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
0.46919	ef6411e3272d4d46b2cf9d6e0623c277	Exp1: Single Distributional Constraint. Balancing demographics can be represented easily through distributional constraints. By using a constraint such as _x ∼ pϕ_female(x) = 0.5, we can target balancing the female biographies in the distribution of all generations. Note that a point-wise objective _x ∼ pϕ_female(x) = 1.0 would maximize the presence of female biographies at the expense of other demographics, inducing bias in the opposite direction. The plot shows how _x ∼ pϕ_female(x) evolves towards the defined expectation: GDC is able to reduce the bias of GPT-2bio to obtain 36.7% female biographies rather than just 7%.	2012.11635	exp_401_distributional.jpg
0.423607	38c8f7b4269e4f2ea95fd28fcca8c5e6	Exp2: Multiple Distributional Constraints This experiment demonstrates the flexibility of in dealing with several distributional constraints at once, even when these constraints have different objectives (increase, decrease, or keep fixed). We challenge the flexibility of by setting four distributional constraints with four arbitrary expectation values targeting ϕ_science and ϕ_art at 40% and ϕ_sports and ϕ_business at 10%. In the figure, from left to right, we can note the increase of ϕ_science and ϕ_art from 1.5% to 20.3% and from 10% to 31.6% respectively. Interestingly, the initial ϕ_business of GPT-2bio (10.9%) is already very close to the desired expectation (10%), and we can see that during the course of the training, keeps this value fixed as it is already satisfying the corresponding target distributional constraint. ϕ_sports initially starts higher than the target distributional constraint 10%, and we can note that succeeds to reduce it from 19.6% to 11.9%.	2012.11635	exp_402_distributional.jpg
0.433375	5841e529bb6945c5ab2d763eccf09e09	Exp3: Hybrid constraints In this experiment, we specify two types of constraints: pointwise with ϕ_art(x) = 1.0 and distributional with ϕ_female(x) = 0.5 (henceforth Hybrid). in a single training procedure is able to increase the expectation of biographies about females from 7.4% to 36.6% and Art professions from 11.4% to 88.6%.	2012.11635	exp_410_distributional.jpg
0.429033	b4f659f64dab4f2b897749b2dc6079a6	Exp4: Hybrid constraints. In this experiment, we specify two types of constraints: pointwise with ϕ_sports(x) = 1.0 and distributional with ϕ_female(x) = 0.5. in a single training procedure is able to increase the expectation of biographies about females from 7.4% to 31.9% and Sports professions from 17.5% to 92.9%.	2012.11635	exp_411_distributional.jpg
0.409264	ee5379dcf0e543f18bf9fb48c15b0956	Exp5: Hybrid constraints. In this experiment, we specify two types of constraints: pointwise with ϕ_business(x) = 1.0 and distributional with ϕ_female(x) = 0.5. in a single training procedure is able to increase the expectation of biographies about females from 7.4% to 37.7% and Business professions from 10.1% to 82.4%.	2012.11635	exp_412_distributional.jpg
0.380032	031704969aed4806adae9f639214bf99	Exp6: Hybrid constraints. In this experiment, we specify two types of constraints: pointwise with ϕ_science(x) = 1.0 and distributional with ϕ_female(x) = 0.5. in a single training procedure is able to increase the expectation of biographies about females from 7.4% to 28.8% and Science professions from 1.2% to 74.7%.	2012.11635	exp_413_distributional.jpg
0.51394	23abc302b7ca495b92a436e51e979139	Case of a pointwise binary requirement r(x)=1: comparison with Reinforce and Ziegler. The curves correspond to different (·,a) levels. The manifold 𝒞 is the set of distributions c s.t. c(x) > 0 → r(x)=1, or, equivalently s.t. _x∼ c r(x) = 1. The curved lines represent increasing levels of the KL divergence (q,a). According to Reinforce, any distribution p_R s.t. _x∼ p_R r(x) =1, that is, any distribution on 𝒞, is optimal. According to Ziegler, to each temperature β > 0 is associated an optimal distribution p_Z = _q β(q,a) - _x∼ q r(x), which does not lie on 𝒞. Our own optimal p is the distribution of minimal that lies on 𝒞.	2012.11635	p_and_reinforce_and-ziegler_2.jpg
0.497317	1f6ab3d042fe4c5b8bb8fd3aec58661b	“Zipf-like” token frequency analysis on sets of 68000 generated samples from each method (only samples strictly satisfying the constraints are kept, for fair comparison). Longer tails mean a lower concentration of mass on the high frequency tokens, and therefore indicate more vocabulary richness. See Appendix <ref> for details.	2012.11635	zipf_main.jpg
0.54272	d98132bde1fe45d59c8b0ea96e1083da	Left-hand cut singularities (solid black curves) in the complex s-plane for the ππ→ππ (a) and π K →π K (b) scattering. In the plot we schematically show the position of the closest left-hand cut singularity (s_L), Adler zero (s_A), threshold (s_th) and the expansion point (s_E). Dashed lines determine the specific form of the conformal map and subsequently the domain of convergence of the conformal expansion in Eq. (<ref>).	2012.11636	1.jpg
0.419986	5086b9e9ec464415a7d6072ea1fa3c93	Results for the ππ→ππ scattering with J=0, I=0 in the single-channel case. Top, central and bottom panels correspond to m_π=physical, 236, 391 MeV, respectively. Left panels show the convergence of the conformal expansion in Eq. (<ref>), central panels show the comparison with the data and right panels show the corresponding Omnès functions. In the phase shift plot for the physical pion mass two curves are shown: fit to the experimental data <cit.> (dashed curve) and fit to the pseudo data from Roy analysis <cit.> (thick curve).	2012.11636	2.jpg
0.466998	4c0f2e42a7be40189818ec56fc9228d2	Comparison between the coupled-channel N/D fits and the Roy-like solution from <cit.>. The dashed curves are the fit solely to the experimental data, while the solid curves take advantage of both the experimental data and the results of Roy (Roy-Steiner) analyses on ππ→ππ ( ππ→ KK̅).	2012.11636	3.jpg
0.432529	a83dcdd1706949e0b4dadee0633714a8	Results for the ππ→ππ, KK̅ scattering with J=0, I=0 in the coupled-channel case. Top, central and bottom panels correspond to 11, 12 and 22 matrix elements, respectively, with 1=ππ and 2=KK̅. Left panels show the convergence of the conformal expansion in Eq. (<ref>), central panels show the comparison with the data, right panels show the elements of the Omnès matrix. In the central plots two curves are shown: fit to the experimental data <cit.> (dashed curve) and fit to the pseudo data from Roy analyses <cit.> (thick curve).	2012.11636	4.jpg
0.426777	4e886d3695e841ffa8970d477d802e44	Results for the π K →π K scattering with J=0, I=1/2 in the single-channel approximation. Top and bottom panels correspond to m_π=physical, 239 MeV, respectively. Left panels show the convergence of the conformal expansion in Eq. (<ref>), central panels show the comparison with the data and right panels show the corresponding Omnès functions. In the phase shift plot for the physical pion mass two curves are shown: fit to the experimental data <cit.> (dashed curve) and fit to the pseudo-data from Roy-Steiner analysis <cit.> (thick curve).	2012.11636	7.jpg
0.428286	e61f477644604949b7a60ce2b0addf23	E(Q_ω) as a function of M(Q_ω) for cubic-quintic ground states in dimension d=3.	2012.11637	NLS35_3d_solmEzoom-eps-converted-to.jpg
0.467461	cfbd33b8c366416e85bc6e42d9e6bf36	Solution to (<ref>) in dimension d=3 for initial data (<ref>) with ω=0.01 and λ=1.001.	2012.11637	NLS35_3d_solom001_1001water-eps-converted-to.jpg
0.449527	05b379e0358545d8be5dc8cafec16954	In-distribution anomaly detection through conditional density estimation. Consider samples of a 1D feature x and a conditional parameter of interest M (left panel), drawn from a smooth Gaussian `background' with a small number of anomalous `signal' events added (inside red circle for clarity). The conditional density values at each data point do not allow the anomaly to be distinguished from the background (center left panel), as they only identify the outliers in low density regions. However, the local over-density anomaly ratio α peaks at the anomalous data points (center right panel), and implementing a minimum cut on the anomaly ratio reveals the anomalous events (right panel).	2012.11638	method_illustration.jpg
