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
0.600779	64d8432aa3274ee49e5f2003a96879f5	EM	2012.15492	after_em.jpg
0.419817	73749b9839634ded88fac412f8c25d8b	k=10	2012.15492	circles_1.jpg
0.414179	6d45b7a77fed4094bc7235c85c4b6348	k=20	2012.15492	circles_2.jpg
0.454782	3d8e75e7f1b2482db5823da6dd743084	k=30	2012.15492	circles_3.jpg
0.566318	20e76f7d10774dc3ab4f2f31fc8a007e	Correction	2012.15492	corrected.jpg
0.444729	5e828b9b6264466191613bb015541a8c	k=3	2012.15492	moons_1.jpg
0.439095	de0d44c3961e4fab98cdf88fe5ddb26b	k=10	2012.15492	moons_2.jpg
0.453261	7aecd2162c594cbfb3fb88d5e9f46f41	k=17	2012.15492	moons_3.jpg
0.429269	74452efea5da47038194d6bb85fe036c	The Super-k algorithm processing steps	2012.15492	processing_steps.jpg
0.431959	69c95a4d386e4c85bc4b324c5b08461b	k=2	2012.15492	random_1.jpg
0.467467	79317e89f31d4799988eea57cdb0e8cb	k=5	2012.15492	random_2.jpg
0.478955	fafb4bb5201e460ca39b2ddcdf914b65	k=8	2012.15492	random_3.jpg
0.471313	e66cfa3581654bb690bb762a1f86d0a2	Class 0	2012.15492	voxelization_0.jpg
0.472111	314fe599a92c4548a73aa05fca559b29	Class 1	2012.15492	voxelization_1.jpg
0.586297	12c1fb44e0fe4780a29e7ad30eb423e7	Voxelization	2012.15492	voxelized.jpg
0.473316	d661b2bf822a4d3b97637345412a1635	Schematic Diagram of our Zero-Shot KD solution. a) We assume access to a pre-trained teacher. b) We adversarially train the generator using out-of-domain data (OOD). c) Finally we use the generated data and the OOD data for KD.	2012.15495	fig_zskd.jpg
0.516594	43bbb7a65dab4d01ad8e021c9cdf0a69	Two cases for quenches in quantum phase transition. (a) E_0≪Λ. Instantaneous quench (IQ) regime, fast quench (FQ) regime, and Kibble-Zurek (KZ) regime. t_KZ∝ E_0^-1(E_0δ)^1/(zν r). (b) E_0>Λ. Instantaneous quench (IQ) regime, non-universal (NU) regime, and Kibble-Zurek (KZ) regime.	2012.15501	figure1.jpg
0.471599	e34159590b124f02aafb3d2c571838b3	The first and second cumulants of the work (absolute value) as a function of the quenching rate for a scalar field with time-dependent mass. The parameters are: E_0=m=0.1, Λ=10. (a) and (c): t_f=0. (b) and (d): t_f→∞. The solid lines denote (κ_n)_re in the Kibble-Zurek regime and fast-quench regime and (κ_n)_su-κ_n in the instantaneous quench regime. The dashed lines denote the fitting curves.	2012.15501	g2.jpg
0.494133	790d78c87fe24872a66dfeaf134da933	The competition between the frozen time t_KZ and the quench time scale δ in the adiabatic-impulse-adiabatic stages. The black solid line denotes the energy gap E_g. (a) the Kibble-Zurek regime. A part of the protocol is inside the impulse stage (t_KZ<δ). (b) the fast quench regime. All the protocol is inside the impulse stage (t_KZ>δ).	2012.15501	kzm-cft.jpg
0.42115	d283b0f8518644b4a42462eeb1a7a114	High-level representation of the AdapterCL.	2012.15504	ADPTCL.jpg
0.471628	d1acf272a4944840b63b280bbdc302a0	In Continual Learning, the model is trained one dataset at the time. In this instance, the model is first trained on data from the hotel domain D_Hotel then on the Taxi D_Taxi and so on. The parameter of the model is updated sequentially based on the loss function L.	2012.15504	CLVANILLA.jpg
0.601604	1c0b48c69cba48c38aad1f4c50e58072	Ablation study on the size of the episodic-memory vs JGA.	2012.15504	DSTMem.jpg
0.527535	400eef6903454528bf2f3c86b5c8593d	Ablation study on the size of the episodic-memory vs the intent accuracy.	2012.15504	IntentMem.jpg
0.427626	733be8c1a34f46218b939bd90a8b0459	Example of input-out pairs, for the four settings, INTENT, DST, NLG and end-to-end (E2E).	2012.15504	example.jpg
0.433025	d6b8da4baf994aff8644abe526277ea2	No Free Lunch in CL. This plot the trade-off between number of parameters added per task and the size of the episodic memory ℳ.	2012.15504	nofreelunch.jpg
0.44416	745c3b2da4d547f990e5289d4851a85d	A three-model homogeneous ensemble with atomic model h, which is shared by the three distinct machines labeling the nodes ŷ_1, ŷ_2, ŷ_3. Predictions are aggregated using the function μ at node ŷ.	2012.15505	homogeneous_ensemble.jpg
0.460672	55d6db6faafb415880131e088335a5ec	A learning network combining a supervised model h_S with target transformation h_T. The supervised composite model receives training features and training target at X and y; when new features are supplied at X, predictions are fetched from ŷ. Nodes z and ŷ are labeled with the same machine, corresponding to a single learned transformation of the target.	2012.15505	target_transformation.jpg
0.462411	9c58e5481f4e462b99fe97f63303be90	The learning network of Figure <ref> with training edges added (shown as dotted arrows).	2012.15505	training_edges.jpg
0.419326	2bc7ac0a8607461faa68600960c8b923	A waveguide with an isosceles trapezoidal ridge of height h_r, base W and base angle π/2-θ. The thickness of the slab is h_s. The dielectric constants of the slab and the ridge, the substrate, and the cladding are ε_1, ε_2, and ε_0 satisfying ε_1 > ε_2 ≥ε_0≥ 1.	2012.15506	structure2.jpg
