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
|
|---|---|---|---|---|
0.414639 | e1aeb2becf334303b5633f785c39068c | eMBB SER vs transmission power in dBm for different URLLC block (segment) size. λ=7 p/msec | 2012.13923 | segmentation.jpg |
0.425869 | 5f81b39ac77a4fa89ea43ccde027246f | Estimated transition rates for male patients aged 30 and 60 years at diagnosis. Expected DCS incidence was modelled as a function of calendar time. Remaining panels are for patients diagnosed in 1995, with excess DCS incidence and other-cause (excluding DCS) mortality before DCS as functions of time since cancer diagnosis, and all-cause mortality after DCS as a function of time since DCS diagnosis. | 2012.13926 | Figure3.jpg |
0.429548 | 7d957b61ea23490a9a1bc917345ef0ba | Probability of being in each state for male patients diagnosed in 1995, aged 30 and 60 years, as a function of time since HL diagnosis. | 2012.13926 | Figure4_stacked.jpg |
0.435706 | 53ed6e1f33474bc989771a1bd7467932 | Proportion excess DCS among total DCS probability at 10 years after diagnosis, among male patients diagnosed in 1985 and 2000, by age at diagnosis. | 2012.13926 | Figure7.jpg |
0.403502 | fa016783e437411d8f5975b2b3ce8c44 | Probability of being in each state for male patients diagnosed in 1995, aged 30 years, as a function of time since diagnosis as predicted from 18 different models with varying degrees of freedom. State definitions: 1 - HL, 2 - Excess DCS, 3 - Expected DCS, 4 - Dead before DCS, 5 - Dead after DCS. | 2012.13926 | Sensitivity_analysis_age30.jpg |
0.436718 | a833defc4ce6428c8f52551df7d29bd0 | Probability of being in each state for male patients diagnosed in 1995, aged 60 years, as a function of time since diagnosis as predicted from 18 different models with varying degrees of freedom. State definitions: 1 - HL, 2 - Excess DCS, 3 - Expected DCS, 4 - Dead before DCS, 5 - Dead after DCS. | 2012.13926 | Sensitivity_analysis_age60.jpg |
0.510639 | b1fdc84c7ce54ea5a4953d253389b63f | Radio stripes network deployed over a football arena. | 2012.13928 | Picture2.jpg |
0.537015 | 0f4fe71158ef4e238788fe7e1ed0a61d | Depiction of processing of data by each AP in each time block. | 2012.13928 | Picture3.jpg |
0.999988 | af9658ec264a4bc48875c650bd3c84ce | Tree architecture of cell-free mMIMO network. | 2012.13928 | Picture4.jpg |
0.459349 | 6d48631f809d4c768b87eb52c9b09668 | Comparison of the proposed OSLP algorithm with centralized LMMSE scheme L=24 and K=10 with the CDF of the SE. | 2012.13928 | plot1_SE_oslpCent.jpg |
0.465443 | 63390f757eb0406eaad6de9462b39760 | Comparison of the proposed OSLP algorithm with other competing algorithms with varying K and fixed L=24, N=1 using the CDF of the SE. | 2012.13928 | plot2_SE_varyK.jpg |
0.433199 | c0eb57c15fde4ca199ad83c44b3146af | The CDFs of the SE for the proposed OSLP algorithm and other competing algorithms at low SNR (with p_k = 1mW ∀ k) for L=24 and N=1. | 2012.13928 | plot3_SEpow1mW.jpg |
0.396873 | 7c9e727f8bc14311824861d61b4f885c | Comparison of MSE performance of the proposed OSLP algorithm with perfect CSI when K=16 and N=1. | 2012.13928 | plot4_msePerCSI.jpg |
0.440908 | b3dc10b7f87541f3bf966996b2e9947f | Comparison of MSE performance of the proposed OSLP algorithm with imperfect CSI when K=10 and N=1. | 2012.13928 | plot5_mseImpCSI50mW.jpg |
0.419879 | 28ea05a28f6b458f8455dbc7cf5179bc | Demonstrating the monotonic increase of the SE with increasing L. | 2012.13928 | plot6_SEvaryLpow50mW.jpg |
0.428646 | fe3d6c01adca41a4be7af0cc6a0e5a4a | Illustration of the percentage of fronthaul saved for the proposed OSLP algorithm and centralized implementation. | 2012.13928 | plotFronthaul.jpg |
0.383434 | 8b33d07de9c246759b734e5a6fb03580 | The architecture of the proposed discriminator. | 2012.13936 | dis.jpg |
0.462025 | 6e0b3a9b655a46d182a9ec4058ce5fbc | The framework of the proposed generalized NR-VQA model. For each frame of the input video, we first utilize the pre-trained VGG16 network to extract the multi-scale features with an attention module. Subsequently, the extracted features are further processed by a fully connected layer to reduce its dimension, followed by a GRU module to acquire the frame-level quality features. We further regularize the frame-level quality features by forcing the features to obey Gaussian distributions via adversarial learning. Finally, a pyramid pooling strategy is utilized for temporal quality aggregation inspired by short-term and long-term memory effects. | 2012.13936 | flow.jpg |