0.513366	c19e740f50c244218ccfc61712be55e9	Hyperfine interactions of magnetic adatoms on ultrathin films: (a) Fermi contact and (b) dipolar contributions. The free-atom values are given for comparison. The number to the left in the legend indicates the number of layers in the film. The adsorption position is denoted with T for atop the anion and with B for the bridge position (for h–BN and Cu_2N only T was considered). The possible impact of corrections from van der Waals interactions was explored for the h–BN case.	2012.11639	fig1.jpg
0.389285	456161fd46b841c9aa238d245e99a44a	Basic properties of isolated 3d transition metal atoms. (a) Spin magnetic moment, (b) spin-polarized energy levels and (c) Fermi contact contribution to the hyperfine interaction of isolated 3d transition metal atoms.	2012.11639	fig2.jpg
0.471437	c5e281636874450eb017ae935fb92de1	Geometry of adatoms on MgO, NaF and NaCl. Top and side views of (a) adatom stacked on top of anion and (b) adatom in the bridge position. The adatom is represented by a blue sphere, the anion by a red sphere and the cation by an orange sphere. The definition of the distance d between the adatom and anion and the cation-anion-cation bond angle α is also illustrated.	2012.11639	fig3.jpg
0.451744	1a5035a3af6d42fc928118628c857b1f	Magnetic properties of adatoms on MgO ultra-thin films, placed on top of oxygen (a–c) and on the bridge position (d–f). (a,d) Spin magnetic moment, (b,e) Fermi contact and (c,f) dipolar contributions to the hyperfine interaction. The number to the left of MgO in the legend indicates the number of layers in the ultra-thin film. The gray dash line represent the free-atom results. The black crosses in panels (b,c) and (e,f) refers to the experimental results for Fe and Ti from Ref. and for Cu from Ref. .	2012.11639	fig4.jpg
0.456675	8a2f8a45fcda4a2b9388858813644c0e	Thickness dependence of the electronic structure of a V adatom on MgO in the oxygen-top position. (a) One monolayer of MgO. (b) Two monolayers of MgO. (c) Free atom.	2012.11639	fig5.jpg
0.443271	66260eac924542838b84077a2b3e4303	Magnetic properties of adatoms on NaF and NaCl ultra-thin films, placed on top of anion (a–c) and on the bridge position (d–f). (a,d) Spin magnetic moment, (b,e) Fermi contact and (c,f) dipolar contributions to the hyperfine interaction. The number to the left of NaF and NaCl in the legend indicates the number of layers in the ultra-thin film. The gray dash line represent the free-atom results.	2012.11639	fig6.jpg
0.450604	3a5986f0ad0642e2837cc851220c17dd	Magnetic properties of top-stacked adatoms on Cu_2N and h–BN. (a) Spin magnetic moment, (b) Fermi contact and (c) dipolar contributions to the hyperfine interaction. The gray dash line represent the free-atom results.	2012.11639	fig7.jpg
0.483159	720f7432b9254219aeae0b119c952282	The BoB system deployed during field tests in a freshwater reservoir in Singapore.	2012.11641	Fig-1.jpg
0.513874	581f2feffb924e87b27685b319aaaf40	Qualitative performance for the area coverage	2012.11641	Fig-10.jpg
0.547466	583f9805fc1e4d73962f004c7cb1718a	Average agent reward for each of the 9 agents constituting the multi-agent system	2012.11641	Fig-11.jpg
0.487314	40519372276041bcbf3428439cdaa0a3	Qualitative performance for the area coverage	2012.11641	Fig-12.jpg
0.455493	73e16f90985e43029654bd32e5bf72a9	Average agent reward for each of the 9 agents constituting the multi-agent system	2012.11641	Fig-14.jpg
0.487911	2df8ab5fefdf4c63842ce993f052690c	Average agent reward	2012.11641	Fig-5.jpg
0.503244	47c805ecadb14eed966993cdf38d3f8b	Qualitative performance for the area coverage	2012.11641	Fig-6.jpg
0.484186	bff7b97cf0384cd4a1d69dfbfe35c09b	Average agent reward	2012.11641	Fig-7.jpg
0.545885	5a6f74e472b24165922cdee810f8c87e	Qualitative performance for the area coverage	2012.11641	Fig-8.jpg
0.49491	b569f06728294ec5b7bab5856f8e60ea	Average agent reward for each of the 9 agents constituting the multi-agent system	2012.11641	Fig-9.jpg
0.523185	a3f29be3b4c84255a5311ef812425955	Evaluation protocol for direct comparison of the neural networks performance in the specific environment. The user can compare the performance of own neural network with the state of the art methods.	2012.11643	results.jpg
0.496147	1299c1978e1046e7aac515ecf9b70666	Examples of robotic arms and exchangeable grippers included in . The user can easily add any robot to the environment just by specifying its name and optionally position, orientation and initial configuration.	2012.11643	robots_selection.jpg
0.481041	710a7ba017a345a29cafa84b5ab99529	Examples of robotic workspaces. Top: Pepper robot by a desk (left) and in a maze (right). Middle: A vertical plane workspace (left), and a collaborative workspace with two KUKA arms and a humanoid (right). Bottom: Example of image augmentation during training, where random textures from a database are applied to the objects (left). A pretrained neural network Yolact can be used to preprocess the images (right).	2012.11643	workspaces_crow.jpg
0.367039	7c25a42abe1f450d924f5a7eb147ee37	C_V(ϵ/β_H) vs ϵ/β_H at T=0.	2012.11644	img1.jpg
0.368181	241aed3fe7d84af9a9bd2746421c05c1	C_𝒜(ϵ/β_H) vs ϵ/β_H at T=0.	2012.11644	img2.jpg
0.514217	613da37e8c7e46238913dd0ba3dd9bcb	Comparison between C_V and C_𝒜 at zero temperature. For large ϵ/β_H, the action complexity decays much faster than volume complexity.	2012.11644	img3.jpg
0.399931	9ce0c8fe6e974c148e8e68f82a72711b	C_𝒜(ϵ/β_H) vs ϵ/β_H at finite temperature.	2012.11644	img4.jpg
0.374115	39d0877f09854a549bea1d3a93304a86	Summary of the batch speed assessment. Absolute error values for each variance component estimated using and summarized as a boxplot.	2012.11646	error_plots.jpg
0.55313	d3ac2e531a4a4338a86f424aefdc8334	Regret averaged across all users for each week in the simulated mHealth trial associated with approaches , and .	2012.11646	seb_sim.jpg
0.464552	6a3a317e31e74e00a0ec5da3e2a565ec	A full-duplex device i transmitting to j as it receives from k in-band.	2012.11647	fdx_mmwave_system_type_13_tx_rx_v2.jpg
0.412389	79b9ac2c64254c309a7b441d7ef97a55	Sum spectral efficiency as a function of snr for various = + 20 dB, where κ = 10 dB, = 12 bits, and K_ij = K_ki = 3. While not obvious from this figure, is fairly robust to self-interference with = 12 bits. Stricter choices of = + 20 dB degrade and, thus, the sum +.	2012.11647	results-11-7a_v2_channel-avg_SNR_ij-10-dB_G_ii--55-dB_kappa-10-dB.jpg
0.422442	d6ced733bddb40348d27caa875c2ce0f	Sum spectral efficiency as a function of for various , where κ = 10 dB, K_ij = K_ki = 1, _ij = _ki = -10 dB, and b = 12 bits. As and are relaxed, the system can achieve a greater sum spectral efficiency since transmit performance is less constrained by a limited receive dynamic range.	2012.11647	results-14-2a_channel-avg_kappa-10-dB.jpg
0.420266	ca97f659534f4776ab87f00152edb9ab	Sum spectral efficiency as a function of κ for various and , where K_ij = K_ki = 1, _ij = _ki = -10 dB, and b = 12 bits.	2012.11647	results-14-4a_channel-avg.jpg
0.434869	f48032db06bc4cb680a0056484b2476d	Difference Between the Best ≤4-Average Mixture and the Simple Average Mixture, Eurozone Real Interest Rate	2012.11649	fig_heatmap_diff_subsetleN_avg_uni_real.jpg
0.483311	e445eb5215b34c8bb5a677ffbe68c468	PIT Histograms, Eurozone Inflation	2012.11649	fig_pit_uniform_4panel_v2.jpg
0.460178	d44d76c5fcb04f28ba2b69a4b73feb7d	The concentration u obtained as a PINN solution of ADE (<ref>) at (a) t=2 min, (b) t=6 min, and (c) t=20 min. The absolute point errors with respect to the STOMP solution are shown in subfigures (d–f). The DNN size is 5 × 60.	2012.11658	Plot_2Dt_CAD_K05.jpg
0.433026	260d5f6dd4e44b159679f17b21b54f5b	Same as in Figure <ref>, but with Pe=200. For t = 0.5, the maximum point errors for w=1, 10, and 100 are 2.21 × 10^-3, 1.14 × 10^-3, and 1.07 × 10^-3, respectively.	2012.11658	Plot_2Dt_Pts_Pe200.jpg