0.450019	51aba59c020c4d40badfd9a587c104ff	The EGB spectra observed by Fermi-LAT (red points) and calculated with the blazar luminosity function (dotted, solid and dashed lines). The dotted, dashed and solid lines represent the EGB spectra of no EBL absorption (w/oEBL), with EBL considered but no ALP effect (EBL) and including an ALP effect (ALP), respectively. We have assume an optimistic ICM for the ALP model. The gray band reflects the scatter caused by the uncertainty of the LF parameters. The triangles are the lower limits of the EGB spectrum placed by cumulating detected individual extragalactic TeV sources <cit.>. The yellow line is the upper limit on EGB by requiring the cascade emission not to exceed the EGB data below 100 GeV <cit.>.	2012.15513	egbsp.jpg
0.4848	23ec02e9bc7641c9a3ea9f7a4ae01343	The modulated EGB spectra assuming only 10% (grey solid line) or 30% (grey dotted line) of blazars are located in a dense intracluster environment.	2012.15513	egbsp0.1.jpg
0.441469	52e131f79b574cb39731e13c1bb44428	The ALP parameters sensitive to EGB observations (hatched region for B_0=3 μ G and blue dashed line for B_0=1 μ G) comparing to other ALP constraints in the literature. The purple, red and green regions are the constraints by <cit.>, <cit.> and <cit.>, respectively. The red dash-dot line encompasses the exclusion region by <cit.>. The light blue region is the parameter space where the low gamma-ray opacity of the universe can be explained by the ALPs <cit.>.	2012.15513	fig3.jpg
0.446235	126b8ac73ce4460980d6b916db3754eb	Given a query vector and key vectors, the attention network first computes a weight vector through the scoring function.	2012.15515	att-a.jpg
0.447996	071cd3c86d8845988d86b74a005d433e	Each output vector is computed as a weighted sum of value vectors.	2012.15515	att-b.jpg
0.414304	006669f0c7ea43f19724a8158fa3563c	A running example of the beam-search algorithm.	2012.15515	beam.jpg
0.419112	beb6dd2ceddd4885b446f41220cab9f7	Bidirectional inference.	2012.15515	bi-decoding.jpg
0.456902	5f5f79f3e87748709e274801f1d5e90d		2012.15515	cnn-a.jpg
0.515412	79779d76eb854d2bad9a1ef61e51018e		2012.15515	cnn-b.jpg
0.494965	e2d59ecd2f3a4ba2b9692f4431a9c3e4		2012.15515	cnn.jpg
0.45041	542fa98b54ef4254af17f6f7045f9797	CNN with additional zero paddings.	2012.15515	decnn-a.jpg
0.442342	76c7471315f940b098e1bcbef8c9a3e1	SAN with causal masking.	2012.15515	decnn-b.jpg
0.427776	c898c4e2c4e74048b19d829db987271e	At each decoding step, the attention mechanism dynamically generates a context vector based on the most relevant source representations for predicting the next target word.	2012.15515	encdec-att.jpg
0.490849	4d12f739cb6041fbbddc9d8153dbaa0d		2012.15515	lcnn.jpg
0.510108	9d61380591ec4cc98b14955ea1ff3200	Unidirectional language model pre-training.	2012.15515	lm.jpg
0.533513	530030bef81f4713b24c9c3685926923	An illustration of multi-head attention.	2012.15515	mhatt.jpg
0.423703	a80e800ee5194d3693a2b5069ef2a793	Bi-directional language model pre-training.	2012.15515	mlm.jpg
0.424714	bd45d5d74b524ebab1475ecf99371c4b	Non-autoregressive decoding.	2012.15515	nar-decoding.jpg
0.451293	9fc8363218584328ade1debf3ebd69a4	An overview of the NMT architecture, which consists of embedding layers, a classification layer, an encoder network, and a decoder network. We use different colors to distinguish different languages.	2012.15515	nmt.jpg
0.46589	f6439588c215493fae99ab830a627c79	RNN.	2012.15515	nn-a.jpg
0.443584	871b445b0b724fcda8a3392961be00f3	CNN with kernel width k=3.	2012.15515	nn-b.jpg
0.463371	b298e0bd223d40a5b660650acac6ccf4	SAN.	2012.15515	nn-c.jpg
0.532973	92632a557a6148e8b85b4c2f12531a2b	Deep RNN.	2012.15515	rnn-a.jpg
0.490429	a5f2b5a1955a44049d35c229efc9c3e6	Bidirectional RNN.	2012.15515	rnn-b.jpg
0.520866	d5b447530993493f931b4c9899d79418	Alternating RNN.	2012.15515	rnn-c.jpg
0.444044	73fab80e60a24712a238755549e98aa0	Sequence-to-sequence model pre-training.	2012.15515	s2s.jpg
0.525567	f63c2516e41549b29a99f5dfea412c3c	The Transformer architecture.	2012.15515	transformer.jpg
0.397243	ed9ce398978246a8bd5f284c1e52ed20	Replaced Token Detection pre-training task	2012.15516	electra.jpg
0.424764	0733ef717678487ea2ffc399883cab2e	Overview of the methodology	2012.15518	Figure1.jpg
0.426314	c18d6e0adaab4da69683842e12159832	The technical description of âSpherical Robotâ used in the case study (Source: Wikipedia, accessed on June 8, 2020)	2012.15518	Figure2.jpg
0.495563	eee262f7da5c42d5871e7979d2db545b	Graph visualizations based on A) WordNet, B) ConceptNet, and C) TechNet	2012.15518	Figure3.jpg
0.431231	bcb8094d0de8428ebc9b06dc2250bdb9	The distribution of participants' answers to evaluation questions.	2012.15518	Figure4.jpg
0.32868	7b3959b010c24e86a4cd5d2306c7a1ed	The percentage of participants according to their choice of best graph representation	2012.15518	Figure5.jpg