0.456907 | 2f794710d8a941daab90ae917a6f8f06 | A few representative situations for the search of . The illustrative sensitivity values, ^* = 5, 2.5, 0.5 meV from left to right, are indicated by the horizontal black dashed lines. The black lines enclose the region allowed by neutrino oscillations, assuming normal ordering; the external gray region is forbidden. Values above the experiment sensitivity can be probed by the experiment, while those below remain inaccessible. Therefore, for each allowed value of m_1 and given ^*, there are three possible scenarios: the experiment will make an observation (green); the experiment will not see a signal (red); the experimental outcome will depend on the value of the Majorana phases (white). | 2012.13938 | cases_01.jpg |
0.45845 | 85651b70c6e44b168c53f7b88cb07b10 | Discovery probability as a function of the experimental sensitivities to for the most unfavorable scenario (black solid line, ) and the most favorable one (black dashed line, ). The colored areas express the probability for the three possible outcomes of an experiment: observing a signal even in the worst case scenario (green, observation), not observing a signal even in the best case scenario (red, inaccessibility), and when observing a signal depends on the value of the Majorana phases (white, exploration). | 2012.13938 | flag16_edit.jpg |
0.473733 | 5db2bb3d91b6461ba9dbb36892c7be03 | F_1 on English test dataset for different depths. | 2012.13939 | depths.jpg |
0.466831 | b2d902bdf1a04ccb866d9e5e33edaf2e | F_1 on English test dataset with different number of filters, x-axis is in logarithmic scale. | 2012.13939 | filters.jpg |
0.459836 | 9ce161cac86944f9a308437394c9deb8 | F_1 on English test dataset for different hash pool sizes. | 2012.13939 | hash.jpg |
0.422173 | cb650d04884f4309a729761e7ac9fe22 | Procedure for generating each filter f_i. ℋ_j outputs the component filter by looking up the shared pool E according to the bucket index produced by D_j. | 2012.13939 | hashgeneration.jpg |
0.382527 | 663dcb81aa2b4a17bfd56e518d50b38e | Proposed SRL framework. | 2012.13939 | kash-layout.jpg |
0.501696 | ec0386499dd242d6ada62b0658d6acf0 | F_1 on English test dataset for different number of importance parameters. | 2012.13939 | pij.jpg |
0.402283 | c3a3203c070f4f15872a67ebd7733b92 | Semantic role labeling example. | 2012.13939 | sent.jpg |
0.458222 | 31c68d4d12434410b280d3c0de3548ba | The figure of R(t) as a function of t | 2012.13940 | R_t.jpg |
0.435979 | c691689b7ede4029a35974c604cba2bd | Error e_H (blue, red, yellow) through time and e_H̃ (purple) through time for initial condition in H^23. | 2012.13943 | error-time-nu32.jpg |
0.447952 | fd654ed032e8435586ae3a275bd2d9e7 | Error e_u versus grid size h for the simulation of ground states with a large cubic nonlinearity. | 2012.13943 | error_groundstate.jpg |
0.467116 | 3cd632bbf1644ed7970bf99849381b71 | Left: the 15 links defined on the human body. Frames orientations and naming are based on REP-103 and REP-120. Right: the URDF kinematic model, viewed in RViz. | 2012.13944 | frames+rviz.jpg |
0.497801 | 30352c74814049408016ca16aac133db | Reference signal processing pipeline. Green nodes (left) process facial signals, blue nodes (middle) deal with the body tracking, while purple nodes (right) implement the audio processing part. Light yellow nodes, at the bottom, deal with modalities fusion, and manage the permanent personIDs. Only nodes with a strong border are present in our reference pipeline implementation. | 2012.13944 | pipeline.jpg |
0.495368 | 124f50e9676d4b7984c84bb1f81653e3 | In this situation: A is facing the robot: A gets a unique faceID, a unique bodyID, and a unique personID; B's body is visible to the robot, but not the face: B only gets a bodyID and personID; C is not seen, but heard: C gets a voiceID and a personID. | 2012.13944 | social_situation.jpg |
0.482432 | fb32071747074856a4f26554b0010a10 | The sensitivity of the axion-photon coupling g_aγ of the proposed experiment with a preliminary setup E̅_0∼ 10^3MVm^-1, R∼ 1m, assuming 10 aT magnetic field sensitivity (solid). The limit for 1 aT magnetic field sensitivity (dashed) is also shown. The CAST limit (gray) and the results from ADMX, CAPP and HAYSTAC (green) are shown. The red-line a_QCD denotes the KSVZ QCD-axion model. | 2012.13946 | constraints_w_ring.jpg |
0.448511 | 0e51c8c34522497da4ff3e8550a3892c | Enhancement of the oscillating B-field signal can be achieved by using a superconductor ring-coil pickup. A superconductor ring couples to a N turn coil, then magnetic field are boosted by N*r_2 /r_1, where N is the turnings of the coil. Comparing to a LC circuit, a ring-coil maintains a large bandwidth and does not need to be tuned to the dark matter axion frequency. | 2012.13946 | fig1-2.jpg |
0.468184 | 0abf61ffbfd941069e071b5e6c827608 | An illustration setup for the proposed experiment. The electric field E⃗ permeates the region inside the capacitor. The dark matter axions flowing along the z⃗ axis gives rise to an effective electric current density j⃗_a. The effective current density subsequently induces an oscillating magnetic field which can be picked up by a SQUID or other types of magnetometer. | 2012.13946 | fig1.jpg |
