text stringlengths 54 548k | label stringclasses 4
values | id_ stringlengths 32 32 |
|---|---|---|
The ability to communicate effectively with other agents is part of a necessary skill repertoire of intelligent agents and commonly seen as one of the great achievements of humanity. A number of papers have studied emergent communication in multi-agent settings {{cite:1068e435a362968ce4828b5278a9e07b07420337}}, {{cite:03488a1eb1e864e3cd8be894a0b90be8b599c3ed}}, {{cite:24541e78e21a9a3e232386a35ad68af1685d713a}} {{cite:6c133685ffec40b9c574fc49ab065a0f68416271}}. The most common problem settings studied are communication tasks, in which agents can exchange messages through symbolic (discrete) cheap-talk channels, that have no impact on the reward function or transition dynamics. A common task is the so called referential game, in which a sender observes an intent that needs to be communicated to a listener via a message.
| i | c0688d6de7e337991cd88a582c85786e |
We introduced a multitask learning approach to AQA and showed that MTL performs better than STL because of better generalization which is especially important in AQA and skill assessment since datasets are small. We showed that the representations learned by our MTL models are better able to capture the inherent concept of quality of actions. Our approach is scalable since the supervision required for the auxiliary tasks is readily available from the existing video footage with minimal extra effort compared to just AQA labeling. In addition, state-of-the-art performance was achieved without any finetuning of hyperparameters. Our best performing and recommended model, C3D-AVG-MTL, achieved 90.44% correlation with judged scores which still leaves a small gap to achieve human-experts-level performance (96% {{cite:394f0b97e9b28bdf0cf3b8ff09573bb979ab10f3}}).
| d | 00ee1dbd9fd268257c0cd4b72ab4891a |
We use faster r-cnn {{cite:8f58806319c6825eb82ab6f00988db6abf65fad4}} and mask r-cnn {{cite:1b5f1a753bdf5878defe296a733fccd422540e5b}} algorithms to benchmark the newly created dataset for graphical object detection task in business documents. Experimentally, we observe that the creation of a model trained with iiit-ar-13k dataset achieves better performance than the model trained with the larger existing datasets. From the experiments, we also observe that the model trained with the larger existing datasets achieves the best performance by fine-tuning with a small number (only 1k) of images from iiit-ar-13k dataset.
| i | 86bfd0c17e667c34c197dcf6a16e7707 |
It follows from {{cite:7b1f3bee17a759ae98e1e0f39ea1baddd4cd1ed2}} that
{{formula:7e2101bd-e3d2-49eb-89ae-c24fab1e402d}}
| m | b47b238683896851d61ff388d57c9b81 |
This is reminiscent of the case of asymptotic coherence distillation. As shown in Ref. {{cite:32a5eb1b7cf535247de18d198ff43588621135c6}}, in the asymptotic case, the distillable coherence of {{formula:ed6d6f4d-4ae4-49c8-8846-3494ed0f371f}} is
{{formula:f9a5dcdd-e824-4b00-bf81-98f646a0f80b}}
| d | 34cf21095856e8dcb328cd1cc7f6f3fe |
Related to GAN reconstruction quality is the ability to efficiently find a latent code that corresponds to a target image. Longer optimization can better reconstruct the image, but becomes intractable over a large dataset. In supplementary material, we investigate classification accuracy as a function of the number of optimization steps, and also perform experiments using an alternative inversion method {{cite:2e4e945c1f51cb6dc72ae41ffd0f45e711ab30a2}} on a smaller face GAN, obtaining similar results. Moreover, recent alternative architectures trained specifically for efficiency {{cite:9f7a729ee2e9b5d9664ec68f8ccdaecc9f2f2580}} or invertibility {{cite:148bee9f51b882ebd2ca83bd88a9637125519a8b}} may help further reduce the computational cost of image reconstruction.
| d | 1f5a4920b83838735770f6408562cf8a |
While fitting the LSPM is computationally feasible on the networks considered here, there are many improvements that could be made to shorten run time. For example, a preprocessing step can be used to decide the truncation level of the dimension of the latent space to be fitted. Using a truncation level that is close to the effective dimension typically resulted in better mixing and convergence times as resources are not wasted inferring negligible dimensions. An adaptive procedure as proposed by {{cite:b1c01642a25dd325f1bea158fdbd5e4ecf34065b}} could be used to increase or decrease the dimension as needed. For larger networks, embedding the case-control concepts of {{cite:6b9e10d523b2e7a234cb7b761ce945aaa2a21a55}} and variational Bayesian inference of {{cite:7d8cfbc027f56ec81977d5fa019978c63cb2060d}} in the LSPM would also bring computational gains.
| d | 62f4b6cd00f9f422b4ead384283aabcd |
This appears to be a viable alternative to that which we have
presented in the main text; it is worth briefly outlining the
method. The basic idea is similar to operator (Strang) splitting
and the method is discussing in some detail in
{{cite:15f6660e500abe7b3836ff6f87f9b941118db2da}}, {{cite:5b1bd7cdaf76b203bd5e1057b687cff3bede6870}}. As noted by Boyd the conventional
centered finite difference schemes can easily be modified by using
spectral differentiation matrices, let {{formula:fadfa3bb-12ad-496e-a3f6-dca2139edfbd}} denote the {{formula:2c3bbadf-6871-41c5-bc29-fb7a6f596366}}
Fourier differentiation matrix ({{cite:9796c227277709da72fa267bfdc263ca121da5df}},
{{cite:15f6660e500abe7b3836ff6f87f9b941118db2da}},{{cite:1023f9a25fe6ce2fbe9f9d8cd3940d6bd1c9ec1a}}{{cite:383ca3abf025f6494572f3eb49edfd872335abbe}}). ADI, or at
least the ADI we use here, means that we split each time step into
two and we first deal implicitly with one set of space derivatives
and then in the next half time step with the other. So
{{formula:37de399e-00be-4a02-b7c9-d47a25311233}}
| m | 0586db4f98be114f7d6749e5be892e65 |
FPFH {{cite:a17a9f1d0efe2308236af3ab04af68c599f37b1b}} and FCGF {{cite:1f3bf1a2493134b0d6e0dc664d50993efb79fa67}}: State-of-the-art hand-crafted and learned geometric descriptors used for point cloud matching. We densely extract these descriptors and train a custom shape vocabulary on the global map. We also compare the traditional feature SIFT and use the corresponding depth points for later camera pose recovering.
Semantic visual localization {{cite:697955516b87cc77f459f7f45ea372bbd47a12ca}}: For simplicity, we denote this work SVL.
We reimplemented the method which uses a variational auto-encoder to learn the representation of point observations. We randomly select 20 scenes from the ScanNet dataset and fuse several key-frames to obtain local semantic maps for training, and test its performance on other ScanNet scenes.
| m | a5fdff64256b99dc706f4a3ad31e6cad |
IIR-SNN with and without batch-norm. Recent works have achieved low latency by adopting batch-normalization (bn) suitably in SNNs {{cite:38a774de214c4dd03cf80e357a15558f049b4a88}}, {{cite:984a80d39783b48e92c1bb6ade048cd7db810e21}}. To disentangle the effect of bn from the proposed gradual latency reduction scheme and ensure that achieving convergence with 1 timestep is orthogonal to using bn, we perform ablation studies as shown in Table REF . For both CIFAR10 and CIFAR100, IIR-SNN enables training with T5 to T1 irrespective of using bn during ANN training. Using bn enhances accuracy, but sequential latency reduction can be performed independently from it. Also, {{cite:38a774de214c4dd03cf80e357a15558f049b4a88}} report that using threshold-dependent bn allows reducing latency up to a minimum of 6 timesteps, but IIR-SNN can go up to 1. Note, in IIR-SNN, bn is used only during ANN training and the bn parameters are fused with the weights during ANN-SNN conversion as proposed in {{cite:bc7a84c04a2051d5e82260bf9dd8346c9502c98b}}; bn is not used in SNN domain training, so the activations at each layer output remain binary.
{{table:c3b1a4e7-a930-452f-aed3-4dd51a5f57ca}} | r | 88d614b096ecb7635d073f4a2d3fca87 |
To conduct experiments and determine the effectiveness of our method, we choose three different biomedical image segmentation datasets as the use case. These three different datasets have different types of segmentation masks. At first, we experimented with the 2018 Data Science Bowl (DSB) Challenge dataset {{cite:4fccbfc336dbddbcacee2b2eb63edd43933bb232}} which contains 670 segmented nuclei images. The nuclei images found in DSB 2018 were captured under varying conditions like different cell sizes, magnification, and imaging modality. This variation within the distribution makes the segmentation of nuclei images a challenging problem. Next, we experimented using the ISIC-2018 Challenge {{cite:9bbc4e32288e8b72dffa3e28d8b200ccad4f199d}}, {{cite:662c2523ef5f1a740c655dc3ac58afc76ba36ece}} dataset, which is a skin segmentation dataset. Skin lesion segmentation assists in melanoma detection, melanoma being the most severe form of skin cancer, warrants an automatic skin lesion segmentation system. Therefore, developing automated systems could be helpful in the clinic. Our third dataset used in the experiments was the Kvasir-Instruments {{cite:4abb5ddd94cd3b5d32c30fff63100f7333bf1727}} dataset which is a diagnostic and therapeutic tool segmentation dataset in gastrointestinal endoscopy. Tool segmentation in gastrointestinal images allows tracking of instruments used during endoscopy and could assist robotic and non-robotic surgeries. Developing such an automated segmentation system might help in complex real-time surgeries inside the gastrointestinal tract.
{{table:9f5abf3b-29a5-4425-b84b-91ce239c2467}}{{table:54b09a8f-4393-492a-9a19-6336bd5f150c}}{{table:c21d29af-cca5-4774-acab-f7356b3ae8b1}} | r | 0ec7abab84f96f924e718f18e4fe0590 |
In recent years, several vector charmonium-like states have been observed, such as the {{formula:f28f4f66-f428-4ad0-b270-7775add16e15}} , {{formula:3f6b9034-52fa-4465-b898-e309c87fba38}} ,
{{formula:71390ad3-ff35-47cc-b3db-8bc1a22ff9eb}} , {{formula:d4b10db4-493b-440e-b528-b5214b08878f}} , {{formula:c7e07177-33b1-454c-819c-4391f7784f60}} , {{formula:5315cc4f-98cf-4a0c-9f09-d0ab3e99e123}} , {{formula:b1961e2a-ecc6-4d5d-9e8f-2093029a10d0}} , {{formula:fd54b787-f907-44ad-8754-2e7cb7c4a273}} , {{formula:d215e48b-f38b-4d0c-bf80-b9cdc15599f1}} , {{formula:a765579c-f7d6-4d92-9ea1-43d429ce05c6}} , etc, they cannot be accommodated suitably in the conventional two quark model, we have to introduce additional quark or gluon degrees of freedom to explore their properties {{cite:41060067ca59bd831ee6b6cc67906c6a43138d87}}.
For example, the {{formula:4ff1a0be-05b9-471a-87a3-e660ea7a1bb0}} observed by the BaBar collaboration {{cite:24e3898cc523869130a898335e0a73520537c755}}, the {{formula:c2365369-1a70-4f79-b046-cf2ebf6b1e71}} , {{formula:012626ec-8722-4c50-a721-ec6b0ccfa590}} and {{formula:36b23cc7-b9aa-473f-98a9-567d6fd5d8d2}} observed by the BESIII collaboration {{cite:d2a66525f57542ee0c71eb21565efdd36bdcafc0}}, {{cite:91e2884500eb89c5c8583c6b5ca2716eb64ae554}}, and the {{formula:9649119f-f50f-4dfe-8926-ee61a1a4c75e}} , {{formula:9cce029c-3a0c-4485-8da4-226ace259bee}} , {{formula:048d6b8c-57d9-486c-964e-a10ea29c7111}} observed by the
Belle collaboration {{cite:26dec52e19fe2392255373cabfd40df315d48ad7}}, {{cite:cc1cc6aec9180bfa776f0fbac65904af811fd7b6}}, {{cite:374bfe5b64bd49e9a48c34175cd0b28042c1d16d}} are excellent candidates for the vector tetraquark states. Considering the analogous masses and decay widths, we can take the {{formula:5c9150db-24f5-472f-88ac-62ef1a610c9b}} , {{formula:3a7b3f99-599c-4cfe-b7ba-96bd8a9c8b49}} and {{formula:8c399198-e9ae-4641-ab9a-830999bd3319}} as the same meson, the {{formula:3fca78cd-57f0-4143-8baa-605bd0db3c44}} and {{formula:168cf927-df5e-451a-8078-5a8c0648d8e2}} as the same meson, and the {{formula:4397a03f-e31d-4b7e-98bf-3fed81a7f1fd}} and {{formula:c15f917b-698e-4691-bb7c-42b36d9f95d2}} as the same meson {{cite:41060067ca59bd831ee6b6cc67906c6a43138d87}}.
| i | 8b17a0c13bf0c45a443ebde8697cebbb |
Modified weak formulation.
The nonlinear system () is equipped with non-homogeneous boundary conditions. The analysis in this paper is based on reformulation of () using lifting technique {{cite:5d6772ddeff3897fa8248008aaa93af64f82884b}}
that reduces the problem to a system of nonlinear PDEs with homogeneous boundary conditions.
| r | aed99901a88f522ed830bb189728694c |
Protons (pions) above 200 (300) MeV kinetic energy have simple or no nuclear corrections as long as formation zone effects can be ignored. Either {{formula:5f29e7c9-cfc2-456d-949a-c73891ec57fb}} or transparency as a validation goal is equally correct.
Low energy protons (similar effects will apply to low energy neutrons) have the most sensitivity to nuclear effects. Some of the codes in this study make severe approximations for these particles and can give untrustworthy results.
For example, Pauli blocking is significant at lower energies and its implementation varies widely among the codes.
It is notable that INCL++ has the best nuclear model and has the best agreement with proton {{formula:8ac040ed-7425-463d-9b61-b42c30a222a9}} data. Although all simulations have good agreement with the lowest energy transparency data at 180 MeV {{cite:94bb8e6d9e155f7c39eb4397325c96ae840b5e5c}}, this misses the region with the most sensitivity. The importance of this finding will vary among experiments. Experiments that run at low neutrino energy {{cite:940fb0d473ce7368f4764c02303e9908d17a7dce}} or that need significantly improved accuracy {{cite:e42acdef00b90eb3fcd44586ea80a52180cb4636}} will be most affected. With the threshold for proton detection down to 47 MeV in the MicroBooNE liquid argon detector results {{cite:ebefa64d65c0240f99aaa6d37811e203c22dc13e}}, examination of these effects is becoming possible.