0.432219	a14664af312a449caabf3aea3a2d63d4	Text generated by AraGPT2-Mega. The first sentence is the human-written prompt	2012.15520	GLTR-gen-crop.jpg
0.434202	5f9c9c8e857a424cac72c1a501c39952	Human-Written Text	2012.15520	GLTR-orig-crop.jpg
0.413596	04176376be0941aa99e476516470f126	Survey results showing human error rates on machine generated (left) and human written text (right)	2012.15520	error_rate_machine-crop.jpg
0.433522	7c95605b33844ed5a84012b816855eae	Random unseen context about coronavirus vaccine(top). Followed by two generated samples bu AraGPT2-mega. Generated text 1 (top_p=0.95), Generated text 2 (top_p=1)	2012.15520	mega_corona_vaccine.jpg
0.379414	87289943aa6a412aac00a442d35fff22	Random unseen contexts on miscellaneous topics. Followed by a generated sample by AraGPT2-mega with top_p=0.95	2012.15520	mega_misc.jpg
0.42941	2e7a2cf7a3a84d5e8fee18df3f5ac201	Random unseen contexts about children stories. Followed by a generated sample by AraGPT2-mega with top_p=0.95	2012.15520	mega_story_1.jpg
0.459908	148fa823475d4dfbb7af7d7371642558	The flow of automatic counting key selection. We use c_i and h_i denote contextual features and historical features for impression i, respectively. Symbol u_i, i∈{1, 2, ⋯, n} refers to user i, and uid is short for user ID. Each colored circle is one type of feature. And k_i is one counting key, which is a set of features including user id.	2012.15522	method-overview.jpg
0.496437	f72e226427534c76817cf79de097b00c	Wide & Deep model structure.	2012.15522	nn.jpg
0.407174	e7f483d57f1345e2a919ca11bf2ff50a	The relative RCE improvement over BASE method. Evaluated with logistic regression model in an online learning setting for different features.	2012.15522	rce-rela-lr.jpg
0.512022	43ddc1a9051743dfab40c63facec3082	Schematically plot of the QAHI/SC heterostructure, where a wire-shaped region is removed from the QAHI. The superconductivity is induced by covering a superconductor on the top of the cut region. The red arrows at the edge of QAHI represent the chiral edge states of QAHI. The positions of MZMs γ_1 and γ_2 are denoted by red dots. It is important to note that the MZM at the end of the cut is localized while the MZM at the beginning of the cut is delocalized due to the extended gapless chiral edge modes.	2012.15523	FIG1.jpg
0.402481	c40cdc3dc85a452b8e512a41a9ad31b0	(a) A schematic plot of QAHI/SC geometry. A cut along the y-direction is created in the middle of QAHI. The red arrow denotes the edge states, d denotes the separation between the two edge states. (b) The energy spectrum for a QAHI with a cut, where the edge states near the cut are highlighted in red and the black dashed line indicates the position of chemical potential. Periodic boundary condition in the y-direction is imposed. (b) Schematic of the BdG energy spectrum when the superconducting pairing is introduced onto the edge states. (d) and (e) show the energy E (in unit of Δ) as a function of k_y for a sample with 160 lattice sites in the x-direction. In (d), width of the cut is 30 sites. In (e), the width of the cut is 4 sites. In (b), (d), (e), the magnetization energy M_z=10 meV, the chemical potential μ=3 meV and the pairing potential Δ=1 meV.	2012.15523	FIG2.jpg
0.465276	703834ed06954c57b1a2de3f3f799479	(a) shows the localized wavefunction of MZMs near a cut, where the QAHI/SC sample is 2500 sites × 160 sites and a cut of 2000 sites × 4 sites is created in the middle with one end touching the edge. (b) The simulated tunneling spectra for the QAHI/SC geometry plotted in Fig. <ref>. A quantized 2e^2/h conductance is induced by the MZM. Here, the magnetization energy M_z=10 meV, the chemical potential μ=3 meV, and the temperature T=0.05 K are used.	2012.15523	FIG3.jpg
0.493167	f56a4cedc0c34933917dacbbd3458356	(a) A cross-shape cut on a QAHI/SC heterostructure through which braiding operations of MZMs γ_1 and γ_2 can be performed. Gn (n from 1 to 4) label the four gates made from semiconducting materials. (b) and (c) show the wavefunctions \|ψ\|^2 of MZMs localized at the closed gates with the geometry displayed in (a). G1 and G3 are closed and other gates are open in (b), while G1 and G2 are closed and other gates are open in (c). (d) A schematic picture of a hexon based on the QAHI/SC heterostructure.	2012.15523	FIG4.jpg
0.447943	f5426a4d2f8e4f63961ca34699317a9a	BANG AR generation. The [MASK] can attend to its previous golden tokens to predict y_4.	2012.15525	BANG-AR.jpg
0.506994	0c48faa09e69400c85247a196e60267f	BANG NAR generation. The [MASK] can only attend to its previous [MASK]s to predict y_3.	2012.15525	BANG-NAR.jpg
0.396157	e4bdaa028fc94b91876276afd918e445	BANG semi-NAR generation. BANG semi-NAR inference supports predicting first several tokens step by step then generating rest tokens at one timestep. Here we present the inference procedure. The training procedure is same to Figure <ref>.	2012.15525	BANG-semiNAR.jpg