0.402831 | 0e96b0f45ec843698d85806105f9c97d | Schematic diagram of the proposed COVID-19 patient admission model under limited medical resources. Each node represents a possible state of the patients. The specific definition of each state and the transition paths among them (light blue and light red arrows with time representing the average delay of the transition) are described in detail in the text. The various quantities P_·→· denote the fractions of the state transition paths, and the time delays follow a normal distribution. The light red dotted arrows indicate the two admission paths of ICU, where ICU admission I and II have the first and second priority, respectively. The light blue dotted arrows indicate the two treatment paths of GW, which have a lower priority than ICU admission: GW admission I and II have the third and fourth priority, respectively. | 2012.13953 | figure1.jpg |
0.49587 | 5227d8295b8f45fe925bdc84a7c3dd70 | Simulation results of the daily evolution of death tolls for the two COVID-19 epicenters. (a,b) Actual data of cumulative death toll with time in Wuhan and Lombardy, respectively. The insets show the daily numbers of new deaths. The open black squares and red crosses represent the actual data and simulation results, respectively. The pink shaded area marks the 95% confidence interval. There is a good agreement between the simulated daily number of deaths and the actual data, validating our ten-state model. | 2012.13953 | figure2.jpg |
0.455788 | ebaa32eee7ab4001b61d89e937ab90a9 | Impact of timing of medical resource deployment and resource input on COVID-19 mortalities. Each panel shows the number of deaths (color coded) in the parameter plane of deployment time and resource input: panels (a,b) are for Wuhan, China, and (c,d) are for Lombardy region, Italy. In panels (a,c), the ICU resource amount is fixed and the resource that varies is GW. In panels (b,d), the GW resource amount is fixed and the resources that vary are ICU. The red hexagon indicates the true death toll under the actual resource deployment, and the white circles mark the contour of the actual death toll in the parameter plane. | 2012.13953 | figure3.jpg |
0.446502 | 303c1ae37a264b9fb25cc174eb5d4a02 | Evolution of death toll with time for different age groups. (a) Simulated evolution trend of the death toll of the three age groups in Wuhan; (b) the actual mortality rate of each age group in Wuhan in comparison with model prediction; (c,d) the corresponding results for Lombardy. The insets in (a) and (c) display the actual daily death toll and model predictions for Wuhan and Lombardy, respectively. In panels (b) and (d), the light blue and pink columns represent the actual and simulated mortality rate of each group, respectively. | 2012.13953 | figure4.jpg |
0.450483 | 82c75679ee584161969c85ad09463e39 | Dependence of death toll on weights. (a,b) Model generated death toll and its dependence on w_1 and w_3 for fixed w_2 = 30 for Wuhan and Lombardy, respectively, where the blue and purple colors correspond to relatively fewer deaths. In the regions marked by Roman numerals, the death toll changes slowly with weights. The insets show the distribution of the number of deaths for different weight combinations, where the numerical values represent the differences in the number of deaths than that from the FCFS treatment strategy. (c,d) Daily number of new deaths under different admission strategies for Wuhan and Lombardy, respectively, where red circles, gray boxes, and blue-purple forks denote the results from the three strategies: preferential treatment for elderly patients, FCFS, and preference for young patients, respectively. Giving priority to young patients results in the lowest daily death curve. | 2012.13953 | figure5.jpg |
0.529462 | f74a1fb169e94a64bb50ae4a0494ad4b | Asymptotic optimization efficiency in China, Korea, Italy, and Spain. Shown is the steady-state efficiency E(∞) versus the cases per ten million people. (a) The result for China based on the parameter values from Wuhan. Inset shows the corresponding results for Italy, South Korea, and Spain. (b) The result for China based on the parameter values from Lombardy and the results for the other three countries are displayed in inset. | 2012.13953 | figure6.jpg |