Pions of any charge at energies where the {{formula:9d73ca61-6e90-4b14-86f0-4cbe5fb177fe}} resonance is important are difficult for all the codes studied here. This is well known {{cite:fe353521ab6f83ddabfb71cc777bbcc52797768e}}, {{cite:7dd5c254dd98cd8acc35319079eec0373af142a2}} and improvements in them are necessary. Although all codes have the effect of the propagating {{formula:341509bc-bba5-4c4a-82c2-f832838d7a0b}} , INCL++ has the additional off-shell effects of the propagating pion which produces a noticeable effect. Some of the codes are more than two standard deviations from the {{formula:a7e315af-dac7-4f65-b818-ff17f1dc6f93}} data. Interesting sensitivity to effects in the pion transparency await data for testing.
| d | 584022f930ccc3b06a089aca3e795f3e |
The proposed model aims at learning a representation that predicts the sequence of encoded motion {{formula:6ed1c04f-9caa-4e83-9694-2cc754ac678d}} over {{formula:548deab9-dbd3-4150-b3b0-c73be8411348}} future time steps given an input image sequence {{formula:197520b1-8409-4e25-b084-74f8ebdf6cf4}} of length
{{formula:4f73fe33-c171-4332-a969-9ec45fc56746}} . Consecutive pair of images were non-rigidly registered applying the B-spline transformation model and optimization based on normalized mutual information as implemented by NiftyReg software {{cite:82496141523d358cfe6a30476d5415a232603cac}}. The resulting two-dimensional displacement fields were encoded using an auxiliary representation space {{formula:2f56e349-f356-4f5e-846f-50cff2a13d22}} = {{formula:7ef2c851-ac6f-4399-b927-22c4d7d702dc}} where {{formula:848f6b7a-b4cf-4ae4-ae3d-deb6467c90a9}} and {{formula:359eaca8-8920-4cf9-ab9f-4bb1728632d1}} is the number of classes. {{formula:3c2060da-1322-4cee-9a35-69d17c4c7a9c}} is a mapping function to encode the displacement fields into labels. To that end, the ranges of values for each vectorial component, i.e. axes {{formula:21b9e464-8448-4b11-a938-a3fdefd4f56e}} and {{formula:42e5b8ae-9172-414a-af64-2b3af6a52bc2}} , were quantized into {{formula:d42fb2bb-9a3e-47e0-bb21-6fac75464b69}} bins. A codebook {{formula:7381c81b-e069-4024-9252-3c9d83e502f0}} was built by assigning a class to each possible combination between the bins of each axis. As the probability distribution of the motion vectors obtained from deformable registration has an approximately-Gaussian shape, bins were selected near to the mean, standard deviation, minimum and maximum distribution values. Using this scheme, we effectively represent the motion observed in the dataset. The motion learning architecture is composed by a multi-scale (MS) encoder, recurrent units and a decoder, as illustrated in Figure REF (a). The MS encoder extracts feature representations at multiple scales through the network: fully, medium and low resolution in order to fully exploit the image features (see Figure REF (b)). The motion learning architecture also contains recurrent units and a fully convolutional spatial decoder. The spatio-temporal features extracted by the MS encoder are extrapolated in time by the convolutional Long Short-Term Memory (LSTM) units and further processed by the spatial decoder to recover the desired dimensions in the form of motion labels. We used a weighted cross entropy loss function as proposed by {{cite:f0ede60cef460fb93e55d5a0da9ecb2be7ca9967}} to promote class rebalancing since the distribution is strongly biased toward classes representing the superior-inferior motion. Adam optimizer with an initial learning rate of {{formula:16d5cfa9-b568-43c4-ab86-37638e7e4af9}} was used. This learning rate was reduced by 2 after 10 epochs without improvements in the validation set accuracy.
{{figure:3fae2248-79f9-4f62-b375-5dd5266d96be}} | m | 3353331ab3ac0399669d4aaf4f6d3d54 |
Two versions of the NN anomaly detection algorithms have been proposed: {{formula:b52a0067-d2c0-494d-b37d-05e67c1e4bb1}} {{cite:ffcab871559279a71924e2449d4405ee60c75ef1}} and {{formula:2a60c9ef-fde6-4460-beb3-adc958512466}} {{cite:ec6455ff9bb58f26cf4caec9cd0e65e294518f8a}}. {{formula:f4a4dab7-7c3a-4ff0-b4c7-b78f10b4be4d}} assigns anomaly score of an instance by computing the distance to its {{formula:e311a71f-ce9d-4a93-b232-e3a23a0384e6}} -nearest-neighbor, whereas {{formula:f5dd5e08-478a-4e1c-8caf-956fe95e0dc1}} takes the average distance over all {{formula:1c9dcf12-40f3-415f-b218-2de0acff184c}} -nearest-neighbors. Both methods are shown to have competitive performance in various comparative studies {{cite:619475bb45327fbed1b2445a13fc1ec151c4511b}}, {{cite:7c132f680546abdc614796dd6420ac50b84c7610}}, {{cite:d042124233f3d48864333af430329a822a943d27}}, {{cite:39fefbd55f90babd9b16f950eda6485c891b37f5}}. In particular, the comparative study developed by Goldstein and Uchida {{cite:619475bb45327fbed1b2445a13fc1ec151c4511b}} is the one of most comprehensive analysis to date that includes the discussion of NN-methods and, at the same time, aligns with the unsupervised anomaly detection setup. However, the authors omit the analysis of ensemble methods, some of which are considered as state-of-the-art algorithms (e.g., {{formula:e6f0d533-a910-4928-93d2-46af80ac43bf}} and {{formula:1232d095-5d1f-4bac-92bb-9cb781a9186b}} ). Emmott et al. {{cite:51c5b70f0765ee4f89a8586266a5cf6fbe0e7bab}} constructed a large corpus (over 20,000) of synthetic benchmark datasets that vary across multiple aspects (e.g., clusteredness, separability, difficulty, etc). The authors evaluate the performance of eight top-performing algorithms, including {{formula:a0235405-341a-4a33-ac49-4360dd0520f9}} and {{formula:371ddbaa-4109-4f3f-ac13-a40447189e4c}} , but omit the analysis of NN-methods. In this section, we provide a comprehensive empirical analysis of NN-methods by comparing {{formula:5e4aa164-f0b9-4364-8d1f-e74235a4ce44}} , {{formula:f995bb47-1c66-45e9-b860-1723726f7325}} , and {{formula:9a248635-1016-461c-97ed-353f23fb5d5a}}{{formula:587da840-c184-4569-b13a-6f9210e80148}} stands for the empirical DTM (see Section ) with {{formula:00e0ddfa-d3fa-47ee-a9d1-9170734a0f46}} . We include its empirical analysis here for comparison purposes. to {{formula:bb2cf200-2220-489b-8e89-1421fd279b67}} , {{formula:e689cd82-203d-4e04-a6ad-e76cd8f6e680}} and {{formula:57125598-aa44-49d5-800c-335d903a0fdb}} on (1) the corpus of synthetic datasets developed in {{cite:51c5b70f0765ee4f89a8586266a5cf6fbe0e7bab}}, (2) 23 real datasets from the ODDS library {{cite:9050ec5eca007f5f93d2ec03991e5a5ccc830106}}, and (3) 6 high dimensional datasets from the UCI library {{cite:f4589c2f5e640bb2d1765925259eef99b2a747e5}}. The code for all our experiments will be made publicly available. In general, no one methodology should be expected to performs well in all possible scenarios. In Section 4 of the Appendix we present different examples in which IForest, LODA, LOF and {{formula:e6fa551b-d73b-4f94-8070-cd617a6d55a0}} perform very differently.
| m | 45dacbeadac0009241c5557397196d90 |
In order to find the underlying reasons for the prediction models' results, we utilize two gradient-based attribution methods, Integrated Gradients (IG) {{cite:2d9100fb0fa65089fdadae16e94bed04d85a9aa5}} and SmoothGrad (SG) {{cite:4049e07d11b8413b9f3fe220a09e0239f26f375e}}, to interpret the event predictions by attributing the prediction results to the input features.
Specifically, IG fulfills this task by calculating the straight-line path-integral of the gradient from a baseline input {{formula:9da3c226-e77c-4933-a760-d2857f6cd517}} to the current input {{formula:5af98161-cc2b-4710-9c07-9b1f21a705f4}} , while SG performs attribution by averaging the gradients of a set of similar inputs generated by adding Gaussian noise to the original input.
However, there is a challenge in that the categorical features
cannot be logically divided or added with noise. Therefore, to apply these attribution methods in our task, we deliberately design an additional embedding layer that maps the categorical features into continuous representations.
| m | 05a765f1d30e57951e8328a48cb8a183 |
While Morcos et al.{{cite:0fca1d9dc1d82de7f2184204c0cbf804d1d0b591}} used direct correlations as pairwise interactions between residues, direct correlations (in liquid theory) are generally different from interactions. In fact, the approach of Morcos et al. may be interpreted as the mean-spherical approximation{{cite:25ad938d5dc4d82a2479a8225f3ad197c68af986}} which is a particular closure condition for solving the Ornstein-Zernike equation. It may be interesting to investigate other choices of closure conditions such as those analogous to, for example, the Percus-Yevick (PY) or hypernetted-chain (HNC) approximations{{cite:25ad938d5dc4d82a2479a8225f3ad197c68af986}}. The HMSA closure{{cite:7e54363910074adf390a09a6ade14ed0e7151266}} is another interesting possibility.
| d | f094aa5176caa1ea906c8778db4df95f |
In this paper we investigate the use of the recently introduced physics-informed neural networks (PINNs). In their seminal paper {{cite:8d40fb485882e30c1fbea532b12126c69167a0f8}} Raissi and co-authors introduce the idea of utilizing modern machine learning ideas as well as the computational frameworks such as TensorFlow {{cite:6acb1fc3885be4120ef6f0f6e65b2ad53cc75a86}} to solve differential equations. Many authors have extended the applicability of the PINN approach to different application areas {{cite:07539cbc5a10a9ad7d9fd6d4a64644ca976c20f1}}, {{cite:a32ac02cc12d25b595968fd81dcfd3a87bd07d07}}, {{cite:09a2b3ece33c63f7be7f3013e056b458825d2d43}}, {{cite:ae6d9384860d3ea6268620cd2664dd8fecdb9c14}}
including fluid dynamics {{cite:0527eeaa06da6bcaa1bbd4575ecfadb7054d65a5}}, {{cite:e8f370a1412b6ac9029c9358fec5761463503169}}, {{cite:49b787ac8319ec6b24ca726cb11d07760edcf3ae}}, {{cite:0330abd8507e1e80cfb53be88b01944c7db2eeaa}}, {{cite:3ae9d93cb802790eba81ff692ff58e506415dc88}}, continuum mechanics and elastodynamics {{cite:15171cb0f7014e175c81ba7e148d02eeab96a6c7}}, {{cite:7e0ccec2c14fbea83e4bff2c67189806bd7c174c}}, {{cite:961efa67f841b7116ad400c0d5326fd6286e03c6}},
inverse problems {{cite:d22c0f4647531b1651e420d871c05093c78f7159}}, {{cite:f1814ca98016757978da91e8db2ff924dd6b6db0}},
fractional advection-diffusion equations
{{cite:1ef3366027843efa5d12e1b5bcf51e7adefff3a0}},
stochastic advection-diffusion-reaction equations
{{cite:ecdb5fe4a15ece868f52b8d331879c5881c3970c}},
stochastic differential equations
{{cite:e260428ab00cb67d5ee78fd9e4e0c0636d376eff}} and
power systems
{{cite:580e2afa2d466ed3aad4561b5c134668d898c414}}.
XPINNs (eXtended PINNs) are introduced in {{cite:22b3a9393823823d95b2b62debaae481f08c875f}} as a generalization of PINNs involving multiple neural networks allowing for parallelization in space and time via domain decomposition, see also {{cite:e659c0a5aff264cedd27764a1d6cea313ed3fa5d}} for a recent review on machine learning approaches in domain decomposition.
| i | 7ce5b80d5ee730408741411cb564bbda |
Modern 3D camera technology allows us to capture 3D point cloud data more accessible than ever {{cite:914ea922c4a60db2c98b222dc39c306dd8b01aa5}}. Now, it is time to adapt 3D point cloud recognition models with LwF capabilities. We identify some key difficulties to address this problem. Firstly, in comparison to image datasets like ImageNet, very large-scale 3D point cloud datasets are not available. 3D datasets usually contain a handful number of classes and instances {{cite:c8c94c9565cae554ac23986f790320bcfff501ac}}, {{cite:40347da0049ec89c3f687fc292ca1483b414682d}}. Secondly, a typical pre-trained model for a 3D recognition system is not as robust as 2D models because of not being trained on a large dataset {{cite:9ed13ef09bd1b7ac683187f1c573ce69234a6f4b}}. Thirdly, 3D point cloud data (especially real scanned objects) contains more noise than 2D image data {{cite:40347da0049ec89c3f687fc292ca1483b414682d}}. This paper investigates how far a 3D point cloud recognition model can obtain LwF capabilities considering all difficulties mentioned above.
| i | f43da1d305c785df9b37f228aa2d951f |
The dominant approaches for ED involve deep neural networks to learn effective features for the input sentences, including separate models {{cite:639f74fe8b128a0bf6ee7ff29f08e491b7dbb58e}} and joint inference models with event argument prediction {{cite:24f0b64a480e6c77f7bd46ecd19189f1e1fe56fd}}. Among those deep neural networks, graph convolutional neural networks (GCN) {{cite:77c31f23a80cea9f57f79e40a0ccf41f744b85b8}} have achieved state-of-the-art performance due to the ability to exploit the syntactic dependency graph to learn effective representations for the words {{cite:97015095fcdc6f1ef88ce6ce6aa0b7af623f0cf2}}, {{cite:27d6cc0d863c1a9e8a669af28cb79cfbfb198d0b}}, {{cite:c8cc173bafa28582fb49b60b2e9740e113c356bd}}. However, two critical issues should be addressed to further improve the performance of such models.
| i | b7045a1179682af410cade6e4e509713 |
Agent: We train three agents {{formula:f64faf7b-fc6b-4002-aeaf-4ed8f96a8b0d}} , {{formula:ba74acc8-362a-4b6b-81d9-2282f35a3417}} and {{formula:8ae948e5-f770-4c40-a44d-352e7ced7520}} with reward functions {{formula:718aa22f-1f51-4a39-a638-b9e88581c838}} , {{formula:0b34c629-b052-4c23-bb40-d68f9bbb9c59}} , {{formula:3c84190d-4e41-44d9-8510-04bcd8db8dba}} respectively. The Q-network contains two hidden layers with 128 nodes each. For the environment, each episode is of length {{formula:8899b9bd-6786-4203-9194-14b1a4a91331}} , the discount rate {{formula:2049c53e-6b7a-42c2-854e-ffe731211ed9}} , {{formula:5fab7fcc-4d1a-43ed-a514-5e16b2dfadfe}} . The decay rate {{formula:8cee808f-89ef-4a87-a1fa-237385d8a30a}} is 0.000008 and 0.000004 for ECD and IEEE data respectively. We update the target network every 25 episodes and the length of the memory pool {{formula:c5c399fb-2cde-4607-b5f7-42c16826d4c1}} for the agent is 75000. The Q-network uses huber loss with Adam optimizer, and learning rate 0.005.
Deep neural networks (NN): The neural network is similar to the Q-network but is trained using binary cross-entropy loss. The optimizer used is Adam and the learning rate is 0.0002. We use the validation data for the early stopping of the training process.
CNN {{cite:621763ccc1d203672473d8d51eca0b5160059421}}: We applied 1-D CNN on the data represented as a feature matrix. The network consists of two convolution layers followed by a dense layer.
LSTM {{cite:8f47f73105c011274c124f379df2e2e7fc51b27f}}: Transaction sequences are first created using a rolling window with a window size of 30. A 2 layer stacked LSTM is used to capture the sequential information. This information is passed to a dense layer for the final prediction.
Random forest classifier (RF) {{cite:60d6cae36e7ff9a1f38bea737fe85f83fa827846}}: It is a popular choice for classification problems. We select the best parameters using a randomized grid search.
XGBoost {{cite:33157159b24f22bf84783235cd68320641ee23b7}}: Parameters for the XGBoost classifier are selected from a randomized grid search.