0.484455	666af28544f74f0891354a778035359b	BANG pretraining. The prediction of y_4 is used as an example. “M-S” and “P-S” represent “Main stream" and “Predicting stream”, respectively. The 1^st predicting stream predicts tokens in AR manner. First tokens in each predicting stream compose NAR prediction.	2012.15525	BANG_pdf.jpg
0.412683	bfb31b54090342698cee123e2295810b	f_0(x)=π/π-2(1-sin(π x))	2012.15527	L2_decay_f1.jpg
0.403716	cfe47d0a0af146578a0ff2fc07b2cd62	f_0(x)=2x	2012.15527	L2_decay_f2.jpg
0.397772	2ca37761327541a9ab55991953683c58	Exponential order of convergence	2012.15527	L2_decay_f2V.jpg
0.535616	acd46f2207f34382bee16e838cf711c0	Stabilization of free energy	2012.15527	energy_decay_f2V.jpg
0.388225	33301b14fe73450297198648ed8e13b5	Convergence to steady state	2012.15527	f2V.jpg
0.464382	7354fc4b5f7645d0ac733806843cd47f	f_0(x)=x^2	2012.15527	phi1b.jpg
0.503497	95020bb45a61419ea0d060d348cc55fb	f_0(x)=max(0,x(0.5-x),(x-0.7)(1-x))	2012.15527	phi2b.jpg
0.468497	fa91106add6747e48eff2e1735566ca5	f_0(x)=1_[0.5,1]	2012.15527	phi3b.jpg
0.400641	ec5afbe1da894a0f8a64b626bcf6cedc	Demo Detection Results on Collected Colonoscopy Dataset. (a) and (b) are video frames of two polyps.	2012.15531	demo.jpg
0.405856	68ff7c1e444643f38a43b89305ba0fea	The flowchart of our Ivy-Net. (a) We create the virtual input x̃_̃ĩ by mixing up the input image with a negative video frame at the pixel level, i.e., sum by random weight λ, without changing its bounding box label ỹ_̃ĩ (Section <ref>). (b) f_j - 1, f_j, f_j + 1 are features of adjacent video frames x_j - 1, x_j, x_j + 1, encoded by the Encoder of polyp detection model. f̂_̂ĵ is estimation of f_j, calculated by f_j - 1 and f_j + 1, e.g. arithmetic mean here. Regularization Loss is calculated from f̂_̂ĵ and f_j by the loss function L_reg.	2012.15531	framework.jpg
0.422514	e7877dc742ec4303af08db622c1d5a5f	Procedure of Finding a Polyp. Modern colonoscopy inserts a long flexible tube from rectum to ileum. That tube contains artificial illuminants, cameras, and channels for air, water as well as devices. Endoscopists observe through cameras during withdrawal of the tube. They can bend and rotate the tube's terminal to adjust the visual field. A typical procedure of finding a polyp into four steps, and CADe is most helpful at Step II.	2012.15531	operation.jpg
0.420873	66d1b6ec6b304c8191e01d32d64e7fde	Encoding the mention using entity markers.	2012.15534	entity_marker.jpg
0.457988	3db3f5492c464131bcab24389e30732f	The retrieval process of HopRetriever for three hops. hop_s,i indicates a beginning jump from the start to e_i is selected based on the initial hidden state 𝐡_s. The selection of hop hop_i,j retrieves the supporting document d_j at the second step. hop_e ends the retrieval process finally.	2012.15534	fig_rnn-crop.jpg
0.454163	10d49fe97a3244d8aa9ec186a168588f	Two examples showing that both structured relation and unstructured fact are needed for complex question answering.	2012.15534	intro_case_2.jpg
0.488772	06f07f570e184eb3b2f9d9e9cc7e85b1	Retrieving Hops over Wikipedia text graph. Documents are retrieved by selecting hops over them iteratively. Each directed arrow implies a mention m_i,j, which reveals how e_i mentions e_j in the document d_i. Hops between entities are indicated by curved arrows. If the mention m_i,j exists between e_i and e_j, the hop hop_i,j is represented based on both m_i,j and the introductory document d_j for retrieval or based on the d_j solely if no mentions exist.	2012.15534	overview_2-crop.jpg
0.486133	de51ffa397fe4ff6bf5e841fba4f5a64	Case 1.	2012.15534	v_case_1.jpg
0.499292	8a3ce7f872024c7384f4bcba7d091415	Case 2.	2012.15534	v_case_2.jpg
0.456204	08e5a65bba1f4784befb698c9011bc35	Case 3.	2012.15534	v_case_3.jpg
0.500814	5cc4790f06484759891f5efa262a2345	The suppression factor β_ rep from Eq. (<ref>) for different temperatures, expressed in the units of the chiral-critical temperature, T_c=T_c(μ_B=0).	2012.15535	all_suppresion_factor.jpg
0.434595	4b797499a68c415181f2aa15364a3754	Schematic representation of Rydberg-ion collisions. The ion is considered the projectile with impact parameter b moving with a velocity v towards the Rydberg atom, placed at the origin of coordinates. The Rydberg electron's position is labeled as 𝐫_1 from the Rydberg core, whereas 𝐫_2 represents its position from the ion. The inset represents a one-dimensional perspective of the charge exchange process: the electron moves through the potential barrier at a given ion-Rydberg distance.	2012.15536	Fig_Ryd_Ion.jpg
0.42123	11fe48e364b94d7ea2d8f6dc0c1f936a	Angular quantum number l' distribution for collisions with n'=n either a) from charge exchange or b) with the electron remaining bound to its original core. Initial state is 90D and with collision energies E_col= 10-10000K as indicated in the legend.	2012.15536	ce_nce_hist_acc_2.jpg