0.513645 | a47c1108c7e64419b9f7b5998934af86 | Validation of our backtracking method. The validation method is based on the data of early laboratory-confirmed cases in Lombardy area (January 28 - February 27, 2020). The incidence data of patients are estimated from the laboratory reported confirmed cases and the distribution of the time delay between the onset of the disease and the reported confirmation. The blue dots and black crosses represent, respectively, the incidence estimation curve and the retrospective survey data of laboratory-confirmed cases in the literature <cit.>. The light blue shaded region represents the 95% confidence interval of the estimation. | 2012.13953 | figureS1.jpg |
0.437773 | 9e29dfb709d74355bbf98a77ed615cdc | Remaining resources of GW and ICU in Wuhan after two admission processes. The ordinate is on a logarithmic scale, and the red and blue dots indicate that the amounts of remaining resources of Wuhan ICU and GW at time t are Δ Q_U(t) and Δ Q_G(t), respectively. The missing areas of the red and blue dots represent the period of Δ Q_U=0 and Δ Q_G=0, respectively. | 2012.13953 | figureS2.jpg |
0.445192 | ebb7029f59ef4521868e1e7753b78579 | Effects of varying ICU deployment plan in Lombardy, Italy. The black curve shows Lombardy's actual ICU medical resource deployment plan, as displayed in Tab. <ref>. The blue and red curves show the two cases where Lombardy's government would delay the deployment of ICU resources by two weeks (=14 and =1) and the ICU resource is expanded by a factor of 1.5 (=0 and =1.5), respectively. | 2012.13953 | figureS3.jpg |
0.480279 | 77368f5606734e7baa3d15f7b62e17f9 | Changes in the number of deaths and overload days for simultaneous GW and ICU resource deployment for Wuhan and Lombardy. (a,b) Numbers of deaths in Wuhan and Lombardy, respectively. (c,d) GW ward overload days in Wuhan and Lombardy, respectively. (e,f) ICU ward overload days in Wuhan and Lombardy, respectively. | 2012.13953 | figureS4.jpg |
0.468158 | e9c73586783449c9b28c597b53a9f94c | Change in overload days for fixed ICU resource deployment and varying GW resource deployment. (a,b) Overload days of GW in Wuhan and Lombardy, respectively. (c,d) Overload days of ICU in Wuhan and Lombardy, respectively. | 2012.13953 | figureS5.jpg |
0.522801 | 0b3f403fde7648fa8820658e048785e0 | Change in overload days for fixed GW resource deployment and varying ICU resource deployment. (a,b) Overload days of GW in Wuhan and Lombardy, respectively. (c,d) Overload days of ICU in Wuhan and Lombardy, respectively. | 2012.13953 | figureS6.jpg |
0.441027 | 093852f79892457086742afa8ba9948c | Mortality rate of different age groups and evolution of the number of deaths. (a) Actual data of mortality rate of different age groups in Wuhan versus model prediction. (b) Change in the number of deaths by age groups in Wuhan. (c) Actual mortality data for all age groups in Lombardy versus model prediction. (d) Number of deaths in Lombardy in each age group. In (a) and (c), the legends from top to bottom are age [0-9] to [80]. In (b) and (d), the light blue and pink cylinders, respectively, represent the actual and simulated mortalities in each group, where the x-axis marks the groups with growing age. | 2012.13953 | figureS7.jpg |
0.431187 | 5f2b4c75f6bc469d8917ad65a00794da | Weighted state age of each age group. (a) Initial state age of the three groups of patients. (b) The state age of the second group is translated by 30 units and fixed as a benchmark. (c) The state age of the first group is translated by 44 units, and that of the third group remains unchanged. (d) The state age of the third group is translated by 44 units, and that of the first group remains unchanged. | 2012.13953 | figureS8.jpg |
0.436933 | 6db15e2d1e444d7eb7b07339a9612a1d | A contractible loop with s=522 in the T model with L=400 is split into two parts (red, of k=90 steps, and black, of s-k steps). By a sequence of conformal transformations, the black part is mapped into a curve in the upper half plane stemming from the origin. | 2012.13956 | loop_seq2.jpg |
0.475497 | 221f86eabbe64ae1bbad21487e553c79 | Extrapolation of α (left) and γ (right) for the different models. The value of both exponents in the bipartite models is clearly different from the one obtained for the monopartite ones. | 2012.13956 | texalpha.jpg |
0.426644 | 2a4289bb14284d04aefded46eb80c3b2 | Example of cycle obtained comparing the minimum cost solution and an excited random solution in the random dimer covering on the honeycomb lattice with 2N=10^6, assuming periodic boundary conditions. In this case, the length of the cycle is S=10746. | 2012.13956 | toro.jpg |
0.482286 | fb3ffc7ed2a9448c8cab227c2e314a24 | Samples from a deep GP at various depths. The deep GP building block for this figure is a singleoutput GP operating on a one-dimensional input domain and using an RBF kernel (lengthscale 0.7) with a zero-mean function. An equally spaced set of evaluation points X from the interval [0, 10] is propagated through the deep GP and the function values f^(l)(X) are shown at increasing depths (l ∈{1, 2, 4, 8}). With depth increasing, function values rapidly change and assume values from a narrow range. The reason is that once samples are mapped to similar values in intermediate layers, they won't assume significantly different values in subsequent layers—see <cit.>. | 2012.13962 | deep_rbf.jpg |