{{table:6ccb49ff-1e85-44ec-abe2-40d08a442381}}{{table:b619173a-a443-48e8-b3c4-0bb7b9c0d1fb}}{{table:3312a737-b3e4-4ddd-9ef3-731164b20cb9}}{{table:3f799a26-3693-4e7c-b4a5-0b9c43f488ef}} | m | 21597d6d6ea2ea028f70240d28556b76 |
In experiments, one type of driving force that is relatively easy
to engineer has the form of harmonic function {{cite:04fc14005ef683f0da757636788ebbb8fa77ad62}}, {{cite:f3880f5520ce1a9c4cfe12a50e5433837fde9b0f}}, which can be
introduced by shaking the lattice back and forth periodically {{cite:892896f20795319def9d9fef402f9ac22889dc55}}, {{cite:b005bd8c5ddc07b7d0dd17c88ae9ab7f46d2c448}}, {{cite:7a62f2ba3443749cdb13f00e4e56b1a8a132b02e}} and described
as {{formula:3c61d31c-f635-40ae-b092-3b96438cc385}} , where {{formula:c097bddf-eef4-4a8f-a01e-dffbb88c79ae}} is the driving amplitude and
{{formula:cbaa3434-4bea-4f92-8c70-d1d3a0e3c6c9}} is the driving frequency of the force. The factor {{formula:f87ab9d7-ef20-43f1-bfdc-03a6757ca830}} in the Peierls phase for such
a harmonic drive reads {{formula:2805cdfb-d893-4b46-a825-701e34d6f697}} according to Eq. (REF ). Working
out the the integrals in Eqs. (REF ) and ()
with the help of the Bessel function expansion {{formula:7b1f5291-69a6-4ac2-abd8-74c473150fd1}} ,
we find {{formula:7bc45b6e-b652-4124-b281-1a693f62248e}} , which is
derived previously in the study of dynamical localization {{cite:892896f20795319def9d9fef402f9ac22889dc55}}.
The Floquet effective Hamiltonian of the system under the high-frequency
harmonic forcing now takes the form
{{formula:5360d68d-9282-464c-abcc-7c6f2cf8c660}}
| m | dc2490d6206912b562631ec95d9da3be |
We proceed by induction.
By construction, for all {{formula:3221e0dc-abdb-42a3-b157-8a699798eb47}} , {{formula:703729e1-300b-4a78-9de8-a080603276fc}} and Equation REF holds when {{formula:9a09385d-04f2-4077-99c8-afc966835c73}} .
Fix {{formula:5bc5e011-6bc0-4f85-ab22-0f0fe5ac5a96}} and suppose that for all {{formula:be256c74-6e5f-42c3-ba73-0f1b591fdc79}} , {{formula:0488107a-f713-4e5f-802c-b10ab34727a2}}
and Equation REF holds.
Let {{formula:05f7b7ad-259f-4d1b-ad0e-ee781c87ea26}} and {{formula:c46d0f69-fedc-4cbc-8465-b8ecf9817d45}} . Define {{formula:fdb4d83e-d757-44e0-b758-2c375d943234}} where the 1 is in the {{formula:51140fbe-968e-41e3-b0e4-0057eee0afd2}} th position.
Choose {{formula:03cb76c3-de39-439d-b939-f1b628ad3122}} such that {{formula:c6eab1f0-1791-42e7-aff9-2d0e72b8af72}} for all {{formula:dbd1ec12-988e-4f89-80c6-9d005d613861}} .
Define {{formula:c278a3f6-abee-44e0-b3f2-2ea673fa04da}} and {{formula:7882478e-62df-459b-aaac-eba911c1fe97}} for {{formula:e4e3d1cf-1721-496e-9e78-2eb162a26055}} , {{formula:d5b033c9-b4bc-40c1-a99f-efaa33ba1668}} .
Note that {{formula:8e8ac3f8-c93d-4639-84c3-ced1ee5d8a09}} for all {{formula:f6c3bf99-7c1a-4972-9333-9427747d0d5e}} by the induction hypothesis.
Define {{formula:318d6934-09fb-460f-b5a6-d7e459e0a592}} .
It is straightforward to verify that {{formula:98009795-ecc8-4f73-bb62-218cfc41f51f}} for all {{formula:599cac89-f42f-4e4f-9578-901180ba08ed}} , {{formula:1a085387-0d63-464c-b375-641224b65cf0}} ,
by using the inequality {{formula:189be063-e0a4-4c6f-bdb1-a4ec3248fa52}} .
Further, {{formula:f56f34da-d463-4dbc-929d-9b3f8485a4ae}} by the induction hypothesis.
Therefore, {{formula:2c1392b7-8ca1-4d85-bd90-f8e0037fe221}} is differentiable and {{formula:7661bb5a-94a9-4503-96b3-e8bed295147c}} for all {{formula:e3e45436-e6fd-42db-b7df-f3932e54d3d8}}
by {{cite:3ed0047e869c51ca149e271e083f076684ff7fd1}}.
| r | 727409e4061f56dff525ad302e981e9c |
Continuous-time dynamic graphs (CTDGs) can be viewed as a set of observations/events {{cite:6fe92d2d45a0306e9e6f4ed19e0bd0cd053889bd}}, and the network evolution information is retained. There are only a few works on CTDG. But recently, more attention has been paid to continuous-time graphs. All three representations of CTDG are described in more detail below.
| m | 2ced82b26622965cfb072d4de48de2f7 |
We used the recurrence analysis method to study the non-linear behaviour of several X-ray sources mentioned in
{{cite:422f2baa4269823c289360567224433e609f35fc}}.
Our present analysis confirmed that the
variability in these sources is significantly governed by the nonlinear dynamics of accretion process.
We confirm that the GRS 1915 and IGR J17091, which show deterministic chaos, and are not the only sources
with that feature.
| d | f749ce97078ef2fa42553167bdfaf5f2 |
Haze removal {{cite:b02ddb9174f9534cc67bf5099d1ef2fd3410b102}} is a classical ill-posed image restoration problem, which plays an important role in intelligent transportation systems, e.g., object detection under haze conditions {{cite:d832afce42216d8164b602bceb2fec1b15a3f855}}, {{cite:f4c3691347ac5f50b4bcf6d5645ffbc821d661c1}}, {{cite:7d5364821f147566fdc85a6d5cc5f8c90c02ee02}}. Haze is defined as some particles such as dust that obscure the clarity of the atmosphere. Dehazing is to remove the veil of haze from a haze image and restore a corresponding haze-free image. In recent years, because the development of deep learning technology has greatly improved the performance of image processing compared with non-learning-based technology, the problem of dehazing attracts more and more attentions in image restoration research community. Various image dehazing methods based on deep learning technology have been proposed, including: (1) Generating medium transmission map {{cite:77aeb099fcf00c2e4ddddc8b72eee3acd45bcdac}} or haze-free image {{cite:f4c3691347ac5f50b4bcf6d5645ffbc821d661c1}}, {{cite:89d7f7b563a34463827a346454591e1516e13ef8}}, {{cite:0f6f9dfafd35857d7cacd0440c11c211e4ce9013}} by a convolutional neural network (CNN); (2) generating transmission map {{cite:e4b447576c4d093b05b0b5df3eb4e20801fce4ce}} or haze-free image {{cite:b7cc48470923d9339de5470eec4e434ec098ca1f}}, {{cite:1c9bb410b219889bb75fbb2475ee87ff683a326f}}, {{cite:0c11fcc24b40d3803aafc30c225717510f1d9a19}} based on encoder-decoder structure without adversary training; (3) reconstructing haze-free image based on generative adversary network (GAN) {{cite:617b0114e5a150b6543db95e41c974ec2dacfd19}}, {{cite:d3198bcbeecdaf0ef0352907ee2452b22ecea191}}, {{cite:708550522d595069c1daefad3be5ed48af0bb8af}}, {{cite:dcf4b899d4871acaf96235ba1b2af8bcc6521e3f}}, {{cite:2ddffdd7d1e3f10cb4ab899132c5db9412278558}}, which are paired image-to-image translation models; (4) reconstructing haze-free image based on cycle GAN (CGAN) {{cite:2cc893bc0d99d7579b51ef628c6fe7e046d8deeb}}, {{cite:919e1d697ee50f79eb1306f425f65d4e5b68475d}}, {{cite:f082433b89de27891fcc2d5c0e98fc4eb39bc69a}}, which are unpaired image-to-image translation models.
| i | 8ec6a2888641c923459204b33590a4d9 |
Electrocardiography (ECG) reads out a spatial map of the time-varying electrical potentials of the heart acquired using electrodes placed at specific locations on the surface of the body. Interpretation of the ECG unveils structural and functional abnormalities of the heart that can aid the noninvasive diagnosis of cardiovascular diseases {{cite:3b65a92c034102fb777d0a998af81aff546a4988}}, {{cite:a7148d1babec3f9dfdceb7618cb20163590c03ce}}, {{cite:4470dcaf488544a90b644e29adab97007cbb5999}}. Importantly, ECG is the most important diagnostic tool for arrhythmia detection {{cite:3458fb3e4f08b3eb1f4069ff27b72235f30ec271}}, {{cite:66c7e4518f3cb73f4a913cd0ac3623c383e6fbc3}}, {{cite:36d3f9c7ec140eb29349f56b8cb8c8d89b84c105}}, {{cite:93613c47587642f6d594b00b0706c8319ddf59d7}}, {{cite:85324b1c4c111f3f2b5d3ee4a43d3dd158c9b7cd}}. As the abnormal heart beats often occur sporadically and are not present at all times, ECG recordings may have to be carried out repeatedly or continuously over an extended period of time, e.g., days with ambulatory Holter devices {{cite:a7148d1babec3f9dfdceb7618cb20163590c03ce}}. Due to the high signal data volume, manual interpretation is time-consuming and susceptible to fatigue-induced error. This has spurred the introduction of automated computer-aided diagnostic systems, which may be based on machine learning. Some machine learning techniques are able to evaluate individual heartbeat signals on ECG records {{cite:2210884c71663fcedc7d230453ece8abc7d15db8}} to complete tasks like classification, localization and prediction.
Among the many explored applications of machine learning for ECG signal analysis, two general problems stand out: regression and classification. Regression is a quantitative prediction task that maps the input data into output consisting of real or continuous values. For ECG data, regression problem can take various forms, including segmentation method for detecting ECG P, Q, R, S, and T waves {{cite:121be436555c5efc04042857a4148138c654b80a}}; reference method for removing noise artifacts from ECG signals {{cite:6c25ea23d50cdb2cfed00758656e2fe6ca329752}}; and increasing the spatial resolution of ECG through lead prediction {{cite:7a8109a4a8475d30a89f8384c91869f154532119}}. On the other hand, classification is a predictive technique that maps the input data to output data (targets, classes or categories) to arrive at the correct class labels to which the input should belong. Examples of works published on ECG dataset classification include labeling heartbeats into one of the five beat classes according to the ANSI/AAMI EC57:1998 standard {{cite:cf4269304fe8f260393ffee921a7b7fb4a87908a}}; classification ECG segments into normal and multiple arrhythmia classes {{cite:a722e5233dd4e9899366fabcb34fabd1ed9e266f}}; and classification of myocardial infarction {{cite:042926f95ab4c2ad4ae7243d95a5994603b94f89}}. In many situations encountered in automated analyses, regression and classification tasks are intertwined, as the former can be used to enhance the performance of the latter, e.g., by mitigating ECG degradation from noise and artifacts as well as missing data from low sampling frequency {{cite:dad2d93790d6453058982e83a57601d08789f426}}.
In this article, we present a novel approach of ECG signal modeling that uses a convolutional neural network (CNN) for both regression and classification. One of the advantages of our method is the flexibility of neural networks, making it possible to adapt a single neural network architecture for multiple tasks. We have exploited this trait to develop a single neural network model that is capable of modeling large parts of the input ECG signal as well as classifying the same data. Moreover, unlike classification, which requires training the model on expert-annotated ECG signal data, the model can accomplish the regression task without the need for data labeling. What is also worth noting, the knowledge gained from the self-supervised regression task can be seamlessly transferred to the downstream classification task. This two-pronged approach offers optionality that may improve diagnostic classification at little additional computational cost. On the other hand, the disadvantage of our approach is the relatively complex method with a lot of hyperparameters that needs tuning, which can require a lot of time and computational resources to optimize. In summary, the novelty of our work is as follows:
| i | cf7ea471c5dda428967314f32a732073 |
In this section, we compare the visual results of our proposed MDDM
with the above mentioned demoiréing methods. The visual results
are shown in Figure REF . The red square
means zoom-in of the image such that we can compare the details of
the results. From left to right are the moiré image, results of
DnCNN {{cite:33d3f0cdcb583179ceae70a81e9774d4fa4881af}}, MSFE {{cite:fe0cdb1f0db9e6d0abfdebe4d7e75982f468640d}},
Sun {{cite:18b01808e1e07fc14a15737a7d7f6fc797244ed5}}, our proposed MDDM and the ground truth. PSNR
and SSIM are shown under the image. The visual results show that our
proposed MDDM is significantly better than other methods. Our MDDM
removes moiré patterns more cleanly, while other methods bring
more artifacts or can't completely remove moiré patterns.
{{table:1d2b59e8-d1f6-4867-9e9d-6af7ccfda399}}{{table:8e3e4c00-9cf2-48b6-9747-5a110c817e31}} | r | d9efb0ff6d488ab98bbe70bbe95dd39f |
We have shown how gapless fermions modes bound to defects or solitons in various dimensions may be detected by computing the index of the Euclidean Dirac operator in the presence of additional background fields. The method involves determining the divergence of a generalized Hall current via a 1-loop Feynman integral, which calculates a topological winding number of the fermion propagator, the field theoretic generalization {{cite:ff6275dba078156a97a3f5caf84182d7f7907d65}}, {{cite:0363b298f236021b0b32430ac4d4fba928a8c62b}} of the TKNN result {{cite:10860b13fbf094d3f27b4d3c9abd1b7fb7a6da9e}}. These currents can be computed for systems without chiral symmetries or anomalies, and generalize the concept of anomaly inflow. It remains to be seen how comprehensive our approach is, whether it can be applied to theories with interactions, and whether this generalized Hall current has any experimental implications in Minkowski spacetime.
| d | ace3b6c1e02a2cc77b2b6c58cfb23559 |
Among the most popular visual grounding tasks is referring expression comprehension (REC), which localizes an object given a referring text {{cite:8cf1160a9f1877fbc2870f128ad21bfb9dc5c1ee}}, {{cite:71b45c73988f615ca2ac6364910d5a2107523400}}, {{cite:85f536630930ea0f155030bdfd4ebe3e842f8a0b}}. This task often requires complex reasoning on prominent objects. A highly related semantic localization task is object detection (DET), which seeks to detect all objects from a predefined set of classes without text inputs {{cite:9745abf5c3a94837f8f7c3556fbe811e98891e68}}, {{cite:a78d2ed8b16c1eeface2bbe552c248cb7420cb3c}}, {{cite:98a491054b17399d71f099ac97761cb73dff3a03}}, {{cite:104fd8b81b7d422ce11d8e6c8781241b362f4f93}}, {{cite:06f980dfcbc9a1bef8f165d8f47451d11ab18bfd}}, {{cite:e51b728d611e99228e594a1f5f86f495e5c8fb91}}. In contrast to REC, this task requires the accurate classification and localization of small, occluded objects. At the intersection of the two is text-based localization {{cite:fde28c2f4a042b69ddff9267b10837fc17a492e1}}, {{cite:6f3885f675b8a44a081acec7c8d49820c0346383}} (LOC), in which a simple category-based text query prompts the model to detect the objects of interest.
| i | b3edf54434af292571f0799feb05d1fe |
The Jacobian conjecture originates from the problem posed by Keller in {{cite:f0cf96ddf35e4f95dd22c7107b849d4383b00272}}. It is the 16th problem in Stephen Smale's list of mathematical problems for the twenty-first century (cf. {{cite:efaa9096e56fb5a335833ba6a4118275b6d63e43}}). Let us recall the precise statement of the Jacobian conjecture.
| i | 6d81971cb1db1175c266700c9423114a |
In the domain of medical image analysis, Pham et al. {{cite:e3e95e780791e5aed275f19e237e4703e59b35ae}} proposed to remap targets to random numbers close to one, finding that this method improved model performance on the CheXpert Dataset {{cite:6fe90101ed27937b46ffe2649ad670309f350fd7}} by approximately 1.4%.