0.600779

64d8432aa3274ee49e5f2003a96879f5

EM

2012.15492

after_em.jpg

0.419817

73749b9839634ded88fac412f8c25d8b

k=10

2012.15492

circles_1.jpg

0.414179

6d45b7a77fed4094bc7235c85c4b6348

k=20

2012.15492

circles_2.jpg

0.454782

3d8e75e7f1b2482db5823da6dd743084

k=30

2012.15492

circles_3.jpg

0.566318

20e76f7d10774dc3ab4f2f31fc8a007e

Correction

2012.15492

corrected.jpg

0.444729

5e828b9b6264466191613bb015541a8c

k=3

2012.15492

moons_1.jpg

0.439095

de0d44c3961e4fab98cdf88fe5ddb26b

k=10

2012.15492

moons_2.jpg

0.453261

7aecd2162c594cbfb3fb88d5e9f46f41

k=17

2012.15492

moons_3.jpg

0.429269

74452efea5da47038194d6bb85fe036c

The Super-k algorithm processing steps

2012.15492

processing_steps.jpg

0.431959

69c95a4d386e4c85bc4b324c5b08461b

k=2

2012.15492

random_1.jpg

0.467467

79317e89f31d4799988eea57cdb0e8cb

k=5

2012.15492

random_2.jpg

0.478955

fafb4bb5201e460ca39b2ddcdf914b65

k=8

2012.15492

random_3.jpg

0.471313

e66cfa3581654bb690bb762a1f86d0a2

Class 0

2012.15492

voxelization_0.jpg

0.472111

314fe599a92c4548a73aa05fca559b29

Class 1

2012.15492

voxelization_1.jpg

0.586297

12c1fb44e0fe4780a29e7ad30eb423e7

Voxelization

2012.15492

voxelized.jpg

0.473316

d661b2bf822a4d3b97637345412a1635

Schematic Diagram of our Zero-Shot KD solution. a) We assume access to a pre-trained teacher. b) We adversarially train the generator using out-of-domain data (OOD). c) Finally we use the generated data and the OOD data for KD.

2012.15495

fig_zskd.jpg

0.516594

43bbb7a65dab4d01ad8e021c9cdf0a69

Two cases for quenches in quantum phase transition. (a) E_0≪Λ. Instantaneous quench (IQ) regime, fast quench (FQ) regime, and Kibble-Zurek (KZ) regime. t_KZ∝ E_0^-1(E_0δ)^1/(zν r). (b) E_0>Λ. Instantaneous quench (IQ) regime, non-universal (NU) regime, and Kibble-Zurek (KZ) regime.

2012.15501

figure1.jpg

0.471599

e34159590b124f02aafb3d2c571838b3

The first and second cumulants of the work (absolute value) as a function of the quenching rate for a scalar field with time-dependent mass. The parameters are: E_0=m=0.1, Λ=10. (a) and (c): t_f=0. (b) and (d): t_f→∞. The solid lines denote (κ_n)_re in the Kibble-Zurek regime and fast-quench regime and (κ_n)_su-κ_n in the instantaneous quench regime. The dashed lines denote the fitting curves.

2012.15501

g2.jpg

0.494133

790d78c87fe24872a66dfeaf134da933

The competition between the frozen time t_KZ and the quench time scale δ in the adiabatic-impulse-adiabatic stages. The black solid line denotes the energy gap E_g. (a) the Kibble-Zurek regime. A part of the protocol is inside the impulse stage (t_KZ<δ). (b) the fast quench regime. All the protocol is inside the impulse stage (t_KZ>δ).

2012.15501

kzm-cft.jpg

0.42115

d283b0f8518644b4a42462eeb1a7a114

High-level representation of the AdapterCL.

2012.15504

ADPTCL.jpg

0.471628

d1acf272a4944840b63b280bbdc302a0

In Continual Learning, the model is trained one dataset at the time. In this instance, the model is first trained on data from the hotel domain D_Hotel then on the Taxi D_Taxi and so on. The parameter of the model is updated sequentially based on the loss function L.

2012.15504

CLVANILLA.jpg

0.601604

1c0b48c69cba48c38aad1f4c50e58072

Ablation study on the size of the episodic-memory vs JGA.

2012.15504

DSTMem.jpg

0.527535

400eef6903454528bf2f3c86b5c8593d

Ablation study on the size of the episodic-memory vs the intent accuracy.

2012.15504

IntentMem.jpg

0.427626

733be8c1a34f46218b939bd90a8b0459

Example of input-out pairs, for the four settings, INTENT, DST, NLG and end-to-end (E2E).

2012.15504

example.jpg

0.433025

d6b8da4baf994aff8644abe526277ea2

No Free Lunch in CL. This plot the trade-off between number of parameters added per task and the size of the episodic memory ℳ.

2012.15504

nofreelunch.jpg

0.44416

745c3b2da4d547f990e5289d4851a85d

A three-model homogeneous ensemble with atomic model h, which is shared by the three distinct machines labeling the nodes ŷ_1, ŷ_2, ŷ_3. Predictions are aggregated using the function μ at node ŷ.

2012.15505

homogeneous_ensemble.jpg

0.460672

55d6db6faafb415880131e088335a5ec

A learning network combining a supervised model h_S with target transformation h_T. The supervised composite model receives training features and training target at X and y; when new features are supplied at X, predictions are fetched from ŷ. Nodes z and ŷ are labeled with the same machine, corresponding to a single learned transformation of the target.

2012.15505

target_transformation.jpg

0.462411

9c58e5481f4e462b99fe97f63303be90

The learning network of Figure <ref> with training edges added (shown as dotted arrows).