0.403868	dbc9a1053c1e4566b5eb8a4c3db32e9e	Snapshots of the PINN solution of the time-dependent ADE (<ref>) with Pe = 50 (left collumns) and the corresponding point errors with respect to the analytical solution (<ref>) (right columns) obtained with: (a) w=w_IC = w_BC = 1, (b) w=w_IC = w_BC = 10, and (c) w=w_IC = w_BC = 100. For t = 0.5, the maximum point errors for w=1, 10, and 100 are 2.96 × 10^-3, 1.33 × 10^-3, and 0.97 × 10^-3, respectively.	2012.11658	Plot_2Dt_Pts_Pe50.jpg
0.478735	74821f9891b34de281be11f1a060dc00	Upper row: the PINN solution of the backward ADE (<ref>) at t = 0 with (a) Pe = 1, (b) Pe = 10, and (c) Pe = 200. Lower row: the absolute point errors with respect to the reference solutions. The maximum point errors for these three cases are 8.07 × 10^-3, 1.23 × 10^-2, and 1.40 × 10^-2, respectively.	2012.11658	Plot_2Dt_Pts_back.jpg
0.465322	ab86ded7fd7c4549bab8350a499fc91d	The reference concentration u obtained as a STOMP solution of the ADE (<ref>) at different times.	2012.11658	Plot_2Dt_bCAD_0_10min_C_K05.jpg
0.396537	dbca953692b7467997306345ff39431b	The absolute errors in the PINN solution û(x⃗,t,θ) with respect to the reference field u(x⃗,t) at different times. The DNN size is 5 × 60.	2012.11658	Plot_2Dt_bCAD_0_10min_err_K05.jpg
0.450069	614dea558334424691b0a80ee37a9727	Absolute error in the PINN solution of the backward ADE with data assimilation. The absolute errors computed with respect to the reference u(x⃗,t) field. The DNN size is 5 × 60.	2012.11658	data_assimilation.jpg
0.418507	3feb035a781044c5addc49f7ca0b1eb3	The comparison of the PINN and finite elements solutions for the ADE with the boundary conditions and velocity vector shown in Figure <ref> for Pe=1 and 100. The PINN solution is obtained with N_f = 41 × 41 residual points, the DNN size 3 × 30 and w_BC = 50. The GFEM and SUPG solutions in subfigures a, b, d and e are reproduced from <cit.>.	2012.11658	fig_advdis_skew_updated.jpg
0.479669	ebd513b100534a7aa549c82d38d5fb7d	Boundary conditions for steady-state ADE with the point source located at (0,0). The velocity vector is v⃗=[cos(π/4), sin(π/4)]^T.	2012.11658	sche_ad_skew.jpg
0.479527	4a55699553b84ff081d90739dacc24fb	Pulse schedules corresponding to the (a) scaled circuit and (b) basis circuit, corresponding to gates given in Fig. <ref>, with the end of the scaled schedule marked by the vertical gray dashed line. Here, the `d' and `u' correspond to the (on-resonant) s and (off-resonant) s of Qiskit Pulse, respectively. The pulses are shaded according to phase, with light pulses in-phase and dark pulses in the quadrature phase. The circular arrows correspond to Z-rotations executed in software by changing the phase of subsequent pulses <cit.>. The duration is given in units of sample time, dt = 0.222 ns.	2012.11660	fig-sched-annotate.jpg
0.43305	c9796e5fdeca4029aa5825e89340fb19	Gate sets that implement a single Suzuki-Trotter step t_i of the time evolution. (a) Implementation with 2-qubit basis gates; each set of gates is labeled by the corresponding Hamiltonian terms above. (b) Implementation with analog scaled 2-qubit ECR gates A(θ), where ZX(θ) = R_x(π/2) A(θ) R_x(π/2), as shown in Fig. <ref>c.	2012.11660	fig2-annotate.jpg
0.419645	3951d9d4b18e4e1fb62bbcc6f752b91e	Quantum process tomography of U_yx(θ) as a function of θ on ibmq_athens. Both the double-CNOT circuit (a) and the parametric ZX(θ)-based circuit (b) create U_yx(θ). The circuit in (b) at θ=π/2 is implemented by the pulse schedule shown in (c). Here, each sample lasts 0.222 ns. (d) Fidelity F_(a) (blue up-triangles) and F_(b) (orange down triangles) of quantum circuits (a) and (b), respectively, obtained with QPT. (e) Relative error reduction of the pulse-efficient circuit in (b) over the circuit in (a), i.e. 1-E_(b)/E_(a) where E_(x)=1-F_(x).	2012.11660	majorana_qpt.jpg
0.499569	501351d4728344609fb66f3bf31a780b	Scaling in different dimensions	2012.11662	Fractaldimensionexamplebw.jpg
0.437598	a14638ece943413fbba81a8918092292	Calculating the mesh dimension	2012.11662	Snow_LineFitAnn.jpg
0.511456	ea066e80bfaa44368531f847acdf6dee	A non uniform mesh	2012.11662	Snow_Mesh_Example.jpg
0.387671	84ddf84d868b4f6a84279078a266ee5e	HalfCheetah	2012.11662	cheetah.jpg
0.389821	e3f8e533311e463bac493a4a9dab7f62	HalfCheetah	2012.11662	cheetah.jpg
0.501832	c6d5b4cdb9734267a3e4145a0b1a0da9	Top: mesh sizes vs log of the box size for the cheetah environment. Lower Left: Every five frames overlaid for the an identity policy on the cheetah. Lower Right: Every five frames of cheetah after the lower mesh dimension training.	2012.11662	cheetah_mesh_cor.jpg
0.43681	5da29009ae584e6a9050bf078db97e17	Hopper	2012.11662	hopper.jpg
0.44734	6267097f46794e8e8be12c1cb09b0774	Hopper	2012.11662	hopper.jpg
0.41274	aa8a5974eb784681847510928422f490	High resolution curve of a well behaved policy	2012.11662	meshdim.jpg
0.393872	2b9bb940a23942e88d3e90d11298fc99	Run time resolution curve of a typical policy	2012.11662	meshdimreal.jpg
0.401837	beb639c1b95e4d138378d4b615122934	Walker	2012.11662	walker.jpg
0.432241	c1a8427a76b04abbb65f37255068e043	Walker	2012.11662	walker.jpg
0.53746	1ce5aa2d0c594779919f2a4068a5c1cd	Torque exerted during the sampled episode	2012.11663	act_hist.jpg
0.483385	09e45fa90ccb4fadaaa9e9bf0cc23f59	Observations during the sampled episode	2012.11663	obs_hist.jpg
0.436918	28f6a3edb70e4036b9a70da592f378aa	Reward curve for SSAC and the other algorithms we compare to. the solid line is the smoothed average of episode reward, averaged over four random seeds. The shaded area indicates the best and worst rewards at each epoch across the four seeds. SSAC is shown starting later to account for the time training the gating function alone.	2012.11663	reward.jpg
0.479445	3dcee0076f55452593987a78a7d1ab73	System diagram for the new technique proposed in this paper. Rounded boxes represent learned neural networks, squared boxes represent static, hand crafted functions. The local controller is a hand designed LQR, the swing-up controller is obtained via reinforcement learning, and the gating function is trained as a neural network classifier	2012.11663	system5.jpg
0.390179	1c51473921a144258adefad4fe2899c2	Balance map for TD3, X and Y indicate the initial position for the trial, a black dot indicates that the trial started from that point satisfies equation (<ref>), and red indicates the converse. when these initial conditions were run for SSAC, it satisfied (11) for every initial condition	2012.11663	th_map_td3.jpg
0.410123	212d0a920e9a436ca55b28c3ed4dd730	Attenuation factor of the emitted radiation by γγ-annihilation. Internal absorption is caused by interaction with the photons created in the acceleration region by the nonthermal radiative processes. The external stellar radiation field provides the target photons for stellar absorption.	2012.11664	attenuation.jpg
0.479103	04d649e6ff88425b902732301007e9c6	Rates of acceleration and radiative losses for electrons.	2012.11664	electron_cooling.jpg
0.405023	efa3bccc8b5c4890a95302bf975a236a	Distribution of relativistic electrons in the stationary one-zone model for a highly efficient acceleration.	2012.11664	electron_distribution_1.jpg
0.411754	79f14791e7244112ac0f094b71c63aab	Distribution of relativistic electrons in the stationary one-zone model for a low value of acceleration efficiency.	2012.11664	electron_distribution_2.jpg
0.373131	8b972c93fec84858a3784bfa97bf9ce2	Rates of acceleration and radiative losses for protons.	2012.11664	proton_cooling.jpg
0.429701	518561fc2fbf4f72878bf8d574e40601	Scheme of the interaction between the equatorial wind and the donor star. The wind stagnates at the point where ρ_ w v_ w^2 = P_ rad. The distances from the stagnation point to the black hole and the center of the star are indicated as r_ BH and r_ S, respectively.	2012.11664	sketch2.jpg
