012.11226	Figure5-1.jpg
0.472127	60125dc0d7544bf8beb85265330b465d	Compressed ratio with variation in sparsity of CSC, CSR, CIS and RLC. The picture includes two boundaries: the upper bound refers to Conv 4 of AlexNet (3×3×384) with data parallelism of 8 bits, while the lower bound is based on Conv 1 (11×11×3) represented on 32 bits.	2012.11233	Compressed_v3.jpg
0.402088	b8b478cbad1d42d3b06dc817dbc32cd2	Pseudocode of a Conv layer.	2012.11233	Conv_pseudocode.jpg
0.442688	057c34ad11974acda661d0211c8f64bc	Dense blocks used in DenseNets <cit.>. The output and input of each blue block, that is the series of two convolutions, are concatenated.	2012.11233	DenseBlock.jpg
0.513809	19075e11647a4f179f4d6c833256dd63	Data reuse in an FC layer and in a Conv layer.	2012.11233	FC_Conv_reuse.jpg
0.406914	ee005c9a258840c3ab0ab317b1148dfa	Example of an FC layer (left) and of how it can be modeled by a vector-matrix multiplication (right).	2012.11233	FC_layer.jpg
0.498474	535a365dfeb74f3ba0278269e6d9e334	Pseudocode of an FC layer.	2012.11233	FC_pseudocode.jpg
0.518698	86c13386aa364985ab619393b0c0ac3b	(left) General structure of a hardware accelerator for DNNs and (right) interconnection schemes	2012.11233	HW_acc_NoC.jpg
0.450245	5aff5af6bc9a422f893d79fb97fe858b	Inception Module used in GoogLeNet <cit.>. Convolutions are performed in four parallel branches and the four outputs are concatenated to produce the single output of the Inception module.	2012.11233	InceptionModule.jpg
0.533061	5427b3a0f3144a05bd99456f23681290	No local reuse dataflow.	2012.11233	NLR.jpg
0.541368	9af72aa9dde74fc69e1a2b73ed680b7b	Skip connection modules used in Residual Networks <cit.>. Three convolution are performed in series and a parallel connection is added. In the parallel connection, it is possible to choose between a 1×1 convolution (left) or the identity function, i.e., no operation (right). The results of the two branches are summed.	2012.11233	ResNetBlocks.jpg
0.517182	2a202020d0d6470ab09ff2b9a4b6aa45	Residual module as modified in Squeeze-and-Excitation Networks <cit.>. A skip connection is inserted in parallel to a pooling and two FC layers, and the output of the two branches are multiplied. As in traditional residual modules, a skip connection runs in parallel to the whole block.	2012.11233	SE_ResNet_Block.jpg
0.444178	ea6eebfe663e4be58f344217ecef53c0	Training techniques for SNNs.	2012.11233	SNNtraining.jpg
0.550422	fe46f810f3dd4a0bae7fa6c3e1fee08b	Taxonomy of Adversarial Attacks.	2012.11233	adversarial_examples_taxonomy.jpg
0.418697	acfc5b80f6a8454db3353b7d93f9c45b	Opportunities for employing approximate computing in deep learning.	2012.11233	approx_computing.jpg
0.494934	cdc021aae77948d88b3bff0f4a00ba2e	Model of a biological neuron, adapted from <cit.>.	2012.11233	biological_neuron.jpg
0.538437	a86484f260d741a081f030c213bfec09	Example of required bandwidth per layer executing Alexnet on a device able to perform 100 Gops/s with 100% efficiency.	2012.11233	bw.jpg
0.435859	6b7fab587eb34137a7ac843aa47810ad	Capsule Network <cit.>.	2012.11233	capsnet.jpg
0.533109	bfc91e7bd2334fa2a7672d2344e7f4ae	Graphical representation of the convolution operation in a Conv layer. A sub-tensor of Ifm (red) is multiplied by a sub-tensor of W and the results are accumulated to produce a single value (red) of the Ofm.	2012.11233	conv_layer.jpg
0.417297	f203296bff8941e7979e8ae56e6a96de	Convolution lowering: mapping a convolution to a matrix-matrix multiplication by rearranging the matrices.	2012.11233	conv_lowering.jpg
0.600175	0aaa9c81e2a84741b37d2ddfb18b253e	GFLOPS comparison between different CPUs and GPUs.	2012.11233	cpu_gpu.jpg
0.568979	8d04b536ba844724ba58f833f726e559	GFLOPS/WATT comparison between different CPUs and GPUs.	2012.11233	cpu_gpu_perwatt.jpg
0.442513	f8cfb6f932b148b9bbe5634eeb52a681	Spatial and temporal mapping of the Multiply-and-Accumulate (MACs) operations to the Processing Elements (PEs).	2012.11233	dataflow_mapping.jpg
0.397904	053040f12382439aafc2363d012136da	Development flows that run from the application to the technology and vice-versa.	2012.11233	development.jpg
0.441703	2e5a97feb7de4b8ebe88bb34daac1143	Comparison between training performed in the cloud (top) and on edge devices (bottom).	2012.11233	edge_training.jpg
0.438745	ee8ce3cc8f0f4d54ac08a4e780a6ffe4	Energy breakdown of two state-of-the-art DNNs accelerators.	2012.11233	energy_breakdown.jpg
0.421581	d2713ae1acec4391863dc35da13c2973	(top) 32-bit floating-point representation. (bottom) 8-bit dynamic fixed-point representation.	2012.11233	floating_fixed.jpg
0.468198	136320ec4adf41bfa1422cc9a6ee5531	Examples of (a) a feedforward NN and (b) a recurrent NN.	2012.11233	fwd_recurrent_nn.jpg
0.429661	273b92e07c1144dd8034a4a03867f2e2	An overview of the hardware-aware NAS problem.	2012.11233	hardware_aware_NAS.jpg
0.46111	2b5591c709624fc8a7736b1093454867	High Bandwidth Memory (HBM) scheme.	2012.11233	hbm.jpg
0.4164	2b8925aab49344fea815a60e77faf6bc	Comparison between the structures of a Von Neuman Architecture (left) and an In Memory Computing (IMC) architecture (right).	2012.11233	imc.jpg
0.504952	c8cbe4afe8d94932833bc5fea1f79463	Comparison of power consumption between conventional hardware architectures and neuromorphic architectures.	2012.11233	infographic_introduction.jpg
0.403864	f718d566d61a433c86ea1986bc9eaee6	a) Artificial intelligence overview. b) Hardware platforms comparison<cit.>.	2012.11233	intro_bea_v1.jpg