2012.15505

training_edges.jpg

0.419326

2bc7ac0a8607461faa68600960c8b923

A waveguide with an isosceles trapezoidal ridge of height h_r, base W and base angle π/2-θ. The thickness of the slab is h_s. The dielectric constants of the slab and the ridge, the substrate, and the cladding are ε_1, ε_2, and ε_0 satisfying ε_1 > ε_2 ≥ε_0≥ 1.

2012.15506

structure2.jpg

0.450019

51aba59c020c4d40badfd9a587c104ff

The EGB spectra observed by Fermi-LAT (red points) and calculated with the blazar luminosity function (dotted, solid and dashed lines). The dotted, dashed and solid lines represent the EGB spectra of no EBL absorption (w/oEBL), with EBL considered but no ALP effect (EBL) and including an ALP effect (ALP), respectively. We have assume an optimistic ICM for the ALP model. The gray band reflects the scatter caused by the uncertainty of the LF parameters. The triangles are the lower limits of the EGB spectrum placed by cumulating detected individual extragalactic TeV sources <cit.>. The yellow line is the upper limit on EGB by requiring the cascade emission not to exceed the EGB data below 100 GeV <cit.>.

2012.15513

egbsp.jpg

0.4848

23ec02e9bc7641c9a3ea9f7a4ae01343

The modulated EGB spectra assuming only 10% (grey solid line) or 30% (grey dotted line) of blazars are located in a dense intracluster environment.

2012.15513

egbsp0.1.jpg

0.441469

52e131f79b574cb39731e13c1bb44428

The ALP parameters sensitive to EGB observations (hatched region for B_0=3 μ G and blue dashed line for B_0=1 μ G) comparing to other ALP constraints in the literature. The purple, red and green regions are the constraints by <cit.>, <cit.> and <cit.>, respectively. The red dash-dot line encompasses the exclusion region by <cit.>. The light blue region is the parameter space where the low gamma-ray opacity of the universe can be explained by the ALPs <cit.>.

2012.15513

fig3.jpg

0.446235

126b8ac73ce4460980d6b916db3754eb

Given a query vector and key vectors, the attention network first computes a weight vector through the scoring function.

2012.15515

att-a.jpg

0.447996

071cd3c86d8845988d86b74a005d433e

Each output vector is computed as a weighted sum of value vectors.

2012.15515

att-b.jpg

0.414304

006669f0c7ea43f19724a8158fa3563c

A running example of the beam-search algorithm.

2012.15515

beam.jpg

0.419112

beb6dd2ceddd4885b446f41220cab9f7

Bidirectional inference.

2012.15515

bi-decoding.jpg

0.456902

5f5f79f3e87748709e274801f1d5e90d

2012.15515

cnn-a.jpg

0.515412

79779d76eb854d2bad9a1ef61e51018e

2012.15515

cnn-b.jpg

0.494965

e2d59ecd2f3a4ba2b9692f4431a9c3e4

2012.15515

cnn.jpg

0.45041

542fa98b54ef4254af17f6f7045f9797

CNN with additional zero paddings.

2012.15515

decnn-a.jpg

0.442342

76c7471315f940b098e1bcbef8c9a3e1

SAN with causal masking.

2012.15515

decnn-b.jpg

0.427776

c898c4e2c4e74048b19d829db987271e

At each decoding step, the attention mechanism dynamically generates a context vector based on the most relevant source representations for predicting the next target word.

2012.15515

encdec-att.jpg

0.490849

4d12f739cb6041fbbddc9d8153dbaa0d

2012.15515

lcnn.jpg

0.510108

9d61380591ec4cc98b14955ea1ff3200

Unidirectional language model pre-training.

2012.15515

lm.jpg

0.533513

530030bef81f4713b24c9c3685926923

An illustration of multi-head attention.

2012.15515

mhatt.jpg

0.423703

a80e800ee5194d3693a2b5069ef2a793

Bi-directional language model pre-training.

2012.15515

mlm.jpg

0.424714

bd45d5d74b524ebab1475ecf99371c4b

Non-autoregressive decoding.

2012.15515

nar-decoding.jpg

0.451293

9fc8363218584328ade1debf3ebd69a4

An overview of the NMT architecture, which consists of embedding layers, a classification layer, an encoder network, and a decoder network. We use different colors to distinguish different languages.

2012.15515

nmt.jpg

0.46589

f6439588c215493fae99ab830a627c79

RNN.

2012.15515

nn-a.jpg

0.443584

871b445b0b724fcda8a3392961be00f3

CNN with kernel width k=3.

2012.15515

nn-b.jpg

0.463371

b298e0bd223d40a5b660650acac6ccf4

SAN.

2012.15515

nn-c.jpg

0.532973

92632a557a6148e8b85b4c2f12531a2b

Deep RNN.

2012.15515

rnn-a.jpg

0.490429

a5f2b5a1955a44049d35c229efc9c3e6

Bidirectional RNN.

2012.15515

rnn-b.jpg

0.520866

d5b447530993493f931b4c9899d79418

Alternating RNN.

2012.15515

rnn-c.jpg

0.444044

73fab80e60a24712a238755549e98aa0

Sequence-to-sequence model pre-training.

2012.15515

s2s.jpg

0.525567

f63c2516e41549b29a99f5dfea412c3c

The Transformer architecture.

2012.15515

transformer.jpg

0.397243

ed9ce398978246a8bd5f284c1e52ed20

Replaced Token Detection pre-training task

2012.15516

electra.jpg

0.424764

0733ef717678487ea2ffc399883cab2e

Overview of the methodology

2012.15518

Figure1.jpg

0.426314

c18d6e0adaab4da69683842e12159832

The technical description of âSpherical Robotâ used in the case study (Source: Wikipedia, accessed on June 8, 2020)

2012.15518

Figure2.jpg

0.495563

eee262f7da5c42d5871e7979d2db545b

Graph visualizations based on A) WordNet, B) ConceptNet, and C) TechNet

2012.15518

Figure3.jpg

0.431231

bcb8094d0de8428ebc9b06dc2250bdb9

The distribution of participants' answers to evaluation questions.