0.570172	35b744d5df7d4320ab113ee8d87d5f46	Spectral energy distribution of the nonthermal radiation emitted in the acceleration region. Nonthermal radiative processes of the secondary pairs are considered for the calculation of the total spectrum.	2012.11664	total_SED.jpg
0.47117	de638ddcff534998be006c5e35dd0f6f	Scheme of the considered model of accretion disk and winds in super-Eddington microquasars. Toroidal magnetic fields are expected in the disk. As long as there is no poloidal magnetic field, there is no launching of jets. The figure corresponds to the parameters adopted in Sotomayor Checa & Romero (2019).	2012.11664	wind_disk.jpg
0.392493	e6c8f1c4cc374519b8c233854bb5bae8	Chavissa et al. 2020<cit.> show MIRC-X aperture synthesis imaging of surface of CL Lac, AGB star. The MiRA reconstructed images at 1.52 μm (left), 1.70 μm (middle) and difference between the two (right). These images show presence of a brighter area that largely affects the position of the photocenter and that likely caused Gaia DR2 astrometric error. Reproduced from Chavissa et al. 2020<cit.>.	2012.11666	Chavissa_AGB_image.jpg
0.421479	0d75d3c1663b443eb656cfe3d1d10991	Kraus et al. 2020<cit.> observed the first disk tearing in a protoplanetary hierarchical triple, GW Ori. (A & B) ALMA images with 20mas resolution showing disk and misaligned rings. The MIRC-X measured inner triple stellar orbit allowed to link the detected misaligned ring to the disk tearing effect. (C) The MIRC-X reconstructed images of inner triple over-plotted with the best-fitting orbital model. The white dot marks the center-of-mass of the system. Figures reproduced from Kraus et al. 2020<cit.>.	2012.11666	GW_Ori.jpg
0.430746	6e27891f84534200961a5b13cf2b9ff9	Gardner et al. 2020<cit.> show discovery and measured orbital elements of a 30-day companion to the B-type binary α Del. (Left) The outer binary orbit of α Del as the Ba component moves relative to the A component fixed at the origin. (Right) Detection of an added astrometric wobble due to the presence of the Bb component. Reproduced from Gardner et al. 2020<cit.>.	2012.11666	alpha_Del.jpg
0.492836	68d38dc8c30d48f6807fadc6663b1922	The Tel-AO software layout. The camera link server reads the frames from the WFS Andor EMCCD camera through BitFlow Neon frame grabber and writes them to an instrument shared memory in a circular buffer so that the main WFS main server uses them. The WFS main server reads the frames from the instrument shared memory and calculates the wavefront slopes from each new science frame. Next, the DM commands are computed by multiplying the slopes matrix with the slopes-to-actuator reconstructor matrix. The tip/tilt is computed by taking the average of slopes and then sent to telescope M2 for correction with appropriate gain. The rest higher-order corrections are sent to the ALPAO deformable mirror. The wavefront sensor images and deformable mirror shapes are sent to the client GUI for monitoring and operation.	2012.11667	TELAO-software.jpg
0.452997	e85024670b8a41e8be15aaab7b7b14e8	Photographs of the Tel-AO (left) and Lab-AO systems (right) as built.	2012.11667	Tel-AO_and_Lab-AO.jpg
0.417485	2f1aaf2c6e0a442da4bb9986fab79eb0	USB data data acquisition of frames (black color) in a comparison to the camera-link data acquisition (blue color). The integration is 3ms but the data points are the time difference between two consecutive frame grabs. The USB acquisition has a lot of jitters, in a worst-case, a grab gets 8 frames in a chunk (i.e., 8 frames of delay). In the case of camera-link, there is also jitter but very small.	2012.11667	USB_missing_frames.jpg
0.417306	58dabbafc68645d5a2717aa7b4442e72	A photograph of the main observing GUI at the CHARA Array, . Each controls one telescope and has all the controls for the telescope and its acquisition camera, tip-tilt, Tel-AO, and Lab-AO, including the real-time displays of the acquisition camera and wavefront sensors of Tel-AO and Lab-AO. This GUI is optimized for screen space – all six GUIs fit in a single screen monitor, time-efficient usage, and stability by keeping only the required observing controls together. This photograph is taken with telescope beacon was ON.	2012.11667	obsgtk_6T.jpg
0.50631	a951c735c6af4126aa9e1e244339b58d	Rejection transfer function measured using the telescope beacon as the source of light. The green dotted line is the measured closed-loop transfer function. The dashed line is the modeled open-loop transfer function. The thick black line is the modeled closed-loop transfer function. This model uses the total delay time T_ D is used. The green line cross the measured -3dB closed-loop bandwidth at frequency of 19Hz.	2012.11667	telao_model.jpg
0.422094	2aa2ac4c56404669b13fbf403b758d32	(Left) The architecture of CCID75 reproduced from Schuette et al. (2014)<cit.>. It has 160 (Horizontal) × 160 (Vertical) imaging area – subdivided into twenty 16 (H) × 80 (V) pixel channels. Each channel has a separate readout chain (serial readout register and output amplifier). (Right) Quantum efficiency is around 80% for 0.6-0.9μm wavelengths.	2012.11668	CCID75_architecture.jpg
0.445979	aaffc375f5ae40bd9d6da9941d19fda8	Computer architecture of MAPS AO system. The frame grabbers of the VPyWFS-CCID75 and the IRPyWFS-SAPHIRA and the RJ45 Ethernet of the acquisition camera are installed on the server computer. The INDI instrument drivers, , and , which are installed on computer, read frames from all the cameras. The and compute slopes from the detector images and then send to over a dedicated Ethernet socket in a “WFS loop" operated at the WFS frame rate, 1kHz. driver works on the computer runs a second loop “ASM loopâ independent of the WFS loop and converts the WFS slopes to the ASM commands and send to the ASM actuators to apply correction. The allow engineering with actuators and the does house keeping and safety.	2012.11668	Computer_architecture1.jpg
0.409986	c6b10bdac4424c5f8d42f3e52b983fa4	The computer handles non-real-time Input and Output devices such as motors and actuators, web Graphical User Interfaces (GUI) and telemetry storage. For example, is web socket server. The is an INDI driver for controlling motors. The is an INDI driver allows to play with ASM administrative states.	2012.11668	Computer_architecture2.jpg
0.422416	ed1f7d6e65a24eeaa067d70281c82580	Expected performance of MAPS compared to legacy MMT AO system. Reproduced from Hinz et al. 2018<cit.>. The performance of legacy MMT AO was limited with residual error of 450 nm due to a low number 55 of corrected modes and a slow temporal response 2.7 ms delay. The new design will correct 220 modes with a faster temporal response of 1 ms delay. These efforts will push the residual wavefront error down to <200 nm, i.e., >70% and 60% Strehl ratios at K and H-bands.	2012.11668	Expected_performance_of_MAPS.jpg
0.430401	a9257992c77245438c154badc52c51ab	(a) A laboratory setup to characterize the CCID75 and tip-tilt modulation mirror. (b) A E-727 PI controller is used to control the tip-tilt modulation mirror. (c) The camera link cables are extended several meters using a noiseless VisionLink camera-fiber-link extender system using LC/LC duplex multimode fiber optic patch cables between the CCID75 camera (will be inside the telescope) and the data acquisition computer (will be placed in the computer room) for the data acquisition.	2012.11668	Lab_setup.jpg
0.475577	b03a722fe9534d55b168e724fc88eb43	MAPS on the MMT 6.5m - showing ASM, wavefront sensors unit (W-unit) and science cameras: ARIES and MMTPol. The pyramid wavefront sensor images are read to computer using a camera link. The computed slopes are sent to AO Reconstructor Computer (). The computes the ASM position control commands from the slopes matrix and sends them to the ASM central hub for making corrections. Next ARIES and MMTPol records AO corrected science observations.	2012.11668	MAPS_on_telescope.jpg
0.47741	e209ca900d8743199a651067a22d12fa	A circular modulation with 3000μrad at 1000 Hz is created with the tip-tilt modulation stage by applying 90-degree phase-shift between two axes moving with sine wave.	2012.11668	Modulation_PV.jpg