0.392261	aaa303f95b40478595f180adb7015ff2	Reduction of the number of parameters by splitting a large kernel in a series of two smaller kernels while maintaining the equivalent receptive field equal.	2012.11233	kernel_reduction.jpg
0.415609	a10b9736128e4f1bbd8e7cb7fa47005d	Features, achievable tasks and applications of the three existing ways of training (supervised, unsupervised, and reinforcement)	2012.11233	learning.jpg
0.468545	c52d3a3007cc41039474e34a47ae195c	Loop reordering of the 7-nested loops representation of the Conv layer.	2012.11233	loop_reordering.jpg
0.41083	5fb0b436eba04bfab7ca0a9e8b6df313	Discontinuity in memory accesses when performing convolution on a matrix stored by rows.	2012.11233	mem_discont.jpg
0.455916	17adbcdf002e473a997882b283354333	Timeline of significant neural networks models with the accuracies achieved on the ImageNet dataset <cit.> and the number of parameters.	2012.11233	models_ver.jpg
0.512325	d5344f7a709049ce9526ea0f3c943583	Model of an artificial neuron.	2012.11233	neuron.jpg
0.523233	abcbaffdbd6343a1a960b4cb14bc8069	Comparison of different spiking neuron models.	2012.11233	neuron_models.jpg
0.475985	9827e19fb73f4ca1a5af80433ee0be13	Different solutions <cit.><cit.><cit.> to spatially unroll the computations in an output stationary dataflow.	2012.11233	out_stat_compare.jpg
0.412633	2273afc5eaff4b558fd4793500454c90	Paper outline.	2012.11233	outline_v3.jpg
0.381445	b54d9997763e480393cfaf2ebf690133	(left) Spatial mapping of the operations in an output stationary dataflow. (right) Modified PE for an output stationary accelerator.	2012.11233	output_stat.jpg
0.431336	7d417e39120345f7b6faf311915581df	(top left) Reordered loops for an output stationary dataflow . (top right) PEs array and (bottom) temporal-spatial mapping of the operations in an output stationary accelerator with input broadcasting and weight forwarding.	2012.11233	output_stat_broadin.jpg
0.464613	02d37daf9ab34c7fa4d50bc2afb6a788	Examples of MaxPooling (left) and AveragePooling (right) layers.	2012.11233	pooling.jpg
0.589967	b9aa8851b8324c23bd1372ce48d9bdc2	Various pruning techniques.	2012.11233	pruning.jpg
0.389783	d75d67af01a74ee79fd4ee8a507a8d56	Linear, logarithmic and vector quantization techniques.	2012.11233	quantization_modified.jpg
0.456676	e759d11b78244d38ac5a12a244d6fdeb	Operations mapping in a row stationary dataflow.	2012.11233	row_stat.jpg
0.45393	9d65937bc19c4688aec2dd380e8be44d	Data movements in ShiDianNao accelerator <cit.>.	2012.11233	shidiannao.jpg
0.50027	cdb50662f18749b29b4cf3ba43477784	Basic models of temporal (left) and spatial (right) architectures.	2012.11233	spatial_temporal2.jpg
0.432007	5ebe47ba097d44cd855d55bbbf36934a	Comparison between Rate, Inter-spike interval (ISI), and Time to first spike (TTFS) encoding techniques for SNNs.	2012.11233	spike_codings.jpg
0.36753	1bc4fc231b534d6f8c81a1691d7913ff	Loop tiling technique applied to the 7-nested loops representation of the Conv layer. The outermost loops describe off-chip memory accesses, and the innermost loops determine the dataflow.	2012.11233	tiling.jpg
0.449412	fb30390fc1624212af282042d30cf961	Spatial mapping of the operations in a weight stationary dataflow.	2012.11233	weight_stat.jpg
0.455275	d295d61746084cfa9b526dc31900d960	(top) PEs array and (bottom) temporal-spatial mapping of the operations in a weight stationary accelerator with inputs broadcasting and outputs forwarding.	2012.11233	weight_stat_broad.jpg
0.405031	ead51a02b22f4b2c83fe666d4499e098	(top left) Reordered loops for a dataflow with weights stationary along dimensions C_i and C_o. (top right) Mapping the operations of a Conv layer to a matrix multiplication. (bottom) PEs array in a C_i\|C_o weight stationary accelerator with input and output forwarding.	2012.11233	weight_stat_ck.jpg
0.456766	1a55a223009948f983d1374120d5b631	(top) PEs array and (bottom) temporal-spatial mapping of the operations in a weight stationary accelerator with inputs and outputs forwarding.	2012.11233	weight_stat_forward.jpg
0.463908	21d98429260840279b703987d13753a7	Measured ROI fluorescence (blue dots) against modulation frequency ω_e at α≈ 11.72. The red line is a running mean over 50 points, acting as a guide to the eye. The dashed gray vertical lines indicate estimated maximum excitation for the breathing mode.	2012.11236	fig10_app.jpg
0.397234	a06dd89d93c54757804f72fe33193b87	Resonant light force modulation. (a) Schematic laser beam set-up. The ion crystal is almost uniformly illuminated by a cooling beam k_1 (cyan). The laser beam illuminates the crystals axial extent (z) under an angle of θ≈25 and the transverse extent (x) under and angle of φ≈45. An excitation beam k_2 (orange) is focused down to only on a fraction of the crystal. It hits the crystal z axis under an angle β≈25 and the x axis under the same angle φ as the cooling beam. The excitation beam is amplitude modulated, resulting in a sinusoidal force. The beam sizes are not to scale. (b) Example crystal photos with excited normal modes. Images taken with an EMCCD camera when the modulation frequency of the excitation laser is resonant with a normal mode. Exposure Time: 100; P_m≈20. Top: Breathing mode / (1,0). Bottom: (2,0). Red squares represent possible regions of interest to record fluorescence of single ions.	2012.11236	fig2.jpg