2012.15518

Figure4.jpg

0.32868

7b3959b010c24e86a4cd5d2306c7a1ed

The percentage of participants according to their choice of best graph representation

2012.15518

Figure5.jpg

0.432219

a14664af312a449caabf3aea3a2d63d4

Text generated by AraGPT2-Mega. The first sentence is the human-written prompt

2012.15520

GLTR-gen-crop.jpg

0.434202

5f9c9c8e857a424cac72c1a501c39952

Human-Written Text

2012.15520

GLTR-orig-crop.jpg

0.413596

04176376be0941aa99e476516470f126

Survey results showing human error rates on machine generated (left) and human written text (right)

2012.15520

error_rate_machine-crop.jpg

0.433522

7c95605b33844ed5a84012b816855eae

Random unseen context about coronavirus vaccine(top). Followed by two generated samples bu AraGPT2-mega. Generated text 1 (top_p=0.95), Generated text 2 (top_p=1)

2012.15520

mega_corona_vaccine.jpg

0.379414

87289943aa6a412aac00a442d35fff22

Random unseen contexts on miscellaneous topics. Followed by a generated sample by AraGPT2-mega with top_p=0.95

2012.15520

mega_misc.jpg

0.42941

2e7a2cf7a3a84d5e8fee18df3f5ac201

Random unseen contexts about children stories. Followed by a generated sample by AraGPT2-mega with top_p=0.95

2012.15520

mega_story_1.jpg

0.459908

148fa823475d4dfbb7af7d7371642558

The flow of automatic counting key selection. We use c_i and h_i denote contextual features and historical features for impression i, respectively. Symbol u_i, i∈{1, 2, ⋯, n} refers to user i, and uid is short for user ID. Each colored circle is one type of feature. And k_i is one counting key, which is a set of features including user id.

2012.15522

method-overview.jpg

0.496437

f72e226427534c76817cf79de097b00c

Wide & Deep model structure.

2012.15522

nn.jpg

0.407174

e7f483d57f1345e2a919ca11bf2ff50a

The relative RCE improvement over BASE method. Evaluated with logistic regression model in an online learning setting for different features.

2012.15522

rce-rela-lr.jpg

0.512022

43ddc1a9051743dfab40c63facec3082

Schematically plot of the QAHI/SC heterostructure, where a wire-shaped region is removed from the QAHI. The superconductivity is induced by covering a superconductor on the top of the cut region. The red arrows at the edge of QAHI represent the chiral edge states of QAHI. The positions of MZMs γ_1 and γ_2 are denoted by red dots. It is important to note that the MZM at the end of the cut is localized while the MZM at the beginning of the cut is delocalized due to the extended gapless chiral edge modes.

2012.15523

FIG1.jpg

0.402481

c40cdc3dc85a452b8e512a41a9ad31b0

(a) A schematic plot of QAHI/SC geometry. A cut along the y-direction is created in the middle of QAHI. The red arrow denotes the edge states, d denotes the separation between the two edge states. (b) The energy spectrum for a QAHI with a cut, where the edge states near the cut are highlighted in red and the black dashed line indicates the position of chemical potential. Periodic boundary condition in the y-direction is imposed. (b) Schematic of the BdG energy spectrum when the superconducting pairing is introduced onto the edge states. (d) and (e) show the energy E (in unit of Δ) as a function of k_y for a sample with 160 lattice sites in the x-direction. In (d), width of the cut is 30 sites. In (e), the width of the cut is 4 sites. In (b), (d), (e), the magnetization energy M_z=10 meV, the chemical potential μ=3 meV and the pairing potential Δ=1 meV.

2012.15523

FIG2.jpg

0.465276

703834ed06954c57b1a2de3f3f799479

(a) shows the localized wavefunction of MZMs near a cut, where the QAHI/SC sample is 2500 sites × 160 sites and a cut of 2000 sites × 4 sites is created in the middle with one end touching the edge. (b) The simulated tunneling spectra for the QAHI/SC geometry plotted in Fig. <ref>. A quantized 2e^2/h conductance is induced by the MZM. Here, the magnetization energy M_z=10 meV, the chemical potential μ=3 meV, and the temperature T=0.05 K are used.

2012.15523

FIG3.jpg

0.493167

f56a4cedc0c34933917dacbbd3458356

(a) A cross-shape cut on a QAHI/SC heterostructure through which braiding operations of MZMs γ_1 and γ_2 can be performed. Gn (n from 1 to 4) label the four gates made from semiconducting materials. (b) and (c) show the wavefunctions |ψ|^2 of MZMs localized at the closed gates with the geometry displayed in (a). G1 and G3 are closed and other gates are open in (b), while G1 and G2 are closed and other gates are open in (c). (d) A schematic picture of a hexon based on the QAHI/SC heterostructure.

2012.15523

FIG4.jpg

0.447943

f5426a4d2f8e4f63961ca34699317a9a

BANG AR generation. The [MASK] can attend to its previous golden tokens to predict y_4.

2012.15525

BANG-AR.jpg

0.506994

0c48faa09e69400c85247a196e60267f

BANG NAR generation. The [MASK] can only attend to its previous [MASK]s to predict y_3.

2012.15525

BANG-NAR.jpg

0.396157

e4bdaa028fc94b91876276afd918e445

BANG semi-NAR generation. BANG semi-NAR inference supports predicting first several tokens step by step then generating rest tokens at one timestep. Here we present the inference procedure. The training procedure is same to Figure <ref>.