0.414878 | a7f4da1676d9476d8f529227b3fe8db1 | Graphical model of a deep GP. A single input X determines where to evaluate the first GP over vector-valued functions f^(1)(·) yielding a multivariate random variable f^(1). A single sample from f^(1) determines where to evaluate the next GP over f^(2)(·) resulting in f^(2). Repeating this process until the last GP with index L yields f^(L) from which a random output for X can be sampled. | 2012.13962 | deepgp.jpg |
0.40647 | 3c3a3c52824e4e0ca28a522609b6539f | Graphical models for VI in supervised learning settings. X refers to inputs and y to labels. Unknown functions are depicted via the variable θ (in parameter space view) or f(·) (in function space view). Vanilla VI is illustrated in A) and subject of Section <ref>, whereas latent-variable VI is illustrated in B) and subject of Section <ref>. Latent-variable VI introduces another latent variable h (in addition to θ) to enable more expressive generative models. For the sake of clarity, we present in both cases one general formulation and another one in which the dataset is i.i.d. which is a common assumption in most scenarios—individual training examples are then indexed with n. | 2012.13962 | vi.jpg |
0.410219 | 5c74a9f81f5247e1b39f45dce98e1f80 | Graphical models for deep VI in supervised learning settings. X refers to inputs and y to labels. Random functions are depicted in weight space view through θ-notation (although we could have used alternatively the function space view in f(·)-notation). The general formulation is depicted in A) and the formulation for an i.i.d. dataset in B) where an index n is added to variables that factorize over samples. The input X and the first latent variable h^(1) determine where to evaluate the first random function θ^(1). This yields a random variable f^(1) determining together with h^(2) where to evaluate the second random function θ^(2), and so on. The last layer depicts the likelihood of y conditioned on f^(L). The red triangle on the left indicates (the generative part of) a Bayesian layer. Bayesian layers are stacked on top of one another and serve as building blocks for deep VI. The red rectangle on the right indicates the likelihood layer on top of the last Bayesian layer. Bayesian layers also need to maintain an approximate posterior for approximate inference over their corresponding latent variables (i.e. θ^(l) as well as h^(l) ∀ l in the general formulation, and θ^(l) as well as h_n^(l) ∀ l and ∀ n in the i.i.d. formulation), as explained in the main text in more detail. | 2012.13962 | vi_deep.jpg |
0.462964 | 5ed31f2683ad42439fcba60128adfcb3 | Graphical model of a latent-variable deep sparse GP in an i.i.d. supervised learning setting. Labels and data points are denoted as y_n and X_n respectively. An input X_n determines together with a sample from the first latent variable h_n^(1) where to evaluate the first GP over vector-valued functions f^(1)(·). This yields the random variable f_n^(1), a sample from which together with a sample from the second latent variable h_n^(2) determines where to evaluate the second multioutput GP over f_n^(2)(·), and so forth. A sample from the last random variable f_n^(L) in this cascade finally determines the probability of y_n. Inducing features are denoted as ϕ and contain inducing points Z as a special case (for Dirac features). Here, bold-face notation means that all M inducing features/points are stored in one variable for each layer l and shared across outputs for notational convenience. If we assume that inducing points/features are not shared across outputs, then each output would have its own set of M inducing points/features. Under the ordinary inducing point formulation for example, 𝐙 would then be a three-dimensional tensor (M ×number of input dimensions×number of outputs) rather than a matrix (M ×number of input dimensions). The associated inducing variables are matrices where the first dimension corresponds to the number of inducing features M and the second dimension to the number of outputs (irrespective of whether inducing features are shared across outputs or not). The total number of inducing features M is usually much smaller than the number of training examples N for reasons of computational and memory efficiency. | 2012.13962 | vi_deep_sgp_latentv.jpg |