0.480772	f03c9da5a2c64ef2bd1f6c302b21fabc	Low frequency spectrum as a function of the aspect ratio α. The solid lines are the modes of the harmonic crystal at T=0, the symbols refer to the experimental measurements. Experimental data from Series (A) as empty symbols (blue circles, black diamonds, red thin diamonds) with P^(A)_m=20 and Series (B) as filled symbols (light blue squares) with P^(B)_m=6. For (A): The error bars represent estimated uncertainties in measured frequencies. For (B): the error bars represent fit uncertainties. The ions are laser cooled close to the Doppler limit. From comparison to simulations in section <ref>, we find T≈3.5.	2012.11236	fig3.jpg
0.409926	f99f7160071b426c926382cfdf318106	Example of a FFT signal S_z, Eq. (<ref>). The signal is extracted from the trajectory of the z coordinates of 5 molecular dynamics simulations at α=11.7 and T=0.1. The blue dots show the FFT, the dotted red line is the running mean of the signal over 30 points, the orange lines indicate automatically estimated center and widths of the peaks. The inset shows a zoom up of the first 5 peaks. See text for further details on the simulation.	2012.11236	fig4.jpg
0.419615	537557d2396249aea563b7710b7fb85d	Time evolution of transverse coordinate of ion 15. The top row is for α=11.7 and the bottom row for α=11.85. Closer to the transition at α_c≈ 12.0 the ion changes more often between the two ground state configurations, about which the crystal oscillates.	2012.11236	fig8_app.jpg
0.434879	4124298e759c4a6c8169ea02b54a2fdf	Escape rate from transition-state theory k_TST (lines) and estimate rates from simulations k_est/2 (symbols). Colors and markers are: α=11.7: blue circles & solid line, α=11.8: orange crosses & dotted line, α=11.85: green diamonds & dash-dotted line, α=11.9: red left triangles & dashed line, α=11.95: brown right triangles & double-dash-dotted line.	2012.11236	fig9_app.jpg
0.434552	8fca0b0cca144e10b88bd5bc0a4681a2	a) The quick release mount, that attaches the ALS to the aerosol transport tube. The new ALS and its parts. b) The geometry of the ALS set-up used at the experiment at FLASH.	2012.11237	als.jpg
0.453537	609bd511b2f7404aaa819a941879d6a2	Simulated beam width containing 90 % of the particles for different distances to the injector tip for 220 nm and 88 nm polystyrene spheres. The measured beam widths are marked as single points with a star, see text and <ref> for details.	2012.11237	focus.jpg
0.37945	e8688e0163074efaba3fbdb5fa9ae506	Sum of 20964 diffraction patterns collected during 220 nm polystyrene injection; experimental data is provided in the red boxes (a, d) and compared to the fitted model. The left figure shows a section of the left half (a) and the right half (d) of the upper detector. The corresponding modeled sections (b, c) are flipped at the horizontal axis to allow for a detailed comparison. (e) Radial plot comparing the angularly integrated experimental and simulated data.	2012.11237	model.jpg
0.517994	5e50a771c068438b96f044e0edcb7c7e	Measured and simulated hit fraction for a single x-ray pulse dependent on the transverse injector position relative to the x-ray focus. The data for 88 nm polystyrene was scaled by a factor of 4 for better visibility.	2012.11237	particle_beam.jpg
0.477495	9564bdc3ccb4414d867564452ab2dfe6	Stability of E^* for τ^*=20	2012.11239	E1stable.jpg
0.46879	ff8026cc1f134f0aa087ce8f0774eeaf	Stability of E^* for τ^*=20	2012.11239	E1unstable.jpg
0.389457	b045d142ada342a3a1b1798175fbedd3	Stability of E_0 for τ^*=5.8485	2012.11239	bifurcation.jpg
0.505953	b3c167e1d0e9447687f5f8cf44e91aa4	Isolation probability vs Infected	2012.11239	pi.jpg
0.407411	ee4be445823641a8a8ceb45cee81a383	Impact of temperature(T) on ℛ_0	2012.11239	r.jpg
0.434971	e3374c3ee7e24758ba5b31c00d7119c2	Impact of isolation delay(τ) on ℛ_0	2012.11239	ri.jpg
0.416785	59ab6dba28934e7495350aacf87f90a5	Impact of isolation probability(p) on ℛ_0	2012.11239	rp.jpg
0.421183	4605e298c260411d934a3d2c49b0c93d	System dynamics when R_0 <1 (β = 0.5)	2012.11239	s1.jpg
0.514291	67dbb8390aa74ec1905edc229dfea2d5	System dynamics when R_0 >1 (β = 1.0)	2012.11239	s2.jpg
0.447472	c55db69a47b040388df879a50c51cbd5	Stability of E_0 for τ^*=20	2012.11239	tau20.jpg
0.468468	63051abba9cf464a8e00948d9aa4506f	Stability of E_0 for τ^*=5	2012.11239	tau5.jpg
0.422832	70d97dd0d09542b0aa23f0e38150ce5b	Stability of E_0 for τ = 5 < τ^*=5.8485	2012.11239	tau=5.jpg
0.434989	37e77b13e0fd47bdb2a2a0a620365887	Stability of E_0 for τ =7 > τ^*=5.8485	2012.11239	tau=85.jpg
0.48896	800efeec66b24d66a9ece224862e7235	Isolation delay vs Infected	2012.11239	taui.jpg
0.451968	e6775b12274742199c382dd25fe02bdb	Death vs temperature	2012.11239	td.jpg
0.458771	c987f1fafe4b4c90b19def39c65bffbd	Infected vs temperature	2012.11239	ti.jpg
0.457453	734a1f627e794c03a48ab85b4ba0d0a2	Recovered vs temperature	2012.11239	tr.jpg
0.470984	586583f36f3b4794b6c1b03f18e590ed	Susceptible vs temperature	2012.11239	ts.jpg
0.394968	e6acaadb75cb466399fada2dba1d0f5a	Projection on the (x_1, x_2) plane of the failure domain of the function φ(x) = x_1 - 3x^2_2 - 3 considered in (<ref>) and the samples of every iteration of CE (left) and CE-m^* (right), with dimension n=100. Since φ does not depend on other coordinates, the full failure domain is a tube around this two-dimensional shape.	2012.11241	Samples_parabola.jpg