0.46919

ef6411e3272d4d46b2cf9d6e0623c277

Exp1: Single Distributional Constraint. Balancing demographics can be represented easily through distributional constraints. By using a constraint such as _x ∼ pϕ_female(x) = 0.5, we can target balancing the female biographies in the distribution of all generations. Note that a point-wise objective _x ∼ pϕ_female(x) = 1.0 would maximize the presence of female biographies at the expense of other demographics, inducing bias in the opposite direction. The plot shows how _x ∼ pϕ_female(x) evolves towards the defined expectation: GDC is able to reduce the bias of GPT-2bio to obtain 36.7% female biographies rather than just 7%.

2012.11635

exp_401_distributional.jpg

0.423607

38c8f7b4269e4f2ea95fd28fcca8c5e6

Exp2: Multiple Distributional Constraints This experiment demonstrates the flexibility of in dealing with several distributional constraints at once, even when these constraints have different objectives (increase, decrease, or keep fixed). We challenge the flexibility of by setting four distributional constraints with four arbitrary expectation values targeting ϕ_science and ϕ_art at 40% and ϕ_sports and ϕ_business at 10%. In the figure, from left to right, we can note the increase of ϕ_science and ϕ_art from 1.5% to 20.3% and from 10% to 31.6% respectively. Interestingly, the initial ϕ_business of GPT-2bio (10.9%) is already very close to the desired expectation (10%), and we can see that during the course of the training, keeps this value fixed as it is already satisfying the corresponding target distributional constraint. ϕ_sports initially starts higher than the target distributional constraint 10%, and we can note that succeeds to reduce it from 19.6% to 11.9%.

2012.11635

exp_402_distributional.jpg

0.433375

5841e529bb6945c5ab2d763eccf09e09

Exp3: Hybrid constraints In this experiment, we specify two types of constraints: pointwise with ϕ_art(x) = 1.0 and distributional with ϕ_female(x) = 0.5 (henceforth Hybrid). in a single training procedure is able to increase the expectation of biographies about females from 7.4% to 36.6% and Art professions from 11.4% to 88.6%.

2012.11635

exp_410_distributional.jpg

0.429033

b4f659f64dab4f2b897749b2dc6079a6

Exp4: Hybrid constraints. In this experiment, we specify two types of constraints: pointwise with ϕ_sports(x) = 1.0 and distributional with ϕ_female(x) = 0.5. in a single training procedure is able to increase the expectation of biographies about females from 7.4% to 31.9% and Sports professions from 17.5% to 92.9%.

2012.11635

exp_411_distributional.jpg

0.409264

ee5379dcf0e543f18bf9fb48c15b0956

Exp5: Hybrid constraints. In this experiment, we specify two types of constraints: pointwise with ϕ_business(x) = 1.0 and distributional with ϕ_female(x) = 0.5. in a single training procedure is able to increase the expectation of biographies about females from 7.4% to 37.7% and Business professions from 10.1% to 82.4%.

2012.11635

exp_412_distributional.jpg

0.380032

031704969aed4806adae9f639214bf99

Exp6: Hybrid constraints. In this experiment, we specify two types of constraints: pointwise with ϕ_science(x) = 1.0 and distributional with ϕ_female(x) = 0.5. in a single training procedure is able to increase the expectation of biographies about females from 7.4% to 28.8% and Science professions from 1.2% to 74.7%.

2012.11635

exp_413_distributional.jpg

0.51394

23abc302b7ca495b92a436e51e979139

Case of a pointwise binary requirement r(x)=1: comparison with Reinforce and Ziegler. The curves correspond to different (·,a) levels. The manifold 𝒞 is the set of distributions c s.t. c(x) > 0 → r(x)=1, or, equivalently s.t. _x∼ c r(x) = 1. The curved lines represent increasing levels of the KL divergence (q,a). According to Reinforce, any distribution p_R s.t. _x∼ p_R r(x) =1, that is, any distribution on 𝒞, is optimal. According to Ziegler, to each temperature β > 0 is associated an optimal distribution p_Z = _q β(q,a) - _x∼ q r(x), which does not lie on 𝒞. Our own optimal p is the distribution of minimal that lies on 𝒞.

2012.11635

p_and_reinforce_and-ziegler_2.jpg

0.497317

1f6ab3d042fe4c5b8bb8fd3aec58661b

“Zipf-like” token frequency analysis on sets of 68000 generated samples from each method (only samples strictly satisfying the constraints are kept, for fair comparison). Longer tails mean a lower concentration of mass on the high frequency tokens, and therefore indicate more vocabulary richness. See Appendix <ref> for details.

2012.11635

zipf_main.jpg

0.54272

d98132bde1fe45d59c8b0ea96e1083da

Left-hand cut singularities (solid black curves) in the complex s-plane for the ππ→ππ (a) and π K →π K (b) scattering. In the plot we schematically show the position of the closest left-hand cut singularity (s_L), Adler zero (s_A), threshold (s_th) and the expansion point (s_E). Dashed lines determine the specific form of the conformal map and subsequently the domain of convergence of the conformal expansion in Eq. (<ref>).

2012.11636

1.jpg

0.419986

5086b9e9ec464415a7d6072ea1fa3c93

Results for the ππ→ππ scattering with J=0, I=0 in the single-channel case. Top, central and bottom panels correspond to m_π=physical, 236, 391 MeV, respectively. Left panels show the convergence of the conformal expansion in Eq. (<ref>), central panels show the comparison with the data and right panels show the corresponding Omnès functions. In the phase shift plot for the physical pion mass two curves are shown: fit to the experimental data <cit.> (dashed curve) and fit to the pseudo data from Roy analysis <cit.> (thick curve).

2012.11636

2.jpg

0.466998

4c0f2e42a7be40189818ec56fc9228d2

Comparison between the coupled-channel N/D fits and the Roy-like solution from <cit.>. The dashed curves are the fit solely to the experimental data, while the solid curves take advantage of both the experimental data and the results of Roy (Roy-Steiner) analyses on ππ→ππ ( ππ→ KK̅).

2012.11636

3.jpg

0.432529

a83dcdd1706949e0b4dadee0633714a8

Results for the ππ→ππ, KK̅ scattering with J=0, I=0 in the coupled-channel case. Top, central and bottom panels correspond to 11, 12 and 22 matrix elements, respectively, with 1=ππ and 2=KK̅. Left panels show the convergence of the conformal expansion in Eq. (<ref>), central panels show the comparison with the data, right panels show the elements of the Omnès matrix. In the central plots two curves are shown: fit to the experimental data <cit.> (dashed curve) and fit to the pseudo data from Roy analyses <cit.> (thick curve).

2012.11636

4.jpg

0.426777

4e886d3695e841ffa8970d477d802e44

Results for the π K →π K scattering with J=0, I=1/2 in the single-channel approximation. Top and bottom panels correspond to m_π=physical, 239 MeV, respectively. Left panels show the convergence of the conformal expansion in Eq. (<ref>), central panels show the comparison with the data and right panels show the corresponding Omnès functions. In the phase shift plot for the physical pion mass two curves are shown: fit to the experimental data <cit.> (dashed curve) and fit to the pseudo-data from Roy-Steiner analysis <cit.> (thick curve).

2012.11636

7.jpg

0.428286

e61f477644604949b7a60ce2b0addf23

E(Q_ω) as a function of M(Q_ω) for cubic-quintic ground states in dimension d=3.

2012.11637

NLS35_3d_solmEzoom-eps-converted-to.jpg

0.467461

cfbd33b8c366416e85bc6e42d9e6bf36

Solution to (<ref>) in dimension d=3 for initial data (<ref>) with ω=0.01 and λ=1.001.

2012.11637

NLS35_3d_solom001_1001water-eps-converted-to.jpg

0.449527

05b379e0358545d8be5dc8cafec16954

In-distribution anomaly detection through conditional density estimation. Consider samples of a 1D feature x and a conditional parameter of interest M (left panel), drawn from a smooth Gaussian `background' with a small number of anomalous `signal' events added (inside red circle for clarity). The conditional density values at each data point do not allow the anomaly to be distinguished from the background (center left panel), as they only identify the outliers in low density regions. However, the local over-density anomaly ratio α peaks at the anomalous data points (center right panel), and implementing a minimum cut on the anomaly ratio reveals the anomalous events (right panel).

2012.11638

method_illustration.jpg

0.513366

c19e740f50c244218ccfc61712be55e9

Hyperfine interactions of magnetic adatoms on ultrathin films: (a) Fermi contact and (b) dipolar contributions. The free-atom values are given for comparison. The number to the left in the legend indicates the number of layers in the film. The adsorption position is denoted with T for atop the anion and with B for the bridge position (for h–BN and Cu_2N only T was considered). The possible impact of corrections from van der Waals interactions was explored for the h–BN case.