0.43362 | 7708982f4c30435bb09f9e9290654a3e | Graphical models for VI with sparse GPs in an i.i.d. supervised learning setting. X_n are training examples and y_n are labels. Both, random functions f(·) and inducing variables , are treated as latent variables. Evaluating a random function f(·) at a specific input location X_n yields the random variable f(X_n). The latter determines the likelihood of the label y_n that corresponds to X_n. Inducing variables are associated with inducing points Z (in the inducing point formulation) or with inducing features ϕ (in the interdomain formulation). The bold-face notation means that all M inducing points (or features) and their corresponding inducing variables are collapsed into one variable. This is necessary because the individual elements of the vector are not pairwise independent. For the sake of memory and computational efficiency, the number of inducing points or features M is usually much smaller than the number of training examples N (indexed by n). The graphical model behind ordinary VI with sparse GPs is depicted in A) and subject of Section <ref>. The graphical model behind the corresponding latent-variable formulation, that requires another latent variable h_n, is depicted in B)—this is subject of Section <ref>. Also note that inducing points/features are not part of the generative model: given X_n, is not required to generate y_n under the prior. The technical reason is that the node for f(·) has two outgoing arrows. Inducing points/features are merely part of the approximate posterior to infer f(·) and are hence variational parameters. | 2012.13962 | vi_sgp.jpg |
0.441874 | 4d33b60d5a144665be0b50750b846163 | Graphical models for latent-variable VI in unsupervised learning settings. Unknown functions are denoted as θ (in parameter space view) or f(·) (in function space view). They receive the additional latent variable h as input and map it probabilistically to the observed output y. A) refers to the ordinary setting and B) to the i.i.d. setting where N different examples are index by n. | 2012.13962 | vi_unsup.jpg |
0.397785 | ee076aab198a465db2c6f6e3ee9073be | Quantitative analysis for (a) the speed of convergence and (b) the time efficiency of our Jacobian-based training method. Three learning strategies for computing IK are compared in (b). | 2012.13965 | IKconverge.jpg |
0.474599 | 2ec48fa0408e4447b84c527b97f102e2 | Pipeline of Jacobian-based method to determine the actuation parameters 𝐜_k+1 for the way-point 𝐩_k+1∈ℒ that solves the IK problem of soft robots by minimizing 𝒪(·) in (<ref>). Both the position 𝐩^r in task space and the Jacobian 𝐉^r are effectively estimated by the networks 𝒩_fk, 𝒩_J and 𝒩_s2r obtained from offline training. When 𝒪 < ϵ^2, we regard the convergence as having been achieved (e.g., 0.1% of the workspace width is employed as ϵ). | 2012.13965 | algorithmFrame.jpg |
0.469953 | 0e42e12c42b74a43b72113cee1293f3e | Comparison for studying the performance of sim-to-real learning on 1) numerical simulation-based (our approach) vs. 2) analytical computation-based (ref. <cit.>) training datasets. (a) Visualization of prediction errors in the task space. (b) Histograms show the distributions of prediction errors when using different numbers (mark as TS #) of samples to train 𝒩_s2r. | 2012.13965 | analyticalComparsion.jpg |
0.397222 | 0c5e067d94ec43c889f31899d590c371 | Two hardware setups employed in our experiments to collect data and verify the performance of our method – (a) a soft actuator with multiple chambers that are actuated by an array of syringes (see (c)) and (b) three connected soft fingers that can be actuated individually by proportional pressure regulators (see (d)). | 2012.13965 | experiment_setup_large.jpg |
0.426049 | f7906e226f324788a201afca437e3bbb | Performance of the proposed IK solver when external load is applied to the end-effector, where black shows the target trajectory. | 2012.13965 | externalLoading.jpg |
0.411418 | d96f81dbddff4a3aab16e35975fd62b4 | Interactive positioning results for the soft manipulator with three finger actuators. (a) With a user-specified position given through the software interface, our Jacobian-based method is applied to determine the IK solution. (b) The bar chat presents the tracking errors for different target positions, where the repeatability is also studied and displayed as the range of deviation in tracking errors. | 2012.13965 | fingerPositioningResult.jpg |
0.508118 | 2dbe0e2c55d3438ea8a445cf3ff815fd | (a) Path following result on the soft finger manipulator with the `8'-shape trajectory. (b) Comparison of tracking errors in the task space from the direct IK learning vs. the Jacobian-based iteration by learning (our method). (c) Visualization of IK solutions in the actuator space, where C1, C2 & C3 present the actuation parameters (i.e., pressures) in three different chambers – large variation can be found from the results obtained by direct IK learning. | 2012.13965 | fingerPositioningResultNew.jpg |