0.397471

82a1c5d0510d4be799be8516bf781184

Triggers in detoxification rounds.

2012.11212

poisonous_samples_2.jpg

0.436963

d760ec802fd74542a8219d6e86366991

Samples on GTSRB, VGG-Face and ImageNet before (the first row) and after injecting the DFST triggers (the second row), and after injecting triggers by existing attacks, including patch, Instagram filter and reflection (the third row). We use their default settings for existing attacks.

2012.11212

style_transfer.jpg

0.415758

3d2fc063475a433db703a5cd975714ea

Examples of attentions from SciBERT which attends to semantically similar tokens (training data is from the CS domain).

2012.11213

attention.jpg

0.456452

3a02a25757b64626ac4af12482ab742c

Creation of paragraph-figure pairs used as training instances for our models.

2012.11213

data.jpg

0.454051

e08167739aa64178a2c2f8949244d6e3

Our model for abstract-caption pair scoring. Paragraphs explicitly mentioning figures are paired with the figure captions during training.

2012.11213

model.jpg

0.486995

e838e092bf324f328fb0aa69522a0f6b

(Color online) Normalized (ζ≡ξ/1+ξ) non-Markovianity measure ζ (of the SSS measure ξ given in Eqs. (<ref>) and (<ref>)) with respect to the parameter p, which provides a unified description of the CP-divisible region corresponding to different waiting time distributions i.e., (λ_1 λ_2), and CP-indivisible region corresponding to same waiting time distributions (λ_1 = λ_2) regimes of the QSM process. Here s=1, with a fixed time-interval of T=1 and for the parameter p, we consider the range {0,1/2}. We can show that the semigroup form corresponds to the point where p = 0. That is, when p is this value, then γ = 0 (a constant, for a given value of s). The figure illustrates the effects of deviation from QDS; when p > s^2/8 (for s=1), we get the regime where its non-Markovian by the standard measures. The vertical line at p= 1/8 divides the CP-divisible and CP-indivisible regions of the quantum semi-Markov processes.

2012.11219

semi-Markov-plot.jpg

0.425199

6d4737cbd7f1410f824d1ba7d9fdca60

(Color online) The Holevo information χ as a function of time for various values of the parameter p, which determines the appropriate regions. Here, s=1 in all the considered cases. The bold, orange curve corresponds to CP-indivisible region with p=2; the blue, dashed curve to semi-Markov with p=1/10; and yellow dot-dashed curve to the semigroup evolution with p=1/100.

2012.11219

semi-Markov_von.jpg

0.43226

d998bc30ba3f4849913c7ee8af198cb8

(Color online) Plot of the decay rate of the semi-Markov process (<ref>), in the CP-indivisible regime.

2012.11219

semi-rate.jpg

0.457037

a8148b533f9f4389aa53893d94c8e400

Co-evolutionary interplay of opinions and links. Red (blue) links denote positive (negative) relationships. Among the three triads in the graph, the chosen one, (i, j , k) is circled. We consider the case Q = 1. As agent i flips one of its attributes, A_i^1, this triad becomes balanced and i decreases its individual stress from -5/3 to -3. The opinion vectors of l and m are A_l = A_i and A_m = - A_i, respectively (not shown in the figure). For Q = 2, the same flip of A_i^1 leads to an increase of Δ H^(i)_Q = 2 = 2/3, if one of the two chosen triads contains k, or Δ H^(i)_Q = 2 = 14/3, if none of them contains k. For Q = 3, Δ H^(i)_Q = 3 = 8/3. The higher Q, the more important becomes the role of social balance.

2012.11221

Fig1.jpg

0.418141

1af8ffb322a6405f96eaa70947a71817

(a) Clusters of positive links (red) are identified in the fragmented and balanced phase, characterized by G = 3 opinions, while in (b) a single cluster appears in the cohesive phase for G = 99. N= 100, Q = 2. Note that N^Δ = 2, 3, or 6, depending on the position of the agents on the lattice (free boundary conditions). Clusters are highlighted with colors.

2012.11221

Fig2.jpg

0.469058

627617f6c15c4e7d9276ea53dd6d0f57

: a white-box fuzzer framework for C Programs.