2012.11639

fig1.jpg

0.389285

456161fd46b841c9aa238d245e99a44a

Basic properties of isolated 3d transition metal atoms. (a) Spin magnetic moment, (b) spin-polarized energy levels and (c) Fermi contact contribution to the hyperfine interaction of isolated 3d transition metal atoms.

2012.11639

fig2.jpg

0.471437

c5e281636874450eb017ae935fb92de1

Geometry of adatoms on MgO, NaF and NaCl. Top and side views of (a) adatom stacked on top of anion and (b) adatom in the bridge position. The adatom is represented by a blue sphere, the anion by a red sphere and the cation by an orange sphere. The definition of the distance d between the adatom and anion and the cation-anion-cation bond angle α is also illustrated.

2012.11639

fig3.jpg

0.451744

1a5035a3af6d42fc928118628c857b1f

Magnetic properties of adatoms on MgO ultra-thin films, placed on top of oxygen (a–c) and on the bridge position (d–f). (a,d) Spin magnetic moment, (b,e) Fermi contact and (c,f) dipolar contributions to the hyperfine interaction. The number to the left of MgO in the legend indicates the number of layers in the ultra-thin film. The gray dash line represent the free-atom results. The black crosses in panels (b,c) and (e,f) refers to the experimental results for Fe and Ti from Ref. and for Cu from Ref. .

2012.11639

fig4.jpg

0.456675

8a2f8a45fcda4a2b9388858813644c0e

Thickness dependence of the electronic structure of a V adatom on MgO in the oxygen-top position. (a) One monolayer of MgO. (b) Two monolayers of MgO. (c) Free atom.

2012.11639

fig5.jpg

0.443271

66260eac924542838b84077a2b3e4303

Magnetic properties of adatoms on NaF and NaCl ultra-thin films, placed on top of anion (a–c) and on the bridge position (d–f). (a,d) Spin magnetic moment, (b,e) Fermi contact and (c,f) dipolar contributions to the hyperfine interaction. The number to the left of NaF and NaCl in the legend indicates the number of layers in the ultra-thin film. The gray dash line represent the free-atom results.

2012.11639

fig6.jpg

0.450604

3a5986f0ad0642e2837cc851220c17dd

Magnetic properties of top-stacked adatoms on Cu_2N and h–BN. (a) Spin magnetic moment, (b) Fermi contact and (c) dipolar contributions to the hyperfine interaction. The gray dash line represent the free-atom results.

2012.11639

fig7.jpg

0.483159

720f7432b9254219aeae0b119c952282

The BoB system deployed during field tests in a freshwater reservoir in Singapore.

2012.11641

Fig-1.jpg

0.513874

581f2feffb924e87b27685b319aaaf40

Qualitative performance for the area coverage

2012.11641

Fig-10.jpg

0.547466

583f9805fc1e4d73962f004c7cb1718a

Average agent reward for each of the 9 agents constituting the multi-agent system

2012.11641

Fig-11.jpg

0.487314

40519372276041bcbf3428439cdaa0a3

Qualitative performance for the area coverage

2012.11641

Fig-12.jpg

0.455493

73e16f90985e43029654bd32e5bf72a9

Average agent reward for each of the 9 agents constituting the multi-agent system

2012.11641

Fig-14.jpg

0.487911

2df8ab5fefdf4c63842ce993f052690c

Average agent reward

2012.11641

Fig-5.jpg

0.503244

47c805ecadb14eed966993cdf38d3f8b

Qualitative performance for the area coverage

2012.11641

Fig-6.jpg

0.484186

bff7b97cf0384cd4a1d69dfbfe35c09b

Average agent reward

2012.11641

Fig-7.jpg

0.545885

5a6f74e472b24165922cdee810f8c87e

Qualitative performance for the area coverage

2012.11641

Fig-8.jpg

0.49491

b569f06728294ec5b7bab5856f8e60ea

Average agent reward for each of the 9 agents constituting the multi-agent system

2012.11641

Fig-9.jpg

0.523185

a3f29be3b4c84255a5311ef812425955

Evaluation protocol for direct comparison of the neural networks performance in the specific environment. The user can compare the performance of own neural network with the state of the art methods.

2012.11643

results.jpg

0.496147

1299c1978e1046e7aac515ecf9b70666

Examples of robotic arms and exchangeable grippers included in . The user can easily add any robot to the environment just by specifying its name and optionally position, orientation and initial configuration.

2012.11643

robots_selection.jpg

0.481041

710a7ba017a345a29cafa84b5ab99529

Examples of robotic workspaces. Top: Pepper robot by a desk (left) and in a maze (right). Middle: A vertical plane workspace (left), and a collaborative workspace with two KUKA arms and a humanoid (right). Bottom: Example of image augmentation during training, where random textures from a database are applied to the objects (left). A pretrained neural network Yolact can be used to preprocess the images (right).

2012.11643

workspaces_crow.jpg

0.367039

7c25a42abe1f450d924f5a7eb147ee37

C_V(ϵ/β_H) vs ϵ/β_H at T=0.

2012.11644

img1.jpg

0.368181

241aed3fe7d84af9a9bd2746421c05c1

C_𝒜(ϵ/β_H) vs ϵ/β_H at T=0.

2012.11644

img2.jpg

0.514217

613da37e8c7e46238913dd0ba3dd9bcb

Comparison between C_V and C_𝒜 at zero temperature. For large ϵ/β_H, the action complexity decays much faster than volume complexity.

2012.11644

img3.jpg

0.399931

9ce0c8fe6e974c148e8e68f82a72711b

C_𝒜(ϵ/β_H) vs ϵ/β_H at finite temperature.

2012.11644

img4.jpg

0.374115

39d0877f09854a549bea1d3a93304a86

Summary of the batch speed assessment. Absolute error values for each variance component estimated using and summarized as a boxplot.

2012.11646

error_plots.jpg

0.55313

d3ac2e531a4a4338a86f424aefdc8334

Regret averaged across all users for each week in the simulated mHealth trial associated with approaches , and .

2012.11646

seb_sim.jpg

0.464552

6a3a317e31e74e00a0ec5da3e2a565ec

A full-duplex device i transmitting to j as it receives from k in-band.

2012.11647

fdx_mmwave_system_type_13_tx_rx_v2.jpg

0.412389

79b9ac2c64254c309a7b441d7ef97a55

Sum spectral efficiency as a function of snr for various = + 20 dB, where κ = 10 dB, = 12 bits, and K_ij = K_ki = 3. While not obvious from this figure, is fairly robust to self-interference with = 12 bits. Stricter choices of = + 20 dB degrade and, thus, the sum +.

2012.11647

results-11-7a_v2_channel-avg_SNR_ij-10-dB_G_ii--55-dB_kappa-10-dB.jpg

0.422442

d6ced733bddb40348d27caa875c2ce0f

Sum spectral efficiency as a function of for various , where κ = 10 dB, K_ij = K_ki = 1, _ij = _ki = -10 dB, and b = 12 bits. As and are relaxed, the system can achieve a greater sum spectral efficiency since transmit performance is less constrained by a limited receive dynamic range.

2012.11647

results-14-2a_channel-avg_kappa-10-dB.jpg

0.420266

ca97f659534f4776ab87f00152edb9ab

Sum spectral efficiency as a function of κ for various and , where K_ij = K_ki = 1, _ij = _ki = -10 dB, and b = 12 bits.

2012.11647

results-14-4a_channel-avg.jpg

0.434869

f48032db06bc4cb680a0056484b2476d

Difference Between the Best ≤4-Average Mixture and the Simple Average Mixture, Eurozone Real Interest Rate

2012.11649

fig_heatmap_diff_subsetleN_avg_uni_real.jpg

0.483311

e445eb5215b34c8bb5a677ffbe68c468

PIT Histograms, Eurozone Inflation

2012.11649

fig_pit_uniform_4panel_v2.jpg

0.460178

d44d76c5fcb04f28ba2b69a4b73feb7d

The concentration u obtained as a PINN solution of ADE (<ref>) at (a) t=2 min, (b) t=6 min, and (c) t=20 min. The absolute point errors with respect to the STOMP solution are shown in subfigures (d–f). The DNN size is 5 × 60.

2012.11658

Plot_2Dt_CAD_K05.jpg

0.433026

260d5f6dd4e44b159679f17b21b54f5b

Same as in Figure <ref>, but with Pe=200. For t = 0.5, the maximum point errors for w=1, 10, and 100 are 2.21 × 10^-3, 1.14 × 10^-3, and 1.07 × 10^-3, respectively.