0.418054 | 65ad45cecc4649b38cc4e282611d2886 | Comparison of the behavior in path following by different speeds of motion on the three-chamber soft robot. | 2012.13965 | motionSpeed.jpg |
0.45276 | f244bd30289e4981a08b9d7a9095c09a | The network structure used in our approach to train the kinematic model and the sim-to-real transfer. | 2012.13965 | network.jpg |
0.457132 | c34e03487d9243fcbc509980de05461b | Two different configurations of motion are shown in (a) and (b), both of which are feasible IK solutions for following the `8'-shape path by the soft finger manipulator. The smoothness in motion is guaranteed by the Jacobian-based iteration for both results as shown in (c), where the actuation in every chamber has minimal variation in control parameters between neighboring waypoints. | 2012.13965 | redoundacy.jpg |
0.408692 | e019678b8f5d4719806f11d4504aa1b1 | Experimental study for the performance influence in the sim-to-real network 𝒩_s2r by using (a) different numbers of neurons and (b) datasets in different sizes. With the properly selected complexity of network structure, the over-fitting problem can be avoided. The distance predict error can be controlled within 1% of the workspace width for both setups when a limited number of training samples are used (see (b)). | 2012.13965 | sim2real_training.jpg |
0.531622 | df1de6268a7b47b0a597dca082a05174 | The results of the simulation are employed as training samples (present in black dots) to learn the forward kinematic network 𝒩_fk and the Jacobian network 𝒩_J, where these samples also span the workspace 𝒫^w of a robot. Red dots represent some example points as targets for motion in the workspace. | 2012.13965 | simulation_workingspace.jpg |
0.438388 | 43fc9f8982f241e98b792908e44cb5ac | Results for the path following task passing through the singularity region. Our method can successfully compute feasible smooth motions when meeting singularity. | 2012.13965 | singularityIssue.jpg |
0.381165 | 17082c280e8b4e219993c2ba4be2a245 | Results of path following task on two soft robots with the same design but fabricated with different materials (shown as Robot 1 and Robot 2 in Fig. <ref>(a)). (a) Trajectory of the soft robots by applying IK solutions with (solid line) and without (dash line) the sim-to-real network. (b) Visualized tracking errors on trajectory waypoints. | 2012.13965 | threeChamberTrajectoryResult.jpg |
0.493615 | d2114952b6e745239eacb1ca05772b66 | Results of path following for the 'box' shape trajectory demonstrate the vertical motion of three-chamber soft robot. It can also find that the maximal tracking errors are reduced from 4.5mm to 1.0mm for both robots after applying the sim-to-real network. | 2012.13965 | threeChamberVerticalTrajectory.jpg |
0.499325 | d133fba3f89c49deb0a9ceabd7ba3367 | Comparison of learning results by using a different number of layers as h and a different number of neurons b per layer (i.e., the total number of neurons in a network is hb). Tests are conducted on (a) the robotic setup without redundancy (the three-chamber actuator with m=n=3) vs. (b) the setup with redundancy (the planar finger manipulator having m=3 and n=2). Red crosses indicate the network parameters used in the physical experiment. | 2012.13965 | training_detail_FNN_LSTM.jpg |
0.4119 | 68d659d01b32484eb378947b8f33a0df | Examples of three antivaccine posts on Instagram. Instagram posts contain images (photographs and texts overlaid in images), captions (i.e., a poster's comments), and hashtags. In (a), if only the caption and hashtags are considered, excluding the image (e.g., the syringe and the texts embedded in the image), this post cannot be detected as an antivaccine post. In (b), if only the image that contains a tree is considered, but the caption (e.g., “a highly dangerous vaccine”) is not, this post cannot be predicted as an antivaccine post. In (c), if the image, text in the image, and the caption were considered, but the hashtags (e.g., #cdcunsafevaccine and #vaccineinjuriedchild) are not, this antivaccine post cannot be detected as antivaccine. (“Null” denotes that no textual content is detected in the post, and “OCR” denotes Optical Character Recognition.) | 2012.13968 | Figure1-AntivaxExample10ocr.jpg |
0.43323 | 5859b382eedd4663bcf605dfcf41d7b9 | The overview architecture of the proposed multimodal network for antivaccine message detection. | 2012.13968 | Figure2-ModelArchi10_ocr.jpg |
0.398548 | 5816876eec714284839e551a4772f650 | Example posts that contain #provaccine and have antithetical attitudes towards vaccination. | 2012.13968 | Figure3-Provaccine4ocr.jpg |
0.504213 | 03fcebdb328f4171b13bab2229826469 | The computation procedure of SeTa-attention weight for the hashtag branch. | 2012.13968 | Figure4-AttentionMechanism.jpg |