2012.11223

FuSeBMC_Fig.jpg

0.428126

d2452041b517489ba84ea5c552289a3d

(a) Normal self-energy (red line) and electronic spectral function (green dotted lines) obtained by CDMFT for momenta on the line k = (0,π) - (π/2,π/2). The self-energy shows two isolated poles in the anti-nodal region at ω = ±ω_0 but with a clear asymmetry in weight. In the nodal region the peak at positive energy splits. The pole at negative energy is damped is difficult to follow.(b) Anomalous self-energy obtained by CDMFT for the same momenta. It shows pairs of poles at the same position as in the normal part but with an anti-symmetric weight. The same splitting occurs when going from the anti-nodal to the nodal region. The weight close to k=(π/2,π/2) is close to zero, indicating that the coupling to the hidden-fermion is specific to the anti-nodal region.

2012.11226

Figure1.jpg

0.455874

52ecad716db74308ada14c3650f82710

Fermi surface for free electrons with dispersion given by ϵ_k = -2t ( cos(k_x)+cos(k_y) ) -4t^'cos(k_x)cos(k_y)-μ where t=1, t^'=-0.2 and μ = -0.56. The black arrow indicate the charge order wave-vector taken to link different part of the Fermi surface while the red line shows the path between k=(0,π) and k=(π/2,π/2) where the self energy is computed.

2012.11226

Figure2.jpg

0.477613

63a8b1debe0b49258e205f30df1497bf

a Normal self-energy in the hidden-fermion model of Ref.Sakai16. The asymmetry in the anti-nodal region is well reproduced. As there is only one hidden-fermion coupled to electrons it is not possible to recover the splitting in the nodal region. The weight of the positive energy pole is also maximum for k=(π/2,π/2) in contrast to the numerical results shown in Fig.<ref>. (b) Normal self-energy obtained while considering hidden-fermions due to CDW order with a wave-vector along the x axis. Due to the symmetry between ϵ_k+Q = ϵ_k-Q at k=(0,π) we only have two visible poles with an asymmetry consistent with the CDMFT results. This symmetry is lost when going to the nodal region and we observe a splitting of the poles. The change of sign of ϵ_k-Q close to k=(π/2,π/2) explains the significant weight for the pole at negative energy. We choose here Q_0 = 0.27 π, z = 0.22, Δ_0 = 0.55, V_0 = 0.7 and a numerical broadening factor iη = 0.03i.

2012.11226

Figure3.jpg

0.527603

5089714769de4a7890c3609f6df7e033

(a) Electronic spectral function (full line) and self-energy (dotted line) in the pseudogap (red) and superconducting phase (blue) in the hidden-fermion model. Above T_c, the self-energy is reduced to a single pole at ω = ϵ_k± Q > 0 and the resulting electronic spectral function shows 2 poles. There is no spectral weight at ω=0 but there will be gapless states away from k_x=0 in contrast to experimental observations. (b) Electronic spectral function (full line) and self-energy (dotted line) in the pseudogap phase from a fractionalized PDW at k=(0,π). Green lines show effects of CDW fluctuations (Γ_1) while the red lines show the effect of superconducting fluctuations (Γ_2). Due to the superposition of SC and CDW in the fractionalized PDW idea, the electronic spectral function retains the 4 poles structure seen in the superconducting phase. Finite lifetime for the particle-particle or particle-hole pairs have a much stronger dampening effect in the two poles closest to ω=0 which could explain the observed two peaks structure in CDMFT or the single occupied band in ARPES above T_c. We choose here Q_0 = 0.27 π, z = 0.22, |Ψ_0| = V_0 = 0.15, Γ_1 = Γ_2 = 0.1 and a numerical broadening factor iη = 0.03i.

2012.11226

Figure4.jpg

0.424722

655a5bb9b9104f7a980f03963fd0f814

(a) Electronic spectral function at ω = 0 for the fractionalized PDW model. The white dotted line indicates the non-interacting Fermi surface while the black dotted line indicates the surface of zeros of the Green's function. We see that the Fermi surface is washed out in the antinodal region. We choose here Q_0 = 0.27 π, z = 0.22, |Ψ_0| = 0.15 and a numerical broadening factor iη = 0.03i.

2012.11226

Figure5-1.jpg

0.472127

60125dc0d7544bf8beb85265330b465d

Compressed ratio with variation in sparsity of CSC, CSR, CIS and RLC. The picture includes two boundaries: the upper bound refers to Conv 4 of AlexNet (3×3×384) with data parallelism of 8 bits, while the lower bound is based on Conv 1 (11×11×3) represented on 32 bits.

2012.11233

Compressed_v3.jpg

0.402088

b8b478cbad1d42d3b06dc817dbc32cd2

Pseudocode of a Conv layer.

2012.11233

Conv_pseudocode.jpg

0.442688

057c34ad11974acda661d0211c8f64bc

Dense blocks used in DenseNets <cit.>. The output and input of each blue block, that is the series of two convolutions, are concatenated.

2012.11233

DenseBlock.jpg

0.513809

19075e11647a4f179f4d6c833256dd63

Data reuse in an FC layer and in a Conv layer.

2012.11233

FC_Conv_reuse.jpg

0.406914

ee005c9a258840c3ab0ab317b1148dfa

Example of an FC layer (left) and of how it can be modeled by a vector-matrix multiplication (right).

2012.11233

FC_layer.jpg

0.498474

535a365dfeb74f3ba0278269e6d9e334

Pseudocode of an FC layer.

2012.11233

FC_pseudocode.jpg

0.518698

86c13386aa364985ab619393b0c0ac3b

(left) General structure of a hardware accelerator for DNNs and (right) interconnection schemes

2012.11233

HW_acc_NoC.jpg

0.450245

5aff5af6bc9a422f893d79fb97fe858b

Inception Module used in GoogLeNet <cit.>. Convolutions are performed in four parallel branches and the four outputs are concatenated to produce the single output of the Inception module.