2012.11658

Plot_2Dt_Pts_Pe200.jpg

0.403868

dbc9a1053c1e4566b5eb8a4c3db32e9e

Snapshots of the PINN solution of the time-dependent ADE (<ref>) with Pe = 50 (left collumns) and the corresponding point errors with respect to the analytical solution (<ref>) (right columns) obtained with: (a) w=w_IC = w_BC = 1, (b) w=w_IC = w_BC = 10, and (c) w=w_IC = w_BC = 100. For t = 0.5, the maximum point errors for w=1, 10, and 100 are 2.96 × 10^-3, 1.33 × 10^-3, and 0.97 × 10^-3, respectively.

2012.11658

Plot_2Dt_Pts_Pe50.jpg

0.478735

74821f9891b34de281be11f1a060dc00

Upper row: the PINN solution of the backward ADE (<ref>) at t = 0 with (a) Pe = 1, (b) Pe = 10, and (c) Pe = 200. Lower row: the absolute point errors with respect to the reference solutions. The maximum point errors for these three cases are 8.07 × 10^-3, 1.23 × 10^-2, and 1.40 × 10^-2, respectively.

2012.11658

Plot_2Dt_Pts_back.jpg

0.465322

ab86ded7fd7c4549bab8350a499fc91d

The reference concentration u obtained as a STOMP solution of the ADE (<ref>) at different times.

2012.11658

Plot_2Dt_bCAD_0_10min_C_K05.jpg

0.396537

dbca953692b7467997306345ff39431b

The absolute errors in the PINN solution û(x⃗,t,θ) with respect to the reference field u(x⃗,t) at different times. The DNN size is 5 × 60.

2012.11658

Plot_2Dt_bCAD_0_10min_err_K05.jpg

0.450069

614dea558334424691b0a80ee37a9727

Absolute error in the PINN solution of the backward ADE with data assimilation. The absolute errors computed with respect to the reference u(x⃗,t) field. The DNN size is 5 × 60.

2012.11658

data_assimilation.jpg

0.418507

3feb035a781044c5addc49f7ca0b1eb3

The comparison of the PINN and finite elements solutions for the ADE with the boundary conditions and velocity vector shown in Figure <ref> for Pe=1 and 100. The PINN solution is obtained with N_f = 41 × 41 residual points, the DNN size 3 × 30 and w_BC = 50. The GFEM and SUPG solutions in subfigures a, b, d and e are reproduced from <cit.>.

2012.11658

fig_advdis_skew_updated.jpg

0.479669

ebd513b100534a7aa549c82d38d5fb7d

Boundary conditions for steady-state ADE with the point source located at (0,0). The velocity vector is v⃗=[cos(π/4), sin(π/4)]^T.

2012.11658

sche_ad_skew.jpg

0.479527

4a55699553b84ff081d90739dacc24fb

Pulse schedules corresponding to the (a) scaled circuit and (b) basis circuit, corresponding to gates given in Fig. <ref>, with the end of the scaled schedule marked by the vertical gray dashed line. Here, the `d' and `u' correspond to the (on-resonant) s and (off-resonant) s of Qiskit Pulse, respectively. The pulses are shaded according to phase, with light pulses in-phase and dark pulses in the quadrature phase. The circular arrows correspond to Z-rotations executed in software by changing the phase of subsequent pulses <cit.>. The duration is given in units of sample time, dt = 0.222 ns.

2012.11660

fig-sched-annotate.jpg

0.43305

c9796e5fdeca4029aa5825e89340fb19

Gate sets that implement a single Suzuki-Trotter step t_i of the time evolution. (a) Implementation with 2-qubit basis gates; each set of gates is labeled by the corresponding Hamiltonian terms above. (b) Implementation with analog scaled 2-qubit ECR gates A(θ), where ZX(θ) = R_x(π/2) A(θ) R_x(π/2), as shown in Fig. <ref>c.

2012.11660

fig2-annotate.jpg

0.419645

3951d9d4b18e4e1fb62bbcc6f752b91e

Quantum process tomography of U_yx(θ) as a function of θ on ibmq_athens. Both the double-CNOT circuit (a) and the parametric ZX(θ)-based circuit (b) create U_yx(θ). The circuit in (b) at θ=π/2 is implemented by the pulse schedule shown in (c). Here, each sample lasts 0.222 ns. (d) Fidelity F_(a) (blue up-triangles) and F_(b) (orange down triangles) of quantum circuits (a) and (b), respectively, obtained with QPT. (e) Relative error reduction of the pulse-efficient circuit in (b) over the circuit in (a), i.e. 1-E_(b)/E_(a) where E_(x)=1-F_(x).

2012.11660

majorana_qpt.jpg

0.499569

501351d4728344609fb66f3bf31a780b

Scaling in different dimensions

2012.11662

Fractaldimensionexamplebw.jpg

0.437598

a14638ece943413fbba81a8918092292

Calculating the mesh dimension

2012.11662

Snow_LineFitAnn.jpg

0.511456

ea066e80bfaa44368531f847acdf6dee

A non uniform mesh

2012.11662

Snow_Mesh_Example.jpg

0.387671

84ddf84d868b4f6a84279078a266ee5e

HalfCheetah

2012.11662

cheetah.jpg

0.389821

e3f8e533311e463bac493a4a9dab7f62

HalfCheetah

2012.11662

cheetah.jpg

0.501832

c6d5b4cdb9734267a3e4145a0b1a0da9

Top: mesh sizes vs log of the box size for the cheetah environment. Lower Left: Every five frames overlaid for the an identity policy on the cheetah. Lower Right: Every five frames of cheetah after the lower mesh dimension training.

2012.11662

cheetah_mesh_cor.jpg

0.43681

5da29009ae584e6a9050bf078db97e17

Hopper

2012.11662

hopper.jpg

0.44734

6267097f46794e8e8be12c1cb09b0774

Hopper

2012.11662

hopper.jpg

0.41274

aa8a5974eb784681847510928422f490

High resolution curve of a well behaved policy

2012.11662

meshdim.jpg

0.393872

2b9bb940a23942e88d3e90d11298fc99

Run time resolution curve of a typical policy

2012.11662

meshdimreal.jpg

0.401837

beb639c1b95e4d138378d4b615122934

Walker

2012.11662

walker.jpg

0.432241

c1a8427a76b04abbb65f37255068e043

Walker

2012.11662

walker.jpg

0.53746

1ce5aa2d0c594779919f2a4068a5c1cd

Torque exerted during the sampled episode

2012.11663

act_hist.jpg

0.483385

09e45fa90ccb4fadaaa9e9bf0cc23f59

Observations during the sampled episode

2012.11663

obs_hist.jpg

0.436918

28f6a3edb70e4036b9a70da592f378aa

Reward curve for SSAC and the other algorithms we compare to. the solid line is the smoothed average of episode reward, averaged over four random seeds. The shaded area indicates the best and worst rewards at each epoch across the four seeds. SSAC is shown starting later to account for the time training the gating function alone.

2012.11663

reward.jpg

0.479445

3dcee0076f55452593987a78a7d1ab73

System diagram for the new technique proposed in this paper. Rounded boxes represent learned neural networks, squared boxes represent static, hand crafted functions. The local controller is a hand designed LQR, the swing-up controller is obtained via reinforcement learning, and the gating function is trained as a neural network classifier

2012.11663

system5.jpg

0.390179

1c51473921a144258adefad4fe2899c2

Balance map for TD3, X and Y indicate the initial position for the trial, a black dot indicates that the trial started from that point satisfies equation (<ref>), and red indicates the converse. when these initial conditions were run for SSAC, it satisfied (11) for every initial condition

2012.11663

th_map_td3.jpg

0.410123

212d0a920e9a436ca55b28c3ed4dd730

Attenuation factor of the emitted radiation by γγ-annihilation. Internal absorption is caused by interaction with the photons created in the acceleration region by the nonthermal radiative processes. The external stellar radiation field provides the target photons for stellar absorption.

2012.11664

attenuation.jpg

0.479103

04d649e6ff88425b902732301007e9c6

Rates of acceleration and radiative losses for electrons.

2012.11664

electron_cooling.jpg

0.405023

efa3bccc8b5c4890a95302bf975a236a

Distribution of relativistic electrons in the stationary one-zone model for a highly efficient acceleration.

2012.11664

electron_distribution_1.jpg

0.411754

79f14791e7244112ac0f094b71c63aab

Distribution of relativistic electrons in the stationary one-zone model for a low value of acceleration efficiency.