2012.15525

BANG-semiNAR.jpg

0.484455

666af28544f74f0891354a778035359b

BANG pretraining. The prediction of y_4 is used as an example. “M-S” and “P-S” represent “Main stream" and “Predicting stream”, respectively. The 1^st predicting stream predicts tokens in AR manner. First tokens in each predicting stream compose NAR prediction.

2012.15525

BANG_pdf.jpg

0.412683

bfb31b54090342698cee123e2295810b

f_0(x)=π/π-2(1-sin(π x))

2012.15527

L2_decay_f1.jpg

0.403716

cfe47d0a0af146578a0ff2fc07b2cd62

f_0(x)=2x

2012.15527

L2_decay_f2.jpg

0.397772

2ca37761327541a9ab55991953683c58

Exponential order of convergence

2012.15527

L2_decay_f2V.jpg

0.535616

acd46f2207f34382bee16e838cf711c0

Stabilization of free energy

2012.15527

energy_decay_f2V.jpg

0.388225

33301b14fe73450297198648ed8e13b5

Convergence to steady state

2012.15527

f2V.jpg

0.464382

7354fc4b5f7645d0ac733806843cd47f

f_0(x)=x^2

2012.15527

phi1b.jpg

0.503497

95020bb45a61419ea0d060d348cc55fb

f_0(x)=max(0,x(0.5-x),(x-0.7)(1-x))

2012.15527

phi2b.jpg

0.468497

fa91106add6747e48eff2e1735566ca5

f_0(x)=1_[0.5,1]

2012.15527

phi3b.jpg

0.400641

ec5afbe1da894a0f8a64b626bcf6cedc

Demo Detection Results on Collected Colonoscopy Dataset. (a) and (b) are video frames of two polyps.

2012.15531

demo.jpg

0.405856

68ff7c1e444643f38a43b89305ba0fea

The flowchart of our Ivy-Net. (a) We create the virtual input x̃_̃ĩ by mixing up the input image with a negative video frame at the pixel level, i.e., sum by random weight λ, without changing its bounding box label ỹ_̃ĩ (Section <ref>). (b) f_j - 1, f_j, f_j + 1 are features of adjacent video frames x_j - 1, x_j, x_j + 1, encoded by the Encoder of polyp detection model. f̂_̂ĵ is estimation of f_j, calculated by f_j - 1 and f_j + 1, e.g. arithmetic mean here. Regularization Loss is calculated from f̂_̂ĵ and f_j by the loss function L_reg.

2012.15531

framework.jpg

0.422514

e7877dc742ec4303af08db622c1d5a5f

Procedure of Finding a Polyp. Modern colonoscopy inserts a long flexible tube from rectum to ileum. That tube contains artificial illuminants, cameras, and channels for air, water as well as devices. Endoscopists observe through cameras during withdrawal of the tube. They can bend and rotate the tube's terminal to adjust the visual field. A typical procedure of finding a polyp into four steps, and CADe is most helpful at Step II.

2012.15531

operation.jpg

0.420873

66d1b6ec6b304c8191e01d32d64e7fde

Encoding the mention using entity markers.

2012.15534

entity_marker.jpg

0.457988

3db3f5492c464131bcab24389e30732f

The retrieval process of HopRetriever for three hops. hop_s,i indicates a beginning jump from the start to e_i is selected based on the initial hidden state 𝐡_s. The selection of hop hop_i,j retrieves the supporting document d_j at the second step. hop_e ends the retrieval process finally.

2012.15534

fig_rnn-crop.jpg

0.454163

10d49fe97a3244d8aa9ec186a168588f

Two examples showing that both structured relation and unstructured fact are needed for complex question answering.

2012.15534

intro_case_2.jpg

0.488772

06f07f570e184eb3b2f9d9e9cc7e85b1

Retrieving Hops over Wikipedia text graph. Documents are retrieved by selecting hops over them iteratively. Each directed arrow implies a mention m_i,j, which reveals how e_i mentions e_j in the document d_i. Hops between entities are indicated by curved arrows. If the mention m_i,j exists between e_i and e_j, the hop hop_i,j is represented based on both m_i,j and the introductory document d_j for retrieval or based on the d_j solely if no mentions exist.

2012.15534

overview_2-crop.jpg

0.486133

de51ffa397fe4ff6bf5e841fba4f5a64

Case 1.

2012.15534

v_case_1.jpg

0.499292

8a3ce7f872024c7384f4bcba7d091415

Case 2.

2012.15534

v_case_2.jpg

0.456204

08e5a65bba1f4784befb698c9011bc35

Case 3.

2012.15534

v_case_3.jpg

0.500814

5cc4790f06484759891f5efa262a2345

The suppression factor β_ rep from Eq. (<ref>) for different temperatures, expressed in the units of the chiral-critical temperature, T_c=T_c(μ_B=0).

2012.15535

all_suppresion_factor.jpg

0.434595

4b797499a68c415181f2aa15364a3754

Schematic representation of Rydberg-ion collisions. The ion is considered the projectile with impact parameter b moving with a velocity v towards the Rydberg atom, placed at the origin of coordinates. The Rydberg electron's position is labeled as 𝐫_1 from the Rydberg core, whereas 𝐫_2 represents its position from the ion. The inset represents a one-dimensional perspective of the charge exchange process: the electron moves through the potential barrier at a given ion-Rydberg distance.

2012.15536

Fig_Ryd_Ion.jpg

0.42123

11fe48e364b94d7ea2d8f6dc0c1f936a

Angular quantum number l' distribution for collisions with n'=n either a) from charge exchange or b) with the electron remaining bound to its original core. Initial state is 90D and with collision energies E_col= 10-10000K as indicated in the legend.

2012.15536

ce_nce_hist_acc_2.jpg