2012.11233

InceptionModule.jpg

0.533061

5427b3a0f3144a05bd99456f23681290

No local reuse dataflow.

2012.11233

NLR.jpg

0.541368

9af72aa9dde74fc69e1a2b73ed680b7b

Skip connection modules used in Residual Networks <cit.>. Three convolution are performed in series and a parallel connection is added. In the parallel connection, it is possible to choose between a 1×1 convolution (left) or the identity function, i.e., no operation (right). The results of the two branches are summed.

2012.11233

ResNetBlocks.jpg

0.517182

2a202020d0d6470ab09ff2b9a4b6aa45

Residual module as modified in Squeeze-and-Excitation Networks <cit.>. A skip connection is inserted in parallel to a pooling and two FC layers, and the output of the two branches are multiplied. As in traditional residual modules, a skip connection runs in parallel to the whole block.

2012.11233

SE_ResNet_Block.jpg

0.444178

ea6eebfe663e4be58f344217ecef53c0

Training techniques for SNNs.

2012.11233

SNNtraining.jpg

0.550422

fe46f810f3dd4a0bae7fa6c3e1fee08b

Taxonomy of Adversarial Attacks.

2012.11233

adversarial_examples_taxonomy.jpg

0.418697

acfc5b80f6a8454db3353b7d93f9c45b

Opportunities for employing approximate computing in deep learning.

2012.11233

approx_computing.jpg

0.494934

cdc021aae77948d88b3bff0f4a00ba2e

Model of a biological neuron, adapted from <cit.>.

2012.11233

biological_neuron.jpg

0.538437

a86484f260d741a081f030c213bfec09

Example of required bandwidth per layer executing Alexnet on a device able to perform 100 Gops/s with 100% efficiency.

2012.11233

bw.jpg

0.435859

6b7fab587eb34137a7ac843aa47810ad

Capsule Network <cit.>.

2012.11233

capsnet.jpg

0.533109

bfc91e7bd2334fa2a7672d2344e7f4ae

Graphical representation of the convolution operation in a Conv layer. A sub-tensor of Ifm (red) is multiplied by a sub-tensor of W and the results are accumulated to produce a single value (red) of the Ofm.

2012.11233

conv_layer.jpg

0.417297

f203296bff8941e7979e8ae56e6a96de

Convolution lowering: mapping a convolution to a matrix-matrix multiplication by rearranging the matrices.

2012.11233

conv_lowering.jpg

0.600175

0aaa9c81e2a84741b37d2ddfb18b253e

GFLOPS comparison between different CPUs and GPUs.

2012.11233

cpu_gpu.jpg

0.568979

8d04b536ba844724ba58f833f726e559

GFLOPS/WATT comparison between different CPUs and GPUs.

2012.11233

cpu_gpu_perwatt.jpg

0.442513

f8cfb6f932b148b9bbe5634eeb52a681

Spatial and temporal mapping of the Multiply-and-Accumulate (MACs) operations to the Processing Elements (PEs).

2012.11233

dataflow_mapping.jpg

0.397904

053040f12382439aafc2363d012136da

Development flows that run from the application to the technology and vice-versa.

2012.11233

development.jpg

0.441703

2e5a97feb7de4b8ebe88bb34daac1143

Comparison between training performed in the cloud (top) and on edge devices (bottom).

2012.11233

edge_training.jpg

0.438745

ee8ce3cc8f0f4d54ac08a4e780a6ffe4

Energy breakdown of two state-of-the-art DNNs accelerators.

2012.11233

energy_breakdown.jpg

0.421581

d2713ae1acec4391863dc35da13c2973

(top) 32-bit floating-point representation. (bottom) 8-bit dynamic fixed-point representation.

2012.11233

floating_fixed.jpg

0.468198

136320ec4adf41bfa1422cc9a6ee5531

Examples of (a) a feedforward NN and (b) a recurrent NN.

2012.11233

fwd_recurrent_nn.jpg

0.429661

273b92e07c1144dd8034a4a03867f2e2

An overview of the hardware-aware NAS problem.

2012.11233

hardware_aware_NAS.jpg

0.46111

2b5591c709624fc8a7736b1093454867

High Bandwidth Memory (HBM) scheme.

2012.11233

hbm.jpg

0.4164

2b8925aab49344fea815a60e77faf6bc

Comparison between the structures of a Von Neuman Architecture (left) and an In Memory Computing (IMC) architecture (right).

2012.11233

imc.jpg

0.504952

c8cbe4afe8d94932833bc5fea1f79463

Comparison of power consumption between conventional hardware architectures and neuromorphic architectures.

2012.11233

infographic_introduction.jpg

0.403864

f718d566d61a433c86ea1986bc9eaee6

a) Artificial intelligence overview. b) Hardware platforms comparison<cit.>.

2012.11233

intro_bea_v1.jpg

0.392261

aaa303f95b40478595f180adb7015ff2

Reduction of the number of parameters by splitting a large kernel in a series of two smaller kernels while maintaining the equivalent receptive field equal.

2012.11233

kernel_reduction.jpg

0.415609

a10b9736128e4f1bbd8e7cb7fa47005d

Features, achievable tasks and applications of the three existing ways of training (supervised, unsupervised, and reinforcement)

2012.11233

learning.jpg

0.468545

c52d3a3007cc41039474e34a47ae195c

Loop reordering of the 7-nested loops representation of the Conv layer.

2012.11233

loop_reordering.jpg

0.41083

5fb0b436eba04bfab7ca0a9e8b6df313

Discontinuity in memory accesses when performing convolution on a matrix stored by rows.