2012.11664

electron_distribution_2.jpg

0.373131

8b972c93fec84858a3784bfa97bf9ce2

Rates of acceleration and radiative losses for protons.

2012.11664

proton_cooling.jpg

0.429701

518561fc2fbf4f72878bf8d574e40601

Scheme of the interaction between the equatorial wind and the donor star. The wind stagnates at the point where ρ_ w v_ w^2 = P_ rad. The distances from the stagnation point to the black hole and the center of the star are indicated as r_ BH and r_ S, respectively.

2012.11664

sketch2.jpg

0.570172

35b744d5df7d4320ab113ee8d87d5f46

Spectral energy distribution of the nonthermal radiation emitted in the acceleration region. Nonthermal radiative processes of the secondary pairs are considered for the calculation of the total spectrum.

2012.11664

total_SED.jpg

0.47117

de638ddcff534998be006c5e35dd0f6f

Scheme of the considered model of accretion disk and winds in super-Eddington microquasars. Toroidal magnetic fields are expected in the disk. As long as there is no poloidal magnetic field, there is no launching of jets. The figure corresponds to the parameters adopted in Sotomayor Checa & Romero (2019).

2012.11664

wind_disk.jpg

0.392493

e6c8f1c4cc374519b8c233854bb5bae8

Chavissa et al. 2020<cit.> show MIRC-X aperture synthesis imaging of surface of CL Lac, AGB star. The MiRA reconstructed images at 1.52 μm (left), 1.70 μm (middle) and difference between the two (right). These images show presence of a brighter area that largely affects the position of the photocenter and that likely caused Gaia DR2 astrometric error. Reproduced from Chavissa et al. 2020<cit.>.

2012.11666

Chavissa_AGB_image.jpg

0.421479

0d75d3c1663b443eb656cfe3d1d10991

Kraus et al. 2020<cit.> observed the first disk tearing in a protoplanetary hierarchical triple, GW Ori. (A & B) ALMA images with 20mas resolution showing disk and misaligned rings. The MIRC-X measured inner triple stellar orbit allowed to link the detected misaligned ring to the disk tearing effect. (C) The MIRC-X reconstructed images of inner triple over-plotted with the best-fitting orbital model. The white dot marks the center-of-mass of the system. Figures reproduced from Kraus et al. 2020<cit.>.

2012.11666

GW_Ori.jpg

0.430746

6e27891f84534200961a5b13cf2b9ff9

Gardner et al. 2020<cit.> show discovery and measured orbital elements of a 30-day companion to the B-type binary α Del. (Left) The outer binary orbit of α Del as the Ba component moves relative to the A component fixed at the origin. (Right) Detection of an added astrometric wobble due to the presence of the Bb component. Reproduced from Gardner et al. 2020<cit.>.

2012.11666

alpha_Del.jpg

0.492836

68d38dc8c30d48f6807fadc6663b1922

The Tel-AO software layout. The camera link server reads the frames from the WFS Andor EMCCD camera through BitFlow Neon frame grabber and writes them to an instrument shared memory in a circular buffer so that the main WFS main server uses them. The WFS main server reads the frames from the instrument shared memory and calculates the wavefront slopes from each new science frame. Next, the DM commands are computed by multiplying the slopes matrix with the slopes-to-actuator reconstructor matrix. The tip/tilt is computed by taking the average of slopes and then sent to telescope M2 for correction with appropriate gain. The rest higher-order corrections are sent to the ALPAO deformable mirror. The wavefront sensor images and deformable mirror shapes are sent to the client GUI for monitoring and operation.

2012.11667

TELAO-software.jpg

0.452997

e85024670b8a41e8be15aaab7b7b14e8

Photographs of the Tel-AO (left) and Lab-AO systems (right) as built.

2012.11667

Tel-AO_and_Lab-AO.jpg

0.417485

2f1aaf2c6e0a442da4bb9986fab79eb0

USB data data acquisition of frames (black color) in a comparison to the camera-link data acquisition (blue color). The integration is 3ms but the data points are the time difference between two consecutive frame grabs. The USB acquisition has a lot of jitters, in a worst-case, a grab gets 8 frames in a chunk (i.e., 8 frames of delay). In the case of camera-link, there is also jitter but very small.

2012.11667

USB_missing_frames.jpg

0.417306

58dabbafc68645d5a2717aa7b4442e72

A photograph of the main observing GUI at the CHARA Array, . Each controls one telescope and has all the controls for the telescope and its acquisition camera, tip-tilt, Tel-AO, and Lab-AO, including the real-time displays of the acquisition camera and wavefront sensors of Tel-AO and Lab-AO. This GUI is optimized for screen space – all six GUIs fit in a single screen monitor, time-efficient usage, and stability by keeping only the required observing controls together. This photograph is taken with telescope beacon was ON.

2012.11667

obsgtk_6T.jpg

0.50631

a951c735c6af4126aa9e1e244339b58d

Rejection transfer function measured using the telescope beacon as the source of light. The green dotted line is the measured closed-loop transfer function. The dashed line is the modeled open-loop transfer function. The thick black line is the modeled closed-loop transfer function. This model uses the total delay time T_ D is used. The green line cross the measured -3dB closed-loop bandwidth at frequency of 19Hz.

2012.11667

telao_model.jpg

0.422094

2aa2ac4c56404669b13fbf403b758d32

(Left) The architecture of CCID75 reproduced from Schuette et al. (2014)<cit.>. It has 160 (Horizontal) × 160 (Vertical) imaging area – subdivided into twenty 16 (H) × 80 (V) pixel channels. Each channel has a separate readout chain (serial readout register and output amplifier). (Right) Quantum efficiency is around 80% for 0.6-0.9μm wavelengths.

2012.11668

CCID75_architecture.jpg

0.445979

aaffc375f5ae40bd9d6da9941d19fda8

Computer architecture of MAPS AO system. The frame grabbers of the VPyWFS-CCID75 and the IRPyWFS-SAPHIRA and the RJ45 Ethernet of the acquisition camera are installed on the server computer. The INDI instrument drivers, , and , which are installed on computer, read frames from all the cameras. The and compute slopes from the detector images and then send to over a dedicated Ethernet socket in a “WFS loop" operated at the WFS frame rate, 1kHz. driver works on the computer runs a second loop “ASM loopâ independent of the WFS loop and converts the WFS slopes to the ASM commands and send to the ASM actuators to apply correction. The allow engineering with actuators and the does house keeping and safety.

2012.11668

Computer_architecture1.jpg

0.409986

c6b10bdac4424c5f8d42f3e52b983fa4

The computer handles non-real-time Input and Output devices such as motors and actuators, web Graphical User Interfaces (GUI) and telemetry storage. For example, is web socket server. The is an INDI driver for controlling motors. The is an INDI driver allows to play with ASM administrative states.

2012.11668

Computer_architecture2.jpg

0.422416

ed1f7d6e65a24eeaa067d70281c82580

Expected performance of MAPS compared to legacy MMT AO system. Reproduced from Hinz et al. 2018<cit.>. The performance of legacy MMT AO was limited with residual error of 450 nm due to a low number 55 of corrected modes and a slow temporal response 2.7 ms delay. The new design will correct 220 modes with a faster temporal response of 1 ms delay. These efforts will push the residual wavefront error down to <200 nm, i.e., >70% and 60% Strehl ratios at K and H-bands.

2012.11668

Expected_performance_of_MAPS.jpg

0.430401

a9257992c77245438c154badc52c51ab

(a) A laboratory setup to characterize the CCID75 and tip-tilt modulation mirror. (b) A E-727 PI controller is used to control the tip-tilt modulation mirror. (c) The camera link cables are extended several meters using a noiseless VisionLink camera-fiber-link extender system using LC/LC duplex multimode fiber optic patch cables between the CCID75 camera (will be inside the telescope) and the data acquisition computer (will be placed in the computer room) for the data acquisition.

2012.11668

Lab_setup.jpg

0.475577

b03a722fe9534d55b168e724fc88eb43

MAPS on the MMT 6.5m - showing ASM, wavefront sensors unit (W-unit) and science cameras: ARIES and MMTPol. The pyramid wavefront sensor images are read to computer using a camera link. The computed slopes are sent to AO Reconstructor Computer (). The computes the ASM position control commands from the slopes matrix and sends them to the ASM central hub for making corrections. Next ARIES and MMTPol records AO corrected science observations.

2012.11668

MAPS_on_telescope.jpg

0.47741

e209ca900d8743199a651067a22d12fa

A circular modulation with 3000μrad at 1000 Hz is created with the tip-tilt modulation stage by applying 90-degree phase-shift between two axes moving with sine wave.

2012.11668

Modulation_PV.jpg