Moreover, Xi et al. {{cite:ec9bbf15514964fb7f195d39779759e359a31d4b}} addressed uncertainties of correctness using a spatial label smoothing technique to reduce the need for well-annotated data while still achieving satisfactory performance.
| d | b0b33bc8b5ba466ea0d8c17bab2ffd94 |
Now, with reasonable descriptions of low-lying meson masses at {{formula:cd24f5f6-3b9d-4941-90a5-8f3bcd50445d}} ,
we examine the low temperature thermodynamics.
Shown in Fig.REF are the entropy densities of a neutral meson gas for various {{formula:1a027a88-9ea9-4a55-bd48-211e4f7185d4}} and {{formula:e74a7a3c-f2d7-421f-9891-3d2a48b5a130}} .
We plot the results for the neutral mesons ({{formula:f3dff16d-99ed-46a4-93d2-43ebc4823959}} ), neutral plus charged mesons ({{formula:06e2c6e3-d779-46a0-aab5-088ebb1a5c0e}} ) with relativistic corrections,
and {{formula:9dbcb78a-1943-40cf-9ac4-ffd8338eabc7}} within pure non-relativistic treatments.
They are compared with the lattice results in Ref.{{cite:73f7c0a5ad81b2384684da746a978c4d42162491}}.
As a guideline we also plot the HRG result at {{formula:21a56f77-8a6d-4328-9e3e-a4c2065fb2e8}} which is based on the PDG list for mesonic and baryonic
spectraIn
Ref.{{cite:c5e6ec71f3517d58eda41dee7473ee8e76b1b721}}, the author computed the PDG based HRG entropy at finite {{formula:551663f9-1f67-477f-9a02-cdf7aca19122}} , regarding hadrons as elementary particles.
The resulting entropy density to {{formula:2da50693-22b2-4675-a0e9-4e5fe44d160a}} MeV is found to be very close to the {{formula:a9085a10-4704-4da6-abc6-ef466658c564}} ..
As we have mentioned before, our HRG includes only mesons and the resulting entropy should be smaller than in the lattice.
The baryon masses are {{formula:0a747e65-ae4e-4a60-bb19-c3e022558dc6}} GeV, so we expect the corrections become substantial for {{formula:90694c92-7cc3-4295-9609-f38e2cb23669}} MeV.
| r | 4fdbfc24e9862e5fb2ebe28b0acd5821 |
While none of the automated approaches evaluated comes sufficiently close to HYPE-Style for standalone use, our work still constitutes an initial foray into evaluating style-level attributes of multimodal cross-domain mapping, an area where it remains difficult to use mainstream automated evaluation metrics out of the box. As future work, we plan to both improve the realism of our generative model and explore better methods for evaluation. For instance, the performance of the pre-auxiliary classifier embeddings suggest that we are operating outside the domain of ImageNet, and from this insight, we are inclined to leverage other embedding spaces, e.g. the Mapillary or Cityscapes datasets {{cite:afda1cd2f2b0185892b405915cdca75417bd358b}}, {{cite:e7e4fd39ef58ea10da5c4b80741b5ad765f56801}}, which could provide more suitable street-level scenery features that is similar to ours. The ultimate vision of this work is to create an ML architecture which, given an image from Google StreetView {{cite:d132ad9e773d36e07a4fba818eb698f60da4b44c}} based on a user-chosen location, is able to generate the most realistic image of climate-change induced extreme weather phenomena, given the contextual characteristics of that given image. While representing flooding realistically is the first step to achieve this goal, we later aim to represent other catastrophic events that are being bolstered by climate change (e.g. tropical cyclones or wildfires) using a similar approach.
| d | 696803d0c41f2e440babe9773c05615d |
2) Visual Comparison: Figure REF shows the estimated disparities and corresponding BadPix0.07 maps. Since the proposed OACC can handle occlusions in a fine-grained manner, our OACC-Net performs well on scenes with heavy and complex occlusions (e.g., the nested structures in scene boxes). Besides, our method is also robust to noise and outperforms many state-of-the-art methods {{cite:36cf0fb7e0d16c0b893f6eb844a08ffdf215c808}}, {{cite:725adac0e3293a56c96cf7e95bc5b61843120824}}, {{cite:411f80631cf291a002023e81a1b66f2e2df5a704}}, {{cite:d8c6ef0b5c2edb720d5c5c085bca9d8b1dde1ca2}} on scenes with large noise (e.g., the bottom dots in scene dots).
| m | d8de3d295cc081baf2eacd4a0073b278 |
In this chapter we address these issues from a different point of view, inspired by the landscape of string theory vacua. We consider a large number {{formula:8991753b-5bdc-46eb-9802-69025a77b045}} of sectors contributing to supersymmetry breaking. Large number of sequestered
hidden sectors have also been considered recently in {{cite:b568f60c694be5e610ae137459aa8b05105da0c7}}, {{cite:624f8437a100533e2dc7b2b906686ba0a1e0c10f}}, {{cite:596ba174cd473dd6fc0cbb0119a62d9968e5a693}}, in models with multiple (pseudo)goldstini. Other works which have addressed supersymmetric soft spectrum phenomenology from the landscape following {{cite:98f387e1f9b11c4eda6fa042e3acef0c98c4e6de}} include {{cite:90d2978745ab0c22e9c735cb92235d82eb4d0026}}, {{cite:87cf60dab04d1edd3932dba07e5c479c88491494}}, {{cite:b72aae2fec3a95fce90505974f9af90d2861303f}}, {{cite:d40b8dfe83d44baefc2ef59ea6f9c0ab4cba9e24}}.
In particular a solution to flavor and CP problems in the landscape through heavy first two generations was
proposed in {{cite:8ddf6f26f5ad01e955152e546e0fc04155e34b79}}.
| i | 5b0cf55103de6bda52fdb0e782978a79 |
Image patches that have similar pattern can be spatially far from each other and thus can be collected in the whole image. This so-called nonlocal self-similarity (NSS) prior is the most outstanding priors for image restoration. The seminal work of nonlocal means (NLM) {{cite:8b5c6e4a57b2f662797d71e884443ead3867eac4}} exploited the NSS prior to perform a series of the weighted filtering for image denoising. Due to its effectiveness, a large amount of related developments have been proposed {{cite:49b4f766a0bbc97e16484274a08f9a17ccd393c2}}, {{cite:308ae6c0c01107ddabdd8075bad7b858bd2f86ca}}, {{cite:285adc346e542890304cf4917f6f7d7273e7447d}}, {{cite:3d573d059ec9a7d5feff43a90f1a722350d68d34}}, {{cite:62954aea0aaf49c1c0b8e7ef4ffb75b8aecfa011}}, {{cite:340eca32385a0eefa0bcaf7f73715a81c54cd5b7}}, {{cite:39c2dedbdd18b3168084bb7baf35d079b205f733}}. For instance, BM3D {{cite:340eca32385a0eefa0bcaf7f73715a81c54cd5b7}} exploited nonlocal similar 2D image patches and 3D transform domain collaborative filtering. Marial {{formula:9a9670bb-eb5f-4837-9b3b-b35d4f2befd5}} {{cite:49b4f766a0bbc97e16484274a08f9a17ccd393c2}} considered the idea of NSS by simultaneous sparse coding (SSC). Dong {{formula:d7b53531-cbd2-4f41-93a3-bdf2f19f2bdc}} {{cite:3d573d059ec9a7d5feff43a90f1a722350d68d34}} proposed the nonlocally centralized sparse representation (NCSR) model for image restoration, which obtained the estimation of the sparse coding coefficients of the original image by the principle of NLM {{cite:8b5c6e4a57b2f662797d71e884443ead3867eac4}}, and then according to those estimates, NCSR, centralized the sparse coding coefficients of the observed image to improve the restoration performance. Zhang {{formula:a0c80110-aea3-426c-a37a-5c795604086e}} {{cite:39c2dedbdd18b3168084bb7baf35d079b205f733}} proposed a group-based sparse representation framework for image restoration.
| i | 2c3dd2cb91d38006835352250f5d7c98 |
CTIN was implemented in Pytorch 1.7.1 {{cite:4fd52ca000bd86bb09cbca60b13e3e767f4cf94e}} and trained using Adam optimizer {{cite:94506423cd6a2828a95223f22c0ff63ee5976641}} on NVIDIA RTX 2080Ti GPU. During training, we used an initial learning rate of 0.0005, a weight decay value of {{formula:2e02d1de-07b7-4860-abd8-3aa74d0ccb49}} , and dropouts with a probability of 0.5 for networks in Spatial Encoder and 0.05 for networks in Temporal Decoder. Furthermore, early stopping with 30 patience {{cite:c3bd59f00cfa173f56235247c8b7d4bd595a9408}} is leveraged to avoid overfitting according to model performance on the validation dataset. The extra experimental results and analysis are listed as follows:
| r | 83783c2a6cc59857d2dde33507648391 |
BEiT: Since ImageNet-1k pre-trained weights are not available, we use the official BEiT code release {{cite:1cf8c6642c79d0b7f60a520968c86aae30d23096}} to train ViT-B and ViT-L ourselves for 800 epochs (the default training length used in {{cite:1cf8c6642c79d0b7f60a520968c86aae30d23096}}) on unsupervised ImageNet-1k.
| m | 8e902caa4bb8e7bbe09c98f82d4af1a4 |
and the conclusion that {{formula:2b83ec8e-0640-4771-90d3-6f54fa557a8a}} with {{formula:4f462946-3110-4714-a612-18dfc4c26937}} , see {{cite:da6b89d1de26286b0ad31854113d6fe1dd2dc366}}, we get
{{formula:60fab868-dfab-4dc4-9119-a5f6c21b5f34}}
| r | 4702833b67bcca62cc701e9ef557e625 |
In Ref. {{cite:5522382d1b6e7ac9749027ea4f7f5f62298c8863}}, the process of {{formula:32a1c720-8113-44d3-b830-ca70da863c3e}} is recommended as a discovery channel of {{formula:e4f7ea15-c24c-4249-b007-c2b2843e2155}} , which has been confirmed by the LHCb experiment in Ref. {{cite:d824d5c6cd4a46ba2eb76e2ae08ad259cf1cb355}}.
Similarly here, we suggest the LHCb collaboration to search for the bottom-charm baryons by the modes of {{formula:33ef8aea-e261-446f-8d57-a2fc22655545}} , {{formula:47fae6d9-dfc6-46b4-9154-c236a5b56c87}}
and {{formula:d8b05a34-de46-42be-bd05-0ccc75cc832c}} , respectively, who have the relatively larger branching ratios.
The branching ratios of those three decays are given as:
{{formula:6d171e39-41e9-4915-85b0-0dce125e4240}}
| r | adb9c2b7c662d3d305473e115581d36b |
For the fit of the benchmark simulation, we let the sampler run 50,000 discarded burn-in steps followed by 50,000 sampling steps in 15 different chains using the Metropolis-Hastings algorithm {{cite:0697710fd7866684118c495a680aa9798c455112}}, {{cite:1c713dd3a4d609187175f6f95da757defe291cc3}}. Convergence is ensured by calculating the Gelman-Rubin coefficient {{formula:913a7905-0a20-44c0-bb81-ccbc633d38fb}}{{cite:930c5de9596603e3418439fb6d4a5de72f254351}}. We use the same flat bounded prior distributions for the source parameters as in {{cite:bf362e16338265d8192d7aec39ae8bd33d224f53}}. Each chain runs for around 4-6 hours on a CPU.
| m | 28f9ebbc24e3b525a547716a2c16831b |
The present work has several perspectives and possible extensions.
First, it would be interesting to re-derive the approximation
(REF ) in a more rigorous way and/or by a direct analysis
of the eigenvalue problem, e.g., by matched asymptotic methods. In
fact, our derivation involved three approximations, and it was
difficult to control the accuracy and relevance of each step. Second,
one can deal with multiple small targets. If the sizes of targets are
much smaller than the distances between them and from the outer
reflecting boundary, the approximation (REF ) is expected
to hold. Note that the capacity of the union of small targets is
equal, in the leading order, to the sum of their capacities; the
surface area is also additive. Moreover, Cheviakov and Ward derived
the next-order correction term to the principal eigenvalue for a
configuration of perfect targets {{cite:cfaa82f23efcd9aad6a3738240d798a64ab824b6}}. This correction
term can be used to define the “corrected” capacity {{formula:353117b0-6199-4ea9-82e4-0fd1b6dae0a1}} , as we did
in Eq. (REF ) for a single target. A numerical validation
of this approximation in configurations with multiple targets presents
an important perspective. Finally, one can investigate other surface
reaction mechanisms (beyond the conventional Robin boundary condition)
by using an encounter-based approach
{{cite:ad13e26b0b5efba5a183598ace81f4577ebcee3f}}, {{cite:537f13e98ae8ea49e26aa4f7d9aaedfb87438c05}}, {{cite:9777cb018fce237db27688fd279d4888150331dc}}, {{cite:8d07e0a0c22c7193050a9a8c0597d53212345e24}}. Here, the
explicit dependence of the reactivity parameter {{formula:3098b3ed-011f-41e9-9b24-d56e5bdade90}} may allow to
access various properties of diffusion-mediated surface phenomena.
| d | db9049c78752a669824fe49af119e9ab |
A variety of distributed planning and decision-making problems, including multiplayer games, search and rescue, and infrastructure monitoring, can be modeled as Multi-agent Markov Decision Processes (MMDPs). In such processes, the state transitions and rewards are determined by the joint actions of all of the agents. While there is a substantial body of work on computing such optimal joint policies {{cite:7a76513b55cb33b5b13f8b1db392cb4df277f775}}, {{cite:54ae59c9296426240d56aeba911f7fe3cf0238cf}}, {{cite:0b4f0a741343e1fb00f3c41e3b10f13cc18d56d7}}, a key challenge is that the the total number of states and actions grows exponentially in the number of agents. This increases the complexity of computing an optimal policy, as well as storing and implementing the policy on the agents.
| i | 9eced68c61960b9fdb11648674e92258 |
large visual spatial contexts combined with efficient iterative patch-based training and dense inference. The output of the CNN is often interpeted as the parameter of a conditional distribution. For instance, in {{cite:e65b3902f9a5612182efe0b774f9ab4365f8a8a1}}, {{cite:29f4160fed4ee7e0d26a97c2b63384ac9445b5aa}}, the output at each voxel is the parameter of a Bernoulli conditional distribution. The Poisson distribution is generally well known for modelling counts over time and space, and particularly has been applied to modelling the count of multiple sclerosis lesions over time {{cite:715abd64be517a443fa40a282140dbe9f83224bc}}, {{cite:cdcd53362edc0ea6f5a6f0a626ed08fa596906b5}}. For this reason, we propose the lesion label counts, or equivalently lesion volume, in a predefined patch size is assumed to follow a Poisson distribution conditional on the patch features. The CNNs of {{cite:e65b3902f9a5612182efe0b774f9ab4365f8a8a1}} and {{cite:29f4160fed4ee7e0d26a97c2b63384ac9445b5aa}} use hundreds of thousands of parameters for segmentation. Using CNNs, coupled with good distributional assumptions, should allow for smaller architectures and faster convergence on the counting task.
| i | 27e0b94ac99163fb3e6bec15ba0d793a |
The model is interpretable, and thus admits the use of
prior knowledge, e.g. if we already know some things about an
object. The formulation is also composable in that models
for individual objects can be learned separately, then combined
together at inference time.