2012.11233

mem_discont.jpg

0.455916

17adbcdf002e473a997882b283354333

Timeline of significant neural networks models with the accuracies achieved on the ImageNet dataset <cit.> and the number of parameters.

2012.11233

models_ver.jpg

0.512325

d5344f7a709049ce9526ea0f3c943583

Model of an artificial neuron.

2012.11233

neuron.jpg

0.523233

abcbaffdbd6343a1a960b4cb14bc8069

Comparison of different spiking neuron models.

2012.11233

neuron_models.jpg

0.475985

9827e19fb73f4ca1a5af80433ee0be13

Different solutions <cit.><cit.><cit.> to spatially unroll the computations in an output stationary dataflow.

2012.11233

out_stat_compare.jpg

0.412633

2273afc5eaff4b558fd4793500454c90

Paper outline.

2012.11233

outline_v3.jpg

0.381445

b54d9997763e480393cfaf2ebf690133

(left) Spatial mapping of the operations in an output stationary dataflow. (right) Modified PE for an output stationary accelerator.

2012.11233

output_stat.jpg

0.431336

7d417e39120345f7b6faf311915581df

(top left) Reordered loops for an output stationary dataflow . (top right) PEs array and (bottom) temporal-spatial mapping of the operations in an output stationary accelerator with input broadcasting and weight forwarding.

2012.11233

output_stat_broadin.jpg

0.464613

02d37daf9ab34c7fa4d50bc2afb6a788

Examples of MaxPooling (left) and AveragePooling (right) layers.

2012.11233

pooling.jpg

0.589967

b9aa8851b8324c23bd1372ce48d9bdc2

Various pruning techniques.

2012.11233

pruning.jpg

0.389783

d75d67af01a74ee79fd4ee8a507a8d56

Linear, logarithmic and vector quantization techniques.

2012.11233

quantization_modified.jpg

0.456676

e759d11b78244d38ac5a12a244d6fdeb

Operations mapping in a row stationary dataflow.

2012.11233

row_stat.jpg

0.45393

9d65937bc19c4688aec2dd380e8be44d

Data movements in ShiDianNao accelerator <cit.>.

2012.11233

shidiannao.jpg

0.50027

cdb50662f18749b29b4cf3ba43477784

Basic models of temporal (left) and spatial (right) architectures.

2012.11233

spatial_temporal2.jpg

0.432007

5ebe47ba097d44cd855d55bbbf36934a

Comparison between Rate, Inter-spike interval (ISI), and Time to first spike (TTFS) encoding techniques for SNNs.

2012.11233

spike_codings.jpg

0.36753

1bc4fc231b534d6f8c81a1691d7913ff

Loop tiling technique applied to the 7-nested loops representation of the Conv layer. The outermost loops describe off-chip memory accesses, and the innermost loops determine the dataflow.

2012.11233

tiling.jpg

0.449412

fb30390fc1624212af282042d30cf961

Spatial mapping of the operations in a weight stationary dataflow.

2012.11233

weight_stat.jpg

0.455275

d295d61746084cfa9b526dc31900d960

(top) PEs array and (bottom) temporal-spatial mapping of the operations in a weight stationary accelerator with inputs broadcasting and outputs forwarding.

2012.11233

weight_stat_broad.jpg

0.405031

ead51a02b22f4b2c83fe666d4499e098

(top left) Reordered loops for a dataflow with weights stationary along dimensions C_i and C_o. (top right) Mapping the operations of a Conv layer to a matrix multiplication. (bottom) PEs array in a C_i|C_o weight stationary accelerator with input and output forwarding.

2012.11233

weight_stat_ck.jpg

0.456766

1a55a223009948f983d1374120d5b631

(top) PEs array and (bottom) temporal-spatial mapping of the operations in a weight stationary accelerator with inputs and outputs forwarding.

2012.11233

weight_stat_forward.jpg

0.463908

21d98429260840279b703987d13753a7

Measured ROI fluorescence (blue dots) against modulation frequency ω_e at α≈ 11.72. The red line is a running mean over 50 points, acting as a guide to the eye. The dashed gray vertical lines indicate estimated maximum excitation for the breathing mode.

2012.11236

fig10_app.jpg

0.397234

a06dd89d93c54757804f72fe33193b87

Resonant light force modulation. (a) Schematic laser beam set-up. The ion crystal is almost uniformly illuminated by a cooling beam k_1 (cyan). The laser beam illuminates the crystals axial extent (z) under an angle of θ≈25 and the transverse extent (x) under and angle of φ≈45. An excitation beam k_2 (orange) is focused down to only on a fraction of the crystal. It hits the crystal z axis under an angle β≈25 and the x axis under the same angle φ as the cooling beam. The excitation beam is amplitude modulated, resulting in a sinusoidal force. The beam sizes are not to scale. (b) Example crystal photos with excited normal modes. Images taken with an EMCCD camera when the modulation frequency of the excitation laser is resonant with a normal mode. Exposure Time: 100; P_m≈20. Top: Breathing mode / (1,0). Bottom: (2,0). Red squares represent possible regions of interest to record fluorescence of single ions.

2012.11236

fig2.jpg

0.480772

f03c9da5a2c64ef2bd1f6c302b21fabc

Low frequency spectrum as a function of the aspect ratio α. The solid lines are the modes of the harmonic crystal at T=0, the symbols refer to the experimental measurements. Experimental data from Series (A) as empty symbols (blue circles, black diamonds, red thin diamonds) with P^(A)_m=20 and Series (B) as filled symbols (light blue squares) with P^(B)_m=6. For (A): The error bars represent estimated uncertainties in measured frequencies. For (B): the error bars represent fit uncertainties. The ions are laser cooled close to the Doppler limit. From comparison to simulations in section <ref>, we find T≈3.5.