0.397944 | 8be82724380f4819928c8bc8c4f8e678 | Detecting antivaccine-related Instagram posts for 30 days with the proposed multimodal network. | 2012.13968 | Figure5-tracking_curve.jpg |
0.470681 | 95e1729440cd464184072ef6ec7bca08 | Flow chart of the ensemble method for antivaccine message detection. | 2012.13968 | Figure6-EnsembleMethod.jpg |
0.46601 | d2d1c9d48f784b11b8aec859cc63c67e | Examples of testing failure cases when applying the proposed multimodal network. “Null” denotes no hashtags in the post. | 2012.13968 | Figure7-FailureCases.jpg |
0.413754 | cb4a775ce9a94cd4a3b84ab5f9123e65 | Examples of antivaccine posts that are detected successfully by the proposed multimodal network but missed by single-modal networks. Both visual and textual attention mechanism results are illustrated. Red denotes attention weights for both images and texts. The OCR results are illustrated inside black boxes. “Null” denotes that no textual content is detected in the post. | 2012.13968 | Figure8-MultiRight_SingleWrong.jpg |
0.409898 | ae70773cb11a4a68947dc5b1b4b4ae36 | The schematic diagram of a four mirror cavity with two transversely coupled Bose-Einstein condensates, oriented along x-axis and y-axis. Two lasers, an external pump and a probe, drive the cavity and generate coupling between atomic states after splitting from beam splitter (BS). η (ω_E) and η_p (ω_p) correspond to the intensities (frequencies) of the external pump and probe lasers, respectively. | 2012.13969 | Fig1.jpg |
0.533595 | 72e1292949e94153803c0a50398e8b39 | (a) Absorption Re[ℰ_out] and dispersion Im[ℰ_out] of cavity probe transmission as function of normalized pump-probe detuning Δ_p/κ, for various atom-cavity couplings G_x/ω_x and G_y/ω_y. For black, blue and red curves G_x/ω_x=G_y/ω_y=0, 0.6, and 0.9, respectively. The other parameters are considered as, Δ=0, γ_x=γ_y=0.1κ, ω_x=ω_y=0.1Δ_c, and κ≈0.1×2πkHz. | 2012.13969 | Fig2.jpg |
0.464418 | 5e689d40357449a7a8e08103ba0aa316 | (a) Absorption Re[ℰ_out] versus normalized Δ_p/κ, for G_x/ω_x=G_y/ω_y=0.5 (black), G_x/ω_x=0.5,G_y/ω_y=2 (blue), G_x/ω_x=2,G_y/ω_y=0.5 (red), and G_x/ω_x=G_y/ω_y=2 (brown), at constant ω_x=ω_y=0.1Δ_c. (b) Absorption Re[ℰ_out] versus normalized Δ_p/κ, for ω_x=ω_y=0.1Δ_c (black), ω_x=0.4Δ_c,ω_y=0.1Δ_c (blue), ω_x=0.1Δ_c,ω_y=0.4Δ_c (red), and ω_x=ω_y=0.4Δ_c (black), with constant G_x/ω_x=G_y/ω_y=0.5. The rest of the parameters are the same as mentioned in Fig. <ref>. | 2012.13969 | Fig3.jpg |
0.462086 | 46a49753f7dc43929f7cd172a99416a8 | Fano resonance in absorption Re[ℰ_out] versus normalized Δ_p/κ at G_x/ω_x=G_y/ω_y=0.8. In (a), black, blue, red and brown curves correspond to Δ=0κ, 0.2κ, 0.5κ and 0.8κ, respectively. In (b), black, blue, red and brown curves correspond to Δ=0κ, -0.2κ, -0.5κ and -0.8κ, respectively. Remaining parameters are the same as mentioned in Fig. <ref>. | 2012.13969 | Fig4.jpg |
0.399417 | aacc06a69e624537abb34c6252a1dde9 | Fano resonance in Absorption Re[ℰ_out] versus normalized pump-probe detuning Δ_p/κ and normalized cavity detuning Δ/κ at G_x/ω_x=G_y/ω_y=0.6 (a) and G_x/ω_x=G_y/ω_y=1 (b). Remaining parameters are the same as mentioned in Fig. <ref>. | 2012.13969 | Fig5.jpg |
0.448423 | 875708c6749c46329fd1fdc800449f4b | The slow light behavior with the probe group delay τ_g as a function of external pump laser power P/P_p. The black, blue solid, and blue dashed curves represent G_x/ω_x=G_y/ω_y=0.3, 0.8, and 2, respectively. Remaining parameters are the same as mentioned in Fig. <ref>. | 2012.13969 | Fig6.jpg |
0.482894 | ba82b1588e4d499e92897c3ac2b85016 | The group delay τ_g versus normalized cavity detuning Δ/κ and normalized cavity decay κ/γ, where γ=γ_x=γ_y, with G_x/ω_x=G_y/ω_y=0.3 (a) and G_x/ω_x=G_y/ω_y=1 (b). The shads of blue and red colors corresponds to the distribution of fast and slow light, respectively. Remaining parameters are the same as mentioned in Fig. <ref>. | 2012.13969 | Fig7.jpg |
0.46256 | 68e238d6dae34df4af3d403e77fdefe3 | The network architecture used in our SoSN model. | 2012.13975 | blocks3.jpg |
0.486376 | 1c09e91eee3a451c93cd20ba3c8c2692 | 2012.13975 | bounds1.jpg |
|
0.487842 | 51a98fe46ce247168b4c1b09b9ecc6e3 | 2012.13975 | bounds1b.jpg |
|
0.416055 | da8c5b76a40a4b889d95afba8b90ddee | 2012.13975 | bounds2b.jpg |
|
0.421844 | 9ebd23ee2c004e809de7386b8b23cda5 | 2012.13975 | bounds4b.jpg |
|
0.416185 | a5488196b53b4262a5f09511394b08c0 | 2012.13975 | bounds5b.jpg |
|
0.523207 | 725674e1d731454280f53c90ca602a42 | 2012.13975 | bounds6b.jpg |
|
0.53633 | c50aee65c64f45cd8645bdb7c0c25251 | 2012.13975 | bounds7b.jpg |
|
0.441043 | 52b7d4c60f664cef840903c3e4f06f87 | 2012.13975 | bounds8b.jpg |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.