The variational inference algorithm is obtained directly
from a generative model for the observations {{formula:9e5176e9-20d1-4fd1-8888-fb2d86609c92}} . In contrast
other leading formulations set up an objective to produce clusters in
{{formula:d7834b66-e195-475a-a149-3a95946725fa}} -space.
The interpretable structure of the GCM allows other inference
methods to be used, as demonstrated by our use of RANSAC.
The GCM conforms to the view, as promoted in
{{cite:e933f6ec2e8cb4c5aa2ba66de29c051d64e590cf}}, that the input is regarded as a
set of parts. This formulation ensures that if the parts can
be detected equivariantly, then the inferences for the objects will
also be equivariant. This was demonstrated in the constellations
and faces experiments.
| d | 590ca6d3f4c9e064ca734d24a13c4abe |
In this paper, for Korean-specific table question answering task, we present KO-TaBERT, a new approach to train BERT-based models that learn jointly textual and structured tabular data by converting table structures. To address this, we firstly create two datasets written in Korean language: the tabular dataset contains conversion formats of around 1.4M tables extracted from Korean Wikipedia documents for pre-training language models, and the table question answering dataset for fine-tuning the models. The table question answering dataset consists of 70k pairs of questions and answers, and the questions are generated by crowdsourced workers considering question difficulty. Additionally we introduce how structured tables are converted into sentence formats. The conversion formats play a crucial role for models to learn table structural information effectively without changing embeddings. Second, we follow BERT architecture {{cite:f46b4e52e47144626f1298fa6d0b91deace7a6b7}} to pre-train a language model with the converted strings from millions of tables, and fine-tune models on the table question answering dataset. All resources we create in this study are released via our GitHub repositoryhttps://github.com/LG-NLP/KO_TableQA.
| i | f62894678b83ef958728cb1d445bd3bb |
The study of products of random matrices has been proposed many decades ago by Bellman {{cite:436b30e2a565f78914af0d3a9fe48cfa642e5271}} and by Furstenberg and Kesten {{cite:dd16c23eb5b2a37ce9823b05721ae49449367a60}}. The motivation was to understand properties of the Lyapunov exponents {{cite:dadcaf727684727b016622a53ae8b837d7b349e6}}, {{cite:362e55a8e978943e21faa25b8220d4f255ae478a}} in this toy model for chaotic dynamical systems, where usually the matrix dimension {{formula:99bfc78b-bf7e-4551-9ba7-88c1e8db9532}} is kept fixed and the number of factors {{formula:8be20a1d-19e8-47d0-9cfe-eb057ed09087}} tends to infinity, see {{cite:653ac28109c1defde0c0ac920d48397e82df9373}} for a recent account. The factors are typically taken to be real non-symmetric or complex non-Hermitian matrices with Gaussian distribution of elements, from the real or complex Ginibre ensemble {{cite:47d8a903333a80666156450886520c0af2b3c8fe}}. While Ginibre himself showed the integrability of a single complex ensemble as a determinantal point process, it has taken many joint efforts to put the real Ginibre ensemble on the same footing as a Pfaffian point process, and we refer to {{cite:f37930c0604b43cdf79190f641f703b6b349d3a7}} and references therein for details.
| i | 6fc85ef4d8d8d9da2aa57344d28bfec3 |
The skeletons are a good privacy-protecting source of information about human posture. However, the quality of body joints approximation depends upon the resolution of video frames and the degree of occlusion due to objects or people in the scene {{cite:42abe659391b8c752678f34f1c505ecd1d5cc32d}}. Occluding the appearance of the people using semantic segmentation masks is another way to preserve the privacy of the individuals in a video frame. Similar to the skeleton-based approach, it could remove a person’s identity while maintaining the global context of the scene. Jiawei et al. {{cite:42abe659391b8c752678f34f1c505ecd1d5cc32d}} showed that it is possible to occlude the target-related information in video frames without compromising the overall performance of human action recognition. They suggested that a model trained for human action recognition can be used to extract features for anomaly detection; however, they did not show any results on the anomaly detection task in their paper. Bidstrup et al. {{cite:7ed6ecdd8279c8809a0c4edf9ca1028bcc7ae861}} investigated the use of semantic segmentation to maintain anonymity in video anomaly detection by transforming the individual pixels in a video frame into semantic groups. Their paper was centered around finding the best pretrained model for transforming individual pixels into semantic groups for UCHK Avenue anomaly detection dataset {{cite:dfa4b14db17cc4f3118c53fba5071528b3e459e5}}. However, due to factors like view angle, color scheme, and objects in the scene, it is not clear to obtain a pretrained model that can satisfactorily transform all the pixels in a RGB frame into semantic groups for any given video dataset. Hence, in this paper, we only transform the RGB pixels for the people in the scene into semantic masks to achieve the anonymity of the individuals. When training anomaly detection methods to derive global patterns from singular pixels in RGB space, the presence of semantic boundary instead of pixels for the individuals in the scene could remove unwanted noise related to the appearance of the individuals and help the models focus on the behaviour of the individuals.
| m | 1835b61cb89eedd8636368ecc2e3b6ab |
where {{formula:6b77027d-494d-49a9-ad74-8708d9393c25}} is the learning rate and {{formula:40be0c84-032c-4392-b3ab-e5db9ad4e106}} and {{formula:4ed0bb28-4906-4ef7-a491-cd64a693c80a}} are positive scalars denoting the portions of the speaker and age group adversarial tasks’ gradients backpropagated to update the generator’s parameters. This idea was achieved using the Gradient Reversal Layer (GRL) proposed by Ganin {{cite:f3d845e39903408575e95defbf8e525e53df36c9}}. As shown in Figure REF , the GRL layer works as an identity-mapping in the forward path while in the backpropagation path it reverses and scales gradient by multiplying it with {{formula:90054015-5f1a-4695-94f5-d06fba07e284}}
| m | 1ed3d8426a03652597ad5533ceb28dc5 |
Our above remarks are of course only speculative, and other
possibilities remain. For example, one may have different product
structures that are possible in twistor space (corresponding to
different ways of picking cohomology representatives), but such that
these correspond to different double copies in position
space. In such a case, one could formally define the notion of a
(non-unique) double copy in twistor space by giving (i) a method
for choosing cohomology representatives for scalar, gauge and gravity
fields; (ii) a product formula (or other map) for combining the chosen
representatives. It may then turn out to be the case that only one of
these definitions matches the original double copy for amplitudes, but
the remaining double copies may nevertheless be useful for
something. The relationship between the twistor double copy of
refs. {{cite:6c8b5360e7ce201795ddfeca1a5f0606690a3c21}}, {{cite:6fe0b53c0c3314e1e455a17e9a9ee4c0c34ea1fb}} and the amplitudes double
copy has been very recently addressed in ref. {{cite:81931655d508a0c3e426b7ea9da0a7ec0a71e5e3}},
which showed that classical spacetime fields can be obtained as a
Penrose transform of scattering amplitudes which have been transformed
from momentum to twistor space. The known double copy for amplitudes
would then imply a twistor-space double copy, and exactly how this
relates to the ideas of this paper would be very interesting to
investigate further. Another possibility is that there is no genuine
double copy in twistor space at all, and that the results obtained
thus far in refs. {{cite:6c8b5360e7ce201795ddfeca1a5f0606690a3c21}}, {{cite:6fe0b53c0c3314e1e455a17e9a9ee4c0c34ea1fb}}, {{cite:9bbd9f42b277a9e6094db7ce9eee8dce2f7c32d5}}
are coincidental, and do not generalise further. Our present paper
gives us hope that this is far too pessimistic a conclusion, but also
tells us that further investigation is necessary.
| d | d212abc97c84fa208b08b8282caaabd4 |
In {{cite:d95f9ad59c8aa958355c137694b1339325c26a8e}}, the sunspot oscillations in the chromosphere were investigated in conjunction with SJI filters 1400 Å and 2796 Å. In both the filters, the global period increases from sunspot center to the penumbra. They also found that apparent horizontal velocities decrease from 12 km s{{formula:93a91eb4-71c9-446e-8164-e4b806f46c97}} in the umbra to about 4 km s{{formula:6c39202f-3d46-4c03-84a4-247596fadc56}} in the penumbra . On the basis of inclined field geometry, they have proposed that these oscillations are the signature of magnetoacoustic waves propagating upward. Since they have analyzed sunspot penumbra, they ought to consider the inclined geometry of the field lines and projection to estimate the properties of the slow waves {{cite:78c8f9ae0a4c3918a27370010eb6709f7e59d673}}. However, in our present work, we have analyzed multiple locations in the broad bright and dark regions in the solar TR. We did not require to consider the geometry of an inclined magnetic field, in fact it can be considered as a top part of a wide slab piercing the solar TR and consisting of homogeneous magnetized plasma where the wave is supposed to be evolved and generating its signature in the TR layer.
| d | 539786f4562e30e680df1470c4df6099 |
In addition, we also validate our ResNet50 pretrained with VGGFace2 on the AU detection track. The validation is summarized in Table REF . Details of evaluation metrics can be found in {{cite:fcda7c78069cb2e53ec15131bcc6674ce3abf1c1}}.
{{table:7ed3ef1e-3581-496c-84f3-5b30d1d5d061}} | d | f51d6c17d80b4ea34d08736bbf5c6a8f |
In our model, there are five parameters, i.e., the viscous parameter {{formula:1f2e0700-900b-473d-a1cc-06dce7c22998}} , BH mass {{formula:b9b2bb00-6bb7-41fb-a4ba-04cde356a180}} , dimensionless BH spin {{formula:240da768-3000-4d38-92d9-db0088747d54}} , dimensionless accretion rate {{formula:ad32ec37-9a88-4c74-8507-ca8f090451d5}} [{{formula:ba3e3189-405c-4fa8-8ae5-8c03ec942edc}} ], and power-law index of magnetic fields {{formula:1d757929-d405-4804-ba6f-96b6ff918de6}} . Here we concentrate on influences of {{formula:14cac9ef-8a18-4c79-925e-0c07510fdc6f}} and {{formula:4baf96e9-3ab8-4829-bd4e-6102eea16c49}} , so we fix the viscous parameter, BH mass, and BH spin with the typical values of {{formula:3b0774c7-b059-435a-87e0-ef86b37a7809}} , {{formula:91d390e2-0221-4d7c-a233-9c2797fbad60}} , and {{formula:2f96b714-a12d-4894-90c7-f7a656887637}} , respectively. The BH spin {{formula:7aab2d1e-f1bb-44d8-b75a-316eef14caae}} at an extreme high value is selected because the MC effect is relatively strong in this condition {{cite:1e7a8cdeb784a4c0fa4b685df5cc78f36d5347ba}}, {{cite:9f81b2afeca5afd5dbbf983d85d013994ac66ed7}}, {{cite:acb4ad21eee0cc9318af04fb1b06db60ede61c4a}}. We set the accretion rates {{formula:35301f4d-ac01-47af-a760-602f5575f832}} , 0.5, and 1, corresponding to the low, medium, and high accretion states, respectively; and set the power-law indices {{formula:56abe44a-447b-4521-acf1-6dc384f82644}} and 5As shown in Figure 2(a) of {{cite:acb4ad21eee0cc9318af04fb1b06db60ede61c4a}}, in the case of {{formula:8e1ab9ff-5956-4752-93f0-7e24fb589358}} = 3 with {{formula:69ddd1e6-825f-4835-9b97-5d182b91ebb9}} = 0.9, one can notice that {{formula:a7c64308-4ab8-4cd5-8cb6-a4b7b66a7e84}} = 0 and the outer boundary of the MC region was too small to significantly effect on the disc, so we only discussed the cases of {{formula:03fac94a-7a75-4ac7-87f0-b670c1bcd6b4}} = 4 and 5, which correspond the infinite MC regions., denoting the relatively incompact and compact magnetic field geometries, respectively.
| r | a4b7e83c377914bc14c6b3b66b79a6b7 |
By the generalized dominated convergence theorem, this will imply that {{formula:c0764e05-2d6d-41fd-af02-de10feb13cc7}} and {{formula:1e5ac705-8854-49ef-939d-d7551c02b838}}
{{cite:3ed0047e869c51ca149e271e083f076684ff7fd1}}.
Supposing this for the moment, we show how the result follows.
Since {{formula:41014e9c-7aec-4e88-b3a9-0ed4307df3a8}} , by Equation REF we have
{{formula:99c3b6d0-ae02-4fb7-865e-ed907d34cb77}}
| r | 3a96e6741f067da5ee3b3d2b9011eaac |
Agent in Data Augmentation
To efficiently render depth image, we build an agent that encapsulates Open3D {{cite:0976b935c3946ba9e38a77cb6aa6767ad543dcbe}} GL framework. Our agent contains the implementation of depth projection and depth image augmentation {{cite:6c524388cc94bca2b16bf723c86c46251403d0bf}}. To get more RGB and depth image pair, the pose of the camera is perturbed with random noise to render more depth images. Such perturbation results in image pairs with known ground truth {{formula:792b8db4-a8a5-4b20-bb67-a55136f062f4}} . In our experiments, the highest distance for two poses is about 12 m. In the indoor scenario, such a gap is considered very large because the view could be completely changed. The permutation rotation of the pose is {{formula:28b0190d-dd50-47b9-8361-f9c5521372b0}} and the translation permutation is {{formula:cc5b61ea-2a0d-4d67-87ce-02235350aec1}} . Every initial RGB and depth image pair can be augmented into 50 samples in our experiment.
We use Lego-LOAM LiDAR SLAM {{cite:b5329a2dd572272cb0049d925d626897ab980d59}} and sensor synchronization to get RGB keyframes. In total, we have more than 60000 image pairs of keyframes with ground truth pose difference for parking garages {{formula:a59ffed6-1d2b-4d9a-b449-957188289727}} . Each dataset is divided into {{formula:903ec21c-c26d-4c6a-aab8-4ca59cf8c2ff}} training, {{formula:a5373a53-093f-4468-94a9-b767999e39b7}} validation, and {{formula:5d031f14-65fb-49bc-82d8-72d460721497}} testing.
| r | 132f864d0c11d689fef4b5c28fa87c80 |
where {{formula:081670b4-edc9-435d-97f4-047c57ae87c8}} are polynomial functions or formal power series. We omit equations admitting a symmetry of order {{formula:bc8aa523-d0f2-4363-81bd-8b54cb4b70b6}} , which have been studied in detail in {{cite:27b2c914f4cc5e96de87f3a27f30d1e849e4ef29}}, {{cite:b03931bc1e50321bc5696fc86623b76a51ad4fd7}}.
| r | b742f70707543cf7bd5e5613484f230c |
We close this work with the proof of the following duality result that extend the one of {{cite:534428b852494c871cdb156b64110f4bf6fec525}}.
| d | 0e2d7659bfb7eb754acc2b988f3be3f9 |
Following the formalism in the previous section, we turn to numerical analysis with specific decay modes. At the LHC, about 5% of the total {{formula:f9fcf7e0-26ca-4450-b804-620e2b9eb687}} -hadrons produced are {{formula:2ccfafd3-b2b5-4768-b601-c90a3a886207}} baryons, and both at the LHCb and CMS the muon reconstruction efficiency is comparatively higher than the other two charged leptons. We therefore are interested in the modes {{formula:3800af48-65b5-403f-a8f1-3f9fad84a5c8}} and {{formula:618cb2a6-5344-476f-be4c-91f4ce1037a8}} channels. Since {{formula:3c0f7280-ca26-46e6-8ce1-50510766ff1d}} , the CP-odd phase is {{formula:f73dbc13-faed-444b-b5da-a7174f1dc94b}} . For numerical analysis, form factor parameterizing the {{formula:e15549b3-781d-48c2-a90d-5058659643b1}} and {{formula:18c45004-f0cd-4ab0-bda6-c63d65715f6c}} hadronic matrix elements are taken from the lattice QCD calculations {{cite:30345dda391171ef3c8c5fdbafdfde50201d0cad}}, and we take the decay constant of pion {{formula:29375910-eacc-4793-a6ad-ee714e8ba4f2}} MeV from {{cite:49564e4d30e08fd06f178997765debcce676d906}}.
| r | 7a9cfe4e7c19c3848159ffbfde5855c3 |
We suggested a model of purposeful kinesis with the diffusion coefficient directly dependent on the reproduction coefficient. This model is a straightforward formalisation of the rule: “Let well enough alone”. The well-being is measured by local and instant values of the reproduction coefficient. {{cite:579df23dfe015060093e1044a8805164c1720ab7}} have discussed the problems of definition of instant individual fitness in the context of physiological adaptation. Let us follow here this analysis in brief. The proper Darwinian fitness is defined by the long-time asymptotic of kinetics. It is non-local in time because it is the average reproduction coefficient in a series of generations and does not characterize an instant state of an individual organism {{cite:77b4263319eda9b7adab03b7fd666692c0559a5c}}, {{cite:82f80fb8070b93ff7273c6d5ab678593393ab94b}}, {{cite:7ba4649dfd3f84fe2e60fc8000f7b3aa601c630a}}, {{cite:1680f852e867945e2cc8471d20c52e63391d12f2}}. The synthetic evolutionary approach starts with the analysis of genetic variation and studies the phenotypic effects of that variation on physiology. Then it goes to the performance of organisms in the sequence of generations (with adequate analysis of the environment) and, finally, it has to return to Darwinian fitness.
The ecologists and physiologists are focused, first of all, on the observation of variation in individual performance {{cite:4e47daaf9d83174bf4c63c1603f188ddfca2ba9c}}. In this approach we have to measure the individual performance and then link it to the Darwinian fitness. This link is not obvious. Moreover, the dependence between the individual performance and the Darwinian fitness is not necessarily monotone. (This observation was partially formalized in the theory of {{formula:f88dd59a-ff1f-4dca-ad0b-0c6423a29cfb}} and {{formula:3cd6f063-73c8-459d-8ec8-52ef532afc38}} selection {{cite:10280c638f9e3e8322b088d404fed53d3acfae0c}}, {{cite:0030c43f9d05ffd4ccafde2ad8e6bccacb2957b8}}.) The notion `performance' in ecology is `task–dependent' {{cite:3767210b64847665b212950fd08f26c2d2733919}} and refers to an organism's ability to carry out specific behaviours and tasks: to capture prey, escape predation, obtain mates, etc. Direct instant measurement of Darwinian fitness is impossible but it is possible to measure various instant performances several times and treat them as the components of fitness in the chain of generations. The relations between performance and lifetime fitness are sketched on flow-chart (Fig. REF ) following {{cite:3767210b64847665b212950fd08f26c2d2733919}} with minor changes. Darwinian fitness may be defined as the lifetime fitness averaged in a sequence of generations.
{{figure:f7d8a14c-1019-4d37-9e2b-487e3ad5dd1c}} | d | 2f52b5d9fdb60293249a1decdec425ad |
Computation Cost. Compared to pure ILs, there are two extra cost sources in common meta-game analysis: approximating and solving the meta-game {{cite:b504f366d1761b9700d71586327cb1cb3681275b}}.
In our case, the meta-game is restricted to a local two-action game, where two actions, {{formula:4667feab-89bb-4d94-89eb-a3a20f0f141b}} and {{formula:7f627072-bde5-4f34-8ac7-7940fb8f6a59}} , are close to each other.
Reusing the IID trajectories will some estimation errors {{cite:e4d8c09f21256fe774c65ac27ee61f5171fc4f05}}, but this issue can be eased by large batch size. Then, we can enjoy this proximity property and reduce the meta-game approximation cost (without extra sampling) by reusing the collected trajectories in the IID step.
The next crucial problem is how to solve the {{formula:487e11a7-0a6d-4806-a7c9-cba06c504524}} -agent two-action meta-game, which consists of the {{formula:2b4eee5f-100b-4919-973d-e58c59cbefc9}} entries of each of the {{formula:b08f8873-2c33-40b0-966b-443db298e5bc}} payoff matrices.
Solving this meta-game is much simpler than solving the whole underlying game, which increases exponentially with state size, action size, agent number, and time horizons.
As the general-sum matrix-form game has no fully polynomial time approximation for computing Nash equilibria {{cite:21c04972c7334f1022c0292ba6dc99cfdf83586e}}, it usually costs a great deal to solve the game {{cite:5ace56e5df2a5a402c68675e3e18f718263ead9c}}.
However, as shown in Remark REF , there always exists at least one pure Nash equilibrium in the meta-game, which can be computed in polynomial time {{cite:094ce5e85808c7793ec7d8504f8da8d214b10c1b}}.
Therefore, if we only require an approximated Nash equilibrium, then when {{formula:786401dc-9c5d-4dfb-bc4d-7e7843acee8e}} is small, for example, {{formula:68cfc04a-78a5-4613-90ee-10a9dbec67c6}} , it is affordable to find a meta-game Nash equilibrium with subexponential complexity {{cite:a60a85a8225d01621b8d88639d7081080fac8e92}}.
But this problem still exists when {{formula:1464fad7-107a-432e-8923-4c16e9bc19cf}} is large.
In this case, we can try a mean field approximation {{cite:a069c716f73d5606841ff9983001d4a0db1431ae}} or utilize special payoff structure assumptions (e.g., graphical game {{cite:273590116e107c40326778bbed8d98f01380d92c}}, {{cite:5ace56e5df2a5a402c68675e3e18f718263ead9c}}) in the meta-game to reduce computational complexity.
| d | 2fa1af88ad91d08f64c3ca849182d219 |
Using Corollary 1, we can derive the result of lock-free parallel optimisation algorithm {{cite:dba6dcecf646b25d3eb81c0da5aafcd3da79db59}} and the asynchronous distributed optimisation algorithm {{cite:0c1e208e582718ef7177807a3bc58c421aeaa694}} as particular cases. By setting the number of threads {{formula:188817b7-565f-4cda-9987-5e1d2aa07968}} and the number of local update {{formula:84e8870c-4e17-4b1d-9ef2-27254b7e2a8e}} , we end up with the distributed asynchronous algorithm presented in {{cite:0c1e208e582718ef7177807a3bc58c421aeaa694}}. The convergence bound of Corollary 1 then becomes {{formula:4e190993-d946-4b71-8ecb-4f1c65a6fc84}} which is equivalent to that of Corollary 2 in {{cite:0c1e208e582718ef7177807a3bc58c421aeaa694}}. By synchronising the global learning {{formula:38f5e755-7e6c-46ea-a1e6-03221337a32a}} , setting the master batch size {{formula:2be9453c-d877-4e0f-8da1-a2151ff229c1}} and the number of global iteration {{formula:cf248579-5683-4c83-b33f-b329a137b01d}} , we end up with the parallel lock free algorithm presented in {{cite:dba6dcecf646b25d3eb81c0da5aafcd3da79db59}}. The convergence bound of Corollary 1 then becomes {{formula:827d5346-cc99-4787-9323-cd2624b3a2f9}} which is equivalent to that of Theorem 1 in {{cite:dba6dcecf646b25d3eb81c0da5aafcd3da79db59}}. The experiments below will empirically demonstrate these two parallel and distributed particular cases of DPSGD.
| d | ba4fe11c4022ce477bb173d076f9a120 |
({{formula:a1d3361a-1a57-488b-86be-91c8346c269c}} ): The U-Net implementation used in this work is based on residual units as used in {{cite:d0cb522ee5af42e70dd664e4f6d132209f607379}}, instead of using the classical convolutional blocks, this is meant to to address the degradation as proposed in {{cite:6c2b17f7619a0126987921cb1b2ce50128b5b3d2}}.
({{formula:f954b599-c1f7-425c-b665-b00f1f7535ce}} ): Is an implementation of Mask-RCNN {{cite:66f986c85e7cd1e9b4b4113f4b119e9e2588cc49}} using ResNet50 as backbone. Mask-RCNN is composed of different stages. The fist stage is composed of two networks: a “backbone”, which performs the initial classification of the input given a pretrained network, and a region proposal network.
The second stage of the model consists of different modules which include a network that predicts the bounding boxes, an object classification network and a FCN which generate the masks for each RoI.
| m | a79aec335ec87d214e5a96b405eddd37 |
We assume that the readers are familar with the concept of vertex operator algebras (VOA), modules over VOA and the {{formula:7a93c4f4-b940-41c3-9c20-cf2c494bea03}} theory, see {{cite:d84e93e3b820012e7083eb664caf84eee3a19ce5}}, {{cite:2bcc26e0462d0d29c5eb642fbb3c21e5fdce36df}}, {{cite:0c2cf2d981e84ac15ddbaa666317c9735634857b}}, {{cite:270e91a3348d6c78b4f926763038829a9ad4a11a}} for example.
The fusion rules formula can be stated as the following thoerem:
| i | 0337b8ef88cff836d7ea02cde4923a5f |
always reaches the global minimum after a sufficiently long sampling time (cf. {{cite:20b265cec9b85d4b9016e88f9fcc8c3120a66999}}).
| m | 66598920c5e713b2879cd0bd351e501c |
with {{formula:2a353a47-28c6-4bf8-b0c8-e523ac147ee4}} . It follows that {{formula:04a0f891-1446-4f0d-9823-0413783da47b}} is well defined by (REF ) (see {{cite:5ab021148a81510b3064ba1794c292a879f62a1e}} and references therein). Hence, the problem (REF ) has a unique solution.
| m | 7c30905993b0a15b7c5431a46727947f |
Overall, ethical aspects of applied AI are a hot topic, and we partially cover it in previous work {{cite:a3fb24f4c986d6dc167fa9376ecbf36e08f9eda9}}.
Jobin et al. {{cite:54c6c1a201c113c7cc957581614e435df501bd6a}} found that transparency and justice/fairness were top ethical principles in guidelines for ethical AI {{cite:54c6c1a201c113c7cc957581614e435df501bd6a}}.
The AIDOaRt project strives for AI approaches that are beneficial, safe and responsible {{cite:9dd346cfdda23dc76704fdbe9b9e19ec31c4c00b}}. Furthermore, the concepts of accountability and explainability are highlighted in the conceptual architecture of AIDOaRt (in Figure REF ). However, only one generic requirement partially covers an ethical challenge in that an AI ought to be able to motivate its actions (Mon.6). This falls under the most prevalent principle of transparency in the classification by Jobin et al. Future research could explore how, and how well, AIDOaRt complied with ethical principles.
| d | 9405e95b4621c86b0e95106a4c155324 |
First, no single existing AST-based model can beat the simple token-based models in all tasks. This may indicate that the plain text of source code has revealed strong naturalness of programming language so that applying classic NLP models directly on source code could already bring pretty good performance on most tasks. Indeed, recently a programming language model CodeBert {{cite:ae5e4d17cc74923f39b45ac5fa657159a879f17e}} was trained with the plain text of source code (and natural language sentences) and showed very good performance on many tasks. The high textual similarity in the current benchmark datasets may also facilitate the training of the token-based models. Nevertheless, for a particular task, we can always derive an AST-based model that performs better than the token-based models. This can be verified by the performance of ASTNN and GGNN for code classification, TBCNN and ASTNN for code clone detection, and code2seq for code search. Therefore one may always try to explore the abundant structural information in ASTs to achieve the best performance for a specific task.
| d | 3787bbe51e28641f2b171cb14cc63d4a |
In Fig. REF , we report the density of states (DOS) for {{formula:b90f49b8-391e-4df8-9219-d8d044a15e1b}} and {{formula:e1281d62-55c9-42f8-998e-9f8f2c329d58}} at different temperatures.
For all the temperatures, there is a strong renormalization of the bare band whose width becomes twice smaller (from {{formula:3d69de92-441b-4a14-b4da-92a24395e604}} to roughly {{formula:06cc08da-731a-4696-aff1-def2c794eea4}} ) and move to lower energies (from {{formula:6cfb1052-8b2e-440a-9ab7-cb864d1190ea}} to roughly {{formula:81c28408-ee8e-4fbf-be71-4b16a4e46377}} ). Furthermore high energy satellite bands appears at multiples of the vibrational frequency {{formula:210da7ef-5a54-481a-b411-dd774d6d5055}} {{cite:a271d898cf8efa10b58c84676f5a0ce3010e230d}} providing a DOS extending from {{formula:63cb2bd8-c29b-4cd1-bae3-aa9184668820}} to {{formula:824842e5-c1db-4ad2-911e-089a294e0120}} . These effects can be easily ascribed to the local modes since they survives also at {{formula:720e583f-a244-4757-94b3-c7461ebf1ab5}} , where the effect of non local modes is weak.
The intrisic reduction of the bare band due to local modes provides a simple and direct explanation of the difference in the bandwidth evidenced in the series of oligoacenes from naphthalene (effective band of the order of {{formula:45485557-391a-43a3-b0b7-544ec793894f}} ) to pentacene (effective band of the order of {{formula:3fa0df3c-6361-456b-bbb7-271c1ddbef7b}} ). Indeed, it can be ascribed to the decrease of the reorganization energy with increasing the benzene rings of the single molecules that in turn reduces the renormalization effects {{cite:d5a4fe6b2e19fc4f5109f371f23a05da1ed8e896}}. Within the polaron theory {{cite:a271d898cf8efa10b58c84676f5a0ce3010e230d}}, the narrowing of the main band is related to the spectral weight Z of the quasi-particle, which is estimated to be about {{formula:6e5fc830-c577-40b7-ab19-1c0944ca13c6}} from the calculations. Therefore, our estimate of Z compares favorably with recent ab-initio results for which Z relative to the electron channel is of the order of 0.7 {{cite:fbcbc253500465849934cda8cf2b7b42a40c9e94}}.
| r | 159b735910b3ee763c35bff60db7182a |
Our model based on the EFTofLSS was described in Section and has 10 free parameters. The monopole and quadrupole are both calculated at redshifts {{formula:8ff40bca-216c-4f30-bd21-76f7d4988d2d}} and {{formula:ed9b296f-65c3-4a8d-9a87-f9ce5a786a38}} . For each likelihood calculation, the cosmological parameters are assumed to be the same for the two redshift slices, but the bias, counterterms, and stochastic parameters are allowed to vary. The axion fraction is varied like the cosmological parameters while the axion mass is fixed for each individual calculation. We thus make sixteen independent likelihood calculations: for each of the eight mass bins between {{formula:1e006fe1-8ff4-4489-9dd0-24d9462e7619}} eV and {{formula:02ff2575-0b70-4b40-a862-0bd3acbc6696}} eV, we compute the posterior distribution for two choices of priors. For the first half of the runs, we include only information about galaxy clustering and use a Big Bang Nucleosynthesis (BBN) prior on the baryon density {{formula:c5adeee0-c44b-4481-a69d-e603ed5c5c5d}} . For the second set of runs which combine our setup with existing constraints, we instead impose a CMB prior on all cosmological parameters along with a prior on the axion fraction also from the same CMB analysis. We thus use the pre-computed chains from the Planck likelihood including an axion component directly. This method has been used as an alternative to joint likelihood analysis when combining large-scale structure and CMB data with the EFTofLSS (see Ref. {{cite:0140d312f12b612bab9c62534e4b71791e7e894f}}, {{cite:a6bd553346a1daee7ee35b0b392f74bd8f7a71bb}} where a CMB prior is put on the baryon density and sound horizon).
{{table:3bc7ae92-3037-4740-91b3-6f8d425ff076}} | m | 24a0b576da7db57cc981ffdf7cb8e9e1 |
Each input image is transformed into a feature map divided into {{formula:c8bea453-420b-4e56-960f-f56c10e1b321}} patches.
The {{formula:ca330a1e-03ff-4d02-952c-023ab2a06c4b}} -th patch, with {{formula:2889cfac-5918-4f38-8dee-f23570311239}} is fed to the corresponding column {{formula:a88881d4-0e8e-4ac5-9a0c-834eba415c0f}} , spatially located at coordinates {{formula:c870150d-019e-447f-bba9-3bb220b571c0}} . The subscript {{formula:d8363149-6d33-4cc2-80d2-55017b0a923a}} is omitted in the next equations for better readability.
As shown in Fig. REF , each column {{formula:adbfcc29-2ef0-4fab-b88b-ef1fe3ea25e7}} consists of {{formula:c1522aab-9759-4083-aee9-5acf5c96bdcd}} embedding levels {{formula:6358888a-5444-4673-8acb-237e82c7e829}} connected by a stack of auto-encoders at location {{formula:a5d77099-d3dd-495b-8e36-bbfa4464f3b1}} at time {{formula:1b902433-7211-4ff3-94fc-419c98429541}} , as suggested in {{cite:bfc0300c6ab425d574bc4162f0ed2068c63a77d3}}. The superscript {{formula:5402baa8-7d56-4645-ae1c-41dfbae0c0c6}} is omitted in the next instances of {{formula:627d1180-953e-4df6-83b7-14718bd45c84}} for better readability.
Each level {{formula:f2196f1d-f945-4e4c-9baf-0b344963a80a}} of the column is an embedding vector representation of size {{formula:904cd0c6-b8f8-4d84-8625-ef638d651aee}} . Levels {{formula:41b4c605-0a63-4715-87a9-f1006a394da8}} and {{formula:dbe6257c-852b-4e53-bff6-7baeaac4ddbe}} represent consecutive levels; {{formula:e3a8873d-a18e-4812-90eb-a30a629693fa}} represents a part of the whole {{formula:f51c9956-7fb4-4c55-a90e-35e9e87a32a1}} . We indicate as {{formula:0d1e45fe-e304-44c9-8ffa-641966170c2d}} all the levels {{formula:039f8595-d39b-48e9-83d0-06afcadf2408}} in all columns {{formula:acdce03a-80ab-484e-b2fa-3a0e7c63dcb5}} sharing the same {{formula:aa1aa94b-bcf9-422a-bd5b-a0a012cdbe18}} value and belonging to the same layer {{formula:6b33de24-dd0c-4986-9972-2787e6696376}} . Being {{formula:8d293278-6148-4d54-b1ab-c8604be04491}} the last layer of our architecture at the last time step {{formula:84fde277-f1da-4ab3-9e94-bdfff7c30887}} , it is represented as {{formula:7b8f1f47-41f5-40d0-b1e3-9c8a7c247d93}} .
| m | b424764a4b55fd22f724db2c1aa713bb |
Cases {{formula:2c314fe8-f7dd-4905-8a9c-bed850af8c4d}} and {{formula:25338b0c-cc3b-4926-8473-72e609b983d1}} are treated in the same way. Let us only prove the statement when {{formula:42dc6ee0-cd4b-4dba-9c6b-365cd6d6d0d9}}
Dividing {{formula:4273a335-c2c8-4fda-81a3-a68afcdc3757}} with {{formula:b4f7583e-2907-4a93-bb5f-cb0fa975fb3b}} we can assume that {{formula:424111b5-b548-4d23-81e1-a572e73ea636}} and {{formula:eb9f2c8b-cd6e-41a3-9fbd-8f60296a9cfd}} Take {{formula:29a4556c-e73c-42db-84b4-edcb555c93b4}} If {{formula:3d632676-2608-48da-b735-306fba433488}} for the constant {{formula:c6e92908-337f-47a7-85c6-f1259f31a057}} and function {{formula:f9aaeb66-d4f2-43ce-8458-6cf7beace4a4}} from Lemma REF , we apply the comparison principle ({{cite:da5326c06e249801767b4aeff89757189a5cd6a6}})
on {{formula:c2c2be8a-8b61-4b1e-85af-484e86c15ff6}} and get that {{formula:4843de8a-895f-410f-8332-9ee0d991a1c9}} Taking {{formula:a8632521-72f7-4bf9-9f85-13052d0dedc5}} bigger if necessary, the same holds also if {{formula:2b3e01c9-57dc-423b-b8da-b8fb90e2ecb1}} . Note that thanks to Definition REF the function {{formula:a058387e-0d9c-4605-b7d2-ee61e047e162}} is also comparable to {{formula:a8eee25d-28ec-47e5-8e31-10f00ffc2a87}} . In case {{formula:71bcfd46-db66-46e3-ac7f-188885757c24}} Lemma REF together with comparison principle gives the desired result.
| r | 0fe0d6d55797032e60d1c1e59bfdbcd4 |
The resulting model amplitudes, including additional background terms, were fitted to the available
data for {{formula:b2116efc-c0f1-4de8-a5f5-643d1c0dc76d}} in all charge channels in the region up to the total energy {{formula:19590355-5ed9-4dd2-8fe3-5b377ba3d323}} MeV.
The asymmetries {{formula:5885abec-3481-49f9-9ff6-ce5fa0fd1c7f}} , {{formula:9f2a2f4f-1fe7-444f-b269-1cdfb81141c0}} , {{formula:7e073809-4d2b-4e23-9fdd-bbebc85f0a9b}} , and {{formula:b93649ee-d13b-4e18-be07-54e76fb5ee5b}} obtained in the present work were included
in the fitting procedure. The {{formula:93cfbfcf-ee4c-4613-b860-fc78794becc2}} and {{formula:d9d54b95-4b37-4e25-8583-ac5cf9439c29}} resonances fitted to the available
{{formula:5fc708ab-40ce-4b63-99ff-8e185d670e0f}} data, along with their resonance parameters are listed in
Tables REF and REF .
For the fitting procedure the initial values of the resonance parameters were taken from the
current Particle Data Group listing {{cite:b2a7a6baccf9a8814575bd8861495be65615e486}}. Since the resonance terms are proportional to the
product of the electromagnetic and hadronic couplings in the tables the products
{{formula:92af968d-346e-42fb-bf22-1fd3fe7473b8}} with {{formula:4106d16e-9c9e-4d46-a673-89d6ddecde0d}} and the ratio of
the helicity amplitudes {{formula:df7db63c-c407-4003-8468-d277677c2e3b}} is given. The corresponding total cross section as well as
the target and the beam-target asymmetries are shown in Figs. REF ,REF -REF
as black solid lines.
| r | de9637a6ee68978fd55095ae8e9cbc68 |
In Figure REF , we display some qualitative results under 1/8 protocol on the Cityscapes val set, and all the approaches are based on DeepLabv3+ with ResNet-101 network. Benefited from the proposed framework with a series of components, our method shows more accurate segmentation results than the previous state-of-art method {{cite:fab2bc87d55b9ad5072a9d8531d041f5a881360c}}, especially for the large region with complex texture and needing long-range information.
| r | d9e67e4c39502c0b61410677356e942b |
paragraph4
.5em plus1ex minus.2ex-.5emVariance of interest. As each encoding is the encoder output of an augmented view from an image, the total variance in encodings mainly comes from three types: i) changes to the encoder, ii) changes across images, and iii) changes within a single image. For type i), MoCo {{cite:122e7740674dda970e36a8a0a3e429a1b8732022}} with its momentum encoder is already a major, well-studied asymmetric design that intuitively reduces the target variance across training iterations. For type ii), as Siamese representation learning encourages uniformity {{cite:2cf0f5110236a3a2ced3c5b044c47d6af58b42b1}}, {{cite:035836f9757bfdc5fa3c72e69ecec9162473a0cb}}, the cross-image variance quickly converges to a constant dependent only on encoding dimensions (evidenced in sec:varianceappendix).If encodings are uniformly distributed on the unit hypersphere (due to {{formula:d6f84d7d-c9e5-41c4-898d-3265fe2b7dd6}} normalization), their variance is {{formula:630f9d50-c773-4166-848c-7928e106f319}} where {{formula:e35eb7fb-f3e7-43fc-b122-ef964f9ed2fe}} is the encoding dimension. Therefore, we focus on type iii), , intra-image variance as the main subject of our study. Note that it does not restrict us to design input augmentations as the only means to adjust variance, as will be discussed in sec:syncbn,sec:meanenc.
| m | 7853fff2b2ff14cc6335124fa904df36 |
Our study differs from other approaches in network inference in so far as our aim here is not to infer the existence, or otherwise, of links but rather to infer the most likely network class that led to the observed population-level data resulting from an epidemic spreading on it. As a result, the data needed for inference does not contain node- or link-level information. There are both advantages and disadvantages to such an approach. On the one hand, the computation of the likelihood in our case is more straightforward and the data needed for inference is modest. On the other hand, if more detailed data is available, the proposed model will not be able to capture it nor benefit from it. However, more complex models will need large quantities of detailed data (i.e., in the case of cascades, the data needs to contain cascades starting from, or involving, as many nodes as possible, {{cite:16210c8c0d5ca0a055cf9a4ece73bc7384700344}}) to produce acceptable results with large computational burden. The choice of model and inference will depend on the context.
| d | 3e7f9f40e8be57317167d4a38502f9e7 |
paragraph41ex plus1ex minus.2ex-1emDatasets and Baselines.
We follow the same experiment protocol as prior works. Two datasets, CIFAR-100 {{cite:3163f6944af7c32d0a36e9e8cee4b14f3d1082b0}} and ImageNet {{cite:a97394b04dffce3b7af323c46bd1f59b20e663a1}} with 100 classes selected randomly, are used for experiments. The data samples of these 100 classes comprise a data stream that arrives in a class-incremental manner in 10 phases, each adding 10 new classes to those already received in previous phases. Particularly, we adopt two training scenarios. One is to train the model from scratch, i.e. the model starts with random initialization and learns 100 classes incrementally. The other follows {{cite:e3f130c45602b1cb266402c94c30907672fe721c}} to train the model with the first 50 classes in order to have a reasonably good feature extractor to begin with. Then the remaining 50 classes are split evenly and learned in 5 sequential phases. The baselines include iCaRL {{cite:bbc3ea86e358e1d39809648d903b3d90e7e44982}}, End-to-End {{cite:adb4c0bd274ef12f5987bb7cf858514c6264ac5c}}, BIC {{cite:80ae699cbcbf3932e29d3f1c9f944f54f8157da3}}, Hou19 {{cite:e3f130c45602b1cb266402c94c30907672fe721c}}, and Mnemonics {{cite:2d14efdee812ebad8cf4acde5cc4729f3ca97022}}.
| r | 490ea44b2db558d41586646a87a5b045 |
Existing methods for multi-hop QA have two main strands.
The first is to predict the sequential relation path in a weakly supervised setting {{cite:8297c7c8f487c9c7da2b466a56ddfba7de25d9dd}}, {{cite:629f4f58c37b539515aec46bc14259c95c8b1bfb}}, that is, to learn the intermediate path only based on the final answer.
These works suffer from the convergence issues due to the huge search space, which heavily hinders their performance.
Besides, they are mostly proposed for the label form.
So, it is not clear how to adapt them to the text form, whose search space is even much huger.
The second strand is to collect evidences by using graph neural networks {{cite:026942d833084c802a06d8ef2005e926a8b33a50}}, {{cite:dfbeff474cd713818b92adca1a7b15ff804c649e}}.
They can handle both the two relation forms and achieve state-of-the-art performance.
Although they prevail over the path-based models in performance, they are weak in interpretability since their intermediate reasoning process is black-box neural network layers.
| i | c4af04d0737234bf1492299a6812935c |
which is the counterterm of the asymptotically AdS spaces. Having the total
finite action {{formula:a5967d5b-23b5-46da-86e4-133a811601ec}} at hand, we can use the quasilocal
definition to construct a divergence free stress-energy tensor {{cite:164f92a4d91136edcfac38335f8b1395d087d8b5}}.
Thus we write down the finite stress-energy tensor in {{formula:ad0911d6-6bea-42d7-aadd-e87cd7fa99b1}} -dimensional
Einstein-dilaton gravity with three Liouville-type dilaton potentials (REF ) in the following form
{{formula:858d239b-7338-43df-b1b7-3a9bab3638e8}}
| m | 753ffbc729e5f65e4f327c5597bb941b |
Chignolin. For chignolin, the crystal structure was first obtained from Protein Data Bank (PDB ID: 5AWL) {{cite:b9c4fa19ee89f15299417755ec1159d8a7d7584c}}.
In order to achieve the initial unfolded conformations, one MD simulation in vacuum at 1000 K is conducted for 5 ns.
12 fully extended conformations are randomly chosen for the initial conformations of adaptive RiD.
They are solvated in a (4.2 nm){{formula:71703531-53e0-4d86-bb9a-ff2d26138d82}} dodecahedron box with 1622 water molecules and 2 sodium ions to neutralize the charge.
All simulations are carried out with the CHARMM22* force field {{cite:b6eb63169940d54f05d97a1aac821baddfdc41f8}} and the TIP3P water model {{cite:48cd5a12a38021d9d05cefd4e37be44d5dd53cd5}}. The temperature is maintained at 340 K (in agreement with some previous work {{cite:27d02622e2e0d19d4af119e7ecb0963bd3de61a0}}, {{cite:73074cc19766bf7996548de93136593f8180c8f8}}).
18 backbone dihedral angles are set as CVs.
12 walkers with different conformations of adaptive RiD are conducted for 31 iterations and the biased MD simulation lasted for 5 ns in each iteration.
The CV values are computed every 5 ps.
All other parameters are the same as in Peptoid.
The reference structure of chignolin is generated from a short, conventional MD simulation that started from the X-ray structure (PDB ID: 5AWL).
This reference structure is a better reflection of the actual folded state of the protein in the CHARMM22* force field than the crystal structure.
Since the maximum RMSD between X-ray and the NMR (PDB ID: 2RVD) structures is about 0.18 nm, we define the folded state to be the structure with RMSD {{formula:f4cf481d-9408-4e81-a9d5-c84f61974470}} 0.19 nm from the reference structure, and the unfolded state to be the structure with RMSD {{formula:6bf6b966-e302-4489-913b-402dfb83675e}} 0.3 nm {{cite:7bc7d2bc53ea5134bf9d236235b0f8790491c045}}.
The RMSD is smoothed along the trajectories with a sliding window of width 3 ns to eliminate the overestimate of folding/unfolding rate due to thermal fluctuation.
| m | b2178fda9d13844756352b2d9941be79 |
where {{formula:28ed6b5e-ba70-4875-8c06-c8ff4454d7de}} and {{formula:45210c4e-9fa0-4138-9517-9b4ccf728f15}} are given as functions of {{formula:e8ca5925-4b97-42d3-ba13-a127e3ffb0bd}} by equation (REF ). This relation between thermodynamics and entanglement in harmonic systems at their ground state has been exploited in {{cite:ac94772e8a3db5c0245bf91ba7cefbe89260c560}}. This temperature spectrum is not contradictory to the thermalization hypothesis {{cite:e063f902d087704180042812b1000f769a90a9e3}}, {{cite:92e85379946c75e686c3903d32091232b6894bd7}}. The latter applies only to non-integrable systems.
| d | 7b0bccc4d93fe63b5cf364dd66bd4045 |
Today’s commercial speech recognition systems in both academic and industry fields require ever-increasing volumes of text-annotated speech signals for training. The need for massive data in supervised learning hinders the fast advancement of speech processing research. To tackle this issue, self-supervised learning (SSL) {{cite:b8fa9201f20dee44c6a3ed7fbb66f8421a159c38}} has emerged to reduce the dependency on large labeled data sets. SSL utilizes proxy supervised learning tasks (also called pretext tasks) to obtain training data from the tremendous amount of unlabeled corpora available on the web. Researchers explore various pretext tasks to pre-train large neural networks without labels and transfer the pre-trained networks to solve complicated downstream tasks. Recently, SSL has become one of the research mainstreams in speech processing as well as other machine learning communities such as Natural Language Processing (NLP) {{cite:94e292ba681ae542547e1ca70b752784d57cbe2d}}, {{cite:184ba47d598d1925115f10fac164abbbb0fea925}} and Computer Vision (CV) {{cite:6d04ab37011680498ab46abaa0684b6822c8d027}}, {{cite:b04289f587f36b80f3a8bdecaad1fe4532c69c13}}.
| i | 0b4ce4073906d5449da16fb5a99f3537 |
Looking-to-Listen model {{cite:6d1263852045ae94df8e3e4cf0b69c56f84d92b9}}: A speaker independent audio-visual speech separation model.
Online Visual Augmented (OVA) model {{cite:fedd0c23fad05d83dfdcacf209792ddca9396662}}: A late fusion based visual speech enhancement model, which involves the audio-based component, visual-based component and the augmentation component.
AV(SE){{formula:c3ac019d-2ea6-4dcb-86f2-8433b9771dfe}} model {{cite:0e1cde3054d2e4c59a4a9d1e1ec6cb36daa8fe85}}: An audio-visual squeeze-excite speech enhancement model.
AMFFCN model {{cite:6d2fd87548a764b21b693941beb423736963cc9d}}: A visual speech enhancement approach using attentional multi-layer feature fusion convolution network.
| r | 66bd662b20509b23558e1011d8708033 |
All these algorithms and indeed the general RL study have been hitherto predominantly limited to discrete-time Markov decision processes (MDPs).
It remains a largely uncharted territory to study RL in continuous time with continuous state and action spaces. The few existing papers on RL in the continuous setting are mostly restricted to deterministic systems; see for example {{cite:42f30e742363691f06e8f0361f9ed32152f79c57}}, {{cite:f3cc8519a290015c40cd9f2b4cca063a23153a54}}, {{cite:8fa5d92908109f54b1d84625b38ea35879cea32d}}, {{cite:0c14031d647aaa4e5ee6df209c5b5f5c3659e140}}, {{cite:8c3528fab8fb8973a369c1bf84bb2c0cdbf0fc1b}} where there are no environmental noises.
In real life, however,
examples abound in which an agent can or actually needs to interact with a random
environment at ultra-high frequency or outright continuously, e.g., high-frequency stock trading, autonomous driving, and robots
navigation.
| i | 08123bc022af98c14b479ca56fa94857 |
Existing methods for learning-to-defer to an expert aim to improve the performance of a prediction task e.g. {{cite:670e29118b908ee73095229f0ad55ff9bbdeb0b6}}, {{cite:93113003430fb2f4fa8972313ee059b439c7578a}}, {{cite:f79b4eb5f790040aa6aeba8fb26adaad6a5cd1ec}} by deferring to the expert. These methods defer to experts either based on the confidence of a model prediction or by characterizing the trade-off of paying a cost (to defer) and improving outcomes using human decisions. These approaches do not account for the sequential nature of decision-making settings, nor the non-stationary dynamics over time. Non-stationarity leads to propagated uncertainty that increases with time for longer-term decisions, and we demonstrate ignoring non-stationarity leads to underestimating this propagated uncertainty, resulting in delayed deferrals and worse long-term outcomes.
{{figure:60a7301d-4141-4188-aa29-50c5cf7555d2}} | i | 7fa9454563652e6e6942316acf3c9a62 |
As discussed before, our method achieves some measure of knowledge transfer to restricted discrete action spaces. However, what remains unexplored in this thread is the application of this method for transferring knowledge to expanded action spaces and action spaces that have disjoint components. For target domains that feature the discrete portion of these action spaces, our method can be still be applied since the input for the Q-networks used in these discrete spaces only depend on the state. The same observation applies to the continuous portion of these target domains since the corresponding critic networks already accept inputs from {{formula:a9ea0a52-5c78-4c11-aa37-a161151f3335}} . Another unexplored direction is the application of our method to other underlying RL algorithms. While we study the use of DQNs {{cite:101e9b9a54794cf219d8d5c4ca07098738bcd666}}, DDPG {{cite:58ccde86609f28cebac7d72574a647fecf5b9b9a}}, and TD3 {{cite:d73742dd4fd3d58889de6b25570edd549cecd403}}, the analysis of our method applied to other recent RL frameworks that employ Q-networks would provide further knowledge as to the efficacy of our method. In summary, our method presents a plausible direction for performing knowledge transfer across differing action spaces.
| d | b1fb5e4206f4cea039adec81db75d9b6 |
Empirically, we found more learnable parameters are needed with greater {{formula:c81c5cd0-78e2-428f-b469-a356ac02b0b7}} . Thus, our generator {{formula:7ada3a60-9bd5-4f95-acd0-e13741b107cc}} is a CNN illustrated in Fig. REF with 16 residual blocks and {{formula:d28d1fdd-b21c-4061-a802-05eacce95c6d}} kernels of size {{formula:514dd449-4cec-4980-808f-cd4c04c9c9f7}} per convolutional layer. The discriminator {{formula:df335dfa-0a90-4b6a-94ff-3a69c0c86e6c}} has the same architecture but with only 2 residual blocks and {{formula:35be4aa8-2dbd-4f58-9884-416b8eac58c6}} kernels per convolutional layer. Our final loss function for stage 2 is identical to the loss proposed in {{cite:e5333152de50243af2ae42c05b05b52467eea5b3}} and is calculated on the error in {{formula:b6650d7b-1d02-49b3-a21c-a5421e9c29eb}} . We use the AdamW optimizer {{cite:4dfb87921b075c61aa7fbb5e10b62f5501d91231}} with a learning rate of {{formula:f3de2d42-4976-4dda-824d-8ec373199d43}} and the one-cycle learning rate scheduler {{cite:7dc6c91051ad4cc7abceb8838bc9915ef3c6ca5c}} for {{formula:7f25eeee-4737-4bc1-b8cb-24b82af38aeb}} steps at a batch size of 32.
| m | 52c3d268f964b09b8fa4edeb08ad6e8e |
Parametric denoising methods in the literature consider the noise in PET images to be additive Gaussian {{cite:3a0897edc670a6f654c0bf0e7f862d260f8bd42c}}. However, Gaussian assumption in PET images may result in the further loss of already poor resolution, increased blurring, and altered clinically relevant imaging markers. Recent attempts such as {{cite:51b1fbcb9be1fdb8c993139883933030d1f99a3f}} used a more realistic Poisson-distributed noise assumption in PET images where authors first “Gaussianize” the Poisson measurement followed by unbiased risk estimation based denoising filtering. Gaussianization is achieved by applying a linear transformation such as a square-root and known as variance stabilizing transformation (VST) {{cite:d8e34ad72f2cdc68c97570d59fc4b5a493713ac8}}. However, the algebraic inverse VST used by this denoising method may be sub-optimal. Regarding these difficulties, we proposed a novel approach in this paper to denoise PET images using the optimal noise characteristics and a 3-D structure preserving noise removal filtering.
| i | 88074ba7e5bd985ddd3d1151ca97deec |
A more popular algorithm is {{formula:3898df38-fda8-4d8d-aa7d-0ebd5ea0e9d1}} {{cite:4ac4db40486eb9354afc4550c149ea812e02a757}}, which adds EMA-style heavy ball momentum to rmsprop.
Adam has an optional “bias correction” scheme.
With no bias correction, Adam employs the rmsprop preconditioner rule in Eq (REF ); with bias correction, the Adam preconditioner is:
{{formula:8d837c91-9414-47ea-9b9a-b13a180db17f}}
| m | 01aae7eb925c010e436360466b272507 |
Some methods, such as the wavelet approach in {{cite:b0132a5b37e0b2c848d5c22a5783816c8a605a46}}, allow for a bit more of an unsupervised approach in terms of data sampling, but the results are largely the same.
In theory the input to a LSTM model can also be trained using complete life time sequences.
Additionally, stacked LSTMs may allow a model to learn different temporal dependencies {{cite:fbb85626b738e3651b8257e4925e9710003b1d26}}, which was also tested in this article but showed no improvement.
| d | 0c7a8d6fea56fa8699242f9ea517c0e1 |
The fact that Newton's constant is the natural cutoff is consistent with the idea that the Bekenstein formula for black hole entropy involves a renormalized Newton constant
{{cite:8e2c71be0d9a5b5033accad0ff9cb9d8eda937b4}}, {{cite:142701f5e01af59a71a8d94118c64f7c22382c65}}, {{cite:9c1fbf753e0bc69df6c72d04ca461001fcc09ad6}}. Furthermore it has been argued in {{cite:e33c36b96ce9fa7cc16bc66bf85e47382fdec4a2}} that this naturally happens in theories of induced gravity.
In the past, {{cite:cddd5fd1662df13baefc7d3db8e85e8b563f65cf}} has argued that Einstein's equations follow from thermodynamics, provided the cutoff in the entanglement entropy in a theory of gravity is Newton's constant. {{cite:5a77a189e5fe84e3d87bb6861df205c3f151ba12}} had also conjectured that the entanglement entropy across an arbitrary surface in a theory of gravity saturates the Bekenstein bound. The reasoning of {{cite:2391e56ac338cbaf3ae07a7ce2c2b09cbf3e8049}} is initimately tied with the identification of bulk entanglement with target space entanglement and therefore differs from these other papers in an essential way.
| i | 5fc34ba6f45a1bac8f34997e752901a0 |
The results of the ablation study on the KITTI dataset are shown in Table REF . We can see that the backbone Monodepth2 model {{cite:5de2838f4ded0999b64a2a5cf5fc164e38b7fbbf}} performed the worst without any of our contributions but by changing the architecture to a Siamese encoder- Siamese decoder, the evaluation measures steadily improved. The reason might be that fusing the complementary information between the stereo image pair gave the framework higher chance to generate accurate predicted depth maps. Our MEA and OT modules were all incorporated in the SE-SD architecture. Row 4 shows that the addition of MEA benefits the depth estimation performance in all the evaluation measures, especially on metrics that are sensitive to large depth errors e.g. RMSE. The significantly large improvement of the SE-SD architecture with MEA, is likely due to the epipolar constraint, which allowed the network to learn strong correspondences limited on the same epipolar lines in the rectified stereo images. The impact of the OT-MNL is presented in Row 5, compared to the SE-SD we still can notice a dramatic increase in most of the evaluation metrics. The reason might be that the optimal transport algorithm further improved the MEA by increasing the correct correspondence weights, merging the semantic features while suppressing outliers. In the last row, by combining the backbone with all of our components, the effectiveness of the final framework was significantly improved, as expected, and state-of-the-art results were observed. Besides, although our OT-MEA module was inspired by the MNL, our results outperformed the same SE-SD architecture with MNL. Apart from the performance evaluation measures, we also estimated the number of parameters for each of the examined settings. While all of our proposed components contributed to the overall performance in the self-supervised depth estimation task, the number of parameters was barely increased. We can see from Table REF that our OT-MEA module cost 0.6 million (2.0%) additional parameters compared with the pure SE-SD architecture.
| r | 83b4a18bab5a8fe55de0c3c6709f934b |
E-FT: As described above.
E-LwF {{cite:04a0050e882c01082d37b3cda517b80cb38ed678}}: It aims to guarantee the output embeddings {{formula:2df72b8d-0b90-4636-be12-13372c25c0c4}} of the models belonging to previous tasks is similar with the output embeddings {{formula:96aa87c7-bdd7-4765-845c-9c7a23853e09}} of the current model when given the same input, which is achieved by constraining the parameters update. This leads to the following loss:
{{formula:390b19e6-584a-4458-8508-7b616bbbb235}}
where {{formula:b5c86ab0-7df6-4020-8bc4-774c88de4a15}} refers to the Frobenius norm.
E-EWC {{cite:5c3f99054d276a47ec59f471e9e7e182cd26c3f4}}: It aims to retain the optimal parameters of the former task during current training process.
The objective function of EWC is :
{{formula:0284c967-0604-4c97-aa1f-de4718c8b6fa}}
where {{formula:233d5dd6-b901-4784-80fc-9ed1773e9fc5}} is the Fisher information matrix computed after the previous task {{formula:869d5a27-b78a-4ce8-b66b-fbb1122c638e}} , and the summation goes over all parameters {{formula:3bf02e1c-64ca-4397-93db-342a179980c9}} of the network.
E-MAS {{cite:2a9b460047af87c667f90f159f7005dd60b0f2e9}}: It aims to accumulate an importance measure for each parameter of the network based on how sensitive the predicted output function is to a change in this parameter.
The objective loss is denoted as :
{{formula:fe97e61a-be88-45e4-a59a-12beff4c122c}}
where {{formula:aa00875e-5eab-46c6-a8ff-ee326d83dc99}} is estimated by the sensitivity of the squared {{formula:9c6d462e-ab70-422b-9306-a557e9bd47a2}} norm of the function output to their changes.
These losses can be added to the metric learning loss to prevent forgetting while training embeddings continually:
{{formula:d9be36fa-927d-4d23-82ba-883efff3d3ae}}
where {{formula:2a9c59ed-fe2e-4b7e-b1cb-d189ad8ab0e0}} , {{formula:25c2c6f3-7c72-4ab2-ba5b-540921c81b20}} is trade-off between the metric learning loss and the other losses, which is set to 1, 1e7 and 1e6 respectively.
SDC {{cite:83822818fe4676a37c6eb0b950631fcafccba8ea}}: It aims to approximate the semantic drift of prototypes after training of new task.
The method is complementary to several existing incremental learning methods to improve the performance further.
| m | c60666dcdc87bb7db7bae1e351995614 |
Recently, {{cite:7c7f85db901c6095a14ccd3c98f39ca116a52912}} derived a slightly improved uniform stability bound that implies
{{formula:ae0b7d70-10c2-427f-bea9-7dc2ec010917}}
| r | 3781681dc36bfe436202ef8baca92503 |
Sequential decision making tasks are most commonly formulated as Markov Decision Problems {{cite:501b478ab948ab56c20cfab5d33ac3573825f917}}. An MDP models a world with state transitions that depend on the action an agent may choose. Transitions also yield rewards. Every MDP is guaranteed to have an optimal policy: a state-to-action mapping
that maximises expected long-term reward {{cite:6600a4748f9ae2aa12ffcd7542cd2e8778c491e3}}. Yet, on a given task, it might not be necessary to sense state at each time step in order to optimise performance. For example, even if the hardware allows a car
to sense state and select actions every millisecond, it might suffice on typical roads to do so once every ten milliseconds. The reduction