2012.11236

fig3.jpg

0.409926

f99f7160071b426c926382cfdf318106

Example of a FFT signal S_z, Eq. (<ref>). The signal is extracted from the trajectory of the z coordinates of 5 molecular dynamics simulations at α=11.7 and T=0.1. The blue dots show the FFT, the dotted red line is the running mean of the signal over 30 points, the orange lines indicate automatically estimated center and widths of the peaks. The inset shows a zoom up of the first 5 peaks. See text for further details on the simulation.

2012.11236

fig4.jpg

0.419615

537557d2396249aea563b7710b7fb85d

Time evolution of transverse coordinate of ion 15. The top row is for α=11.7 and the bottom row for α=11.85. Closer to the transition at α_c≈ 12.0 the ion changes more often between the two ground state configurations, about which the crystal oscillates.

2012.11236

fig8_app.jpg

0.434879

4124298e759c4a6c8169ea02b54a2fdf

Escape rate from transition-state theory k_TST (lines) and estimate rates from simulations k_est/2 (symbols). Colors and markers are: α=11.7: blue circles & solid line, α=11.8: orange crosses & dotted line, α=11.85: green diamonds & dash-dotted line, α=11.9: red left triangles & dashed line, α=11.95: brown right triangles & double-dash-dotted line.

2012.11236

fig9_app.jpg

0.434552

8fca0b0cca144e10b88bd5bc0a4681a2

a) The quick release mount, that attaches the ALS to the aerosol transport tube. The new ALS and its parts. b) The geometry of the ALS set-up used at the experiment at FLASH.

2012.11237

als.jpg

0.453537

609bd511b2f7404aaa819a941879d6a2

Simulated beam width containing 90 % of the particles for different distances to the injector tip for 220 nm and 88 nm polystyrene spheres. The measured beam widths are marked as single points with a star, see text and <ref> for details.

2012.11237

focus.jpg

0.37945

e8688e0163074efaba3fbdb5fa9ae506

Sum of 20964 diffraction patterns collected during 220 nm polystyrene injection; experimental data is provided in the red boxes (a, d) and compared to the fitted model. The left figure shows a section of the left half (a) and the right half (d) of the upper detector. The corresponding modeled sections (b, c) are flipped at the horizontal axis to allow for a detailed comparison. (e) Radial plot comparing the angularly integrated experimental and simulated data.

2012.11237

model.jpg

0.517994

5e50a771c068438b96f044e0edcb7c7e

Measured and simulated hit fraction for a single x-ray pulse dependent on the transverse injector position relative to the x-ray focus. The data for 88 nm polystyrene was scaled by a factor of 4 for better visibility.

2012.11237

particle_beam.jpg

0.477495

9564bdc3ccb4414d867564452ab2dfe6

Stability of E^* for τ^*=20

2012.11239

E1stable.jpg

0.46879

ff8026cc1f134f0aa087ce8f0774eeaf

Stability of E^* for τ^*=20

2012.11239

E1unstable.jpg

0.389457

b045d142ada342a3a1b1798175fbedd3

Stability of E_0 for τ^*=5.8485

2012.11239

bifurcation.jpg

0.505953

b3c167e1d0e9447687f5f8cf44e91aa4

Isolation probability vs Infected

2012.11239

pi.jpg

0.407411

ee4be445823641a8a8ceb45cee81a383

Impact of temperature(T) on ℛ_0

2012.11239

r.jpg

0.434971

e3374c3ee7e24758ba5b31c00d7119c2

Impact of isolation delay(τ) on ℛ_0

2012.11239

ri.jpg

0.416785

59ab6dba28934e7495350aacf87f90a5

Impact of isolation probability(p) on ℛ_0

2012.11239

rp.jpg

0.421183

4605e298c260411d934a3d2c49b0c93d

System dynamics when R_0 <1 (β = 0.5)

2012.11239

s1.jpg

0.514291

67dbb8390aa74ec1905edc229dfea2d5

System dynamics when R_0 >1 (β = 1.0)

2012.11239

s2.jpg

0.447472

c55db69a47b040388df879a50c51cbd5

Stability of E_0 for τ^*=20

2012.11239

tau20.jpg

0.468468

63051abba9cf464a8e00948d9aa4506f

Stability of E_0 for τ^*=5

2012.11239

tau5.jpg

0.422832

70d97dd0d09542b0aa23f0e38150ce5b

Stability of E_0 for τ = 5 < τ^*=5.8485

2012.11239

tau=5.jpg

0.434989

37e77b13e0fd47bdb2a2a0a620365887

Stability of E_0 for τ =7 > τ^*=5.8485

2012.11239

tau=85.jpg

0.48896

800efeec66b24d66a9ece224862e7235

Isolation delay vs Infected

2012.11239

taui.jpg

0.451968

e6775b12274742199c382dd25fe02bdb

Death vs temperature

2012.11239

td.jpg

0.458771

c987f1fafe4b4c90b19def39c65bffbd

Infected vs temperature

2012.11239

ti.jpg

0.457453

734a1f627e794c03a48ab85b4ba0d0a2

Recovered vs temperature

2012.11239

tr.jpg

0.470984

586583f36f3b4794b6c1b03f18e590ed

Susceptible vs temperature

2012.11239

ts.jpg

0.394968

e6acaadb75cb466399fada2dba1d0f5a

Projection on the (x_1, x_2) plane of the failure domain of the function φ(x) = x_1 - 3x^2_2 - 3 considered in (<ref>) and the samples of every iteration of CE (left) and CE-m^* (right), with dimension n=100. Since φ does not depend on other coordinates, the full failure domain is a tube around this two-dimensional shape.

2012.11241

Samples_parabola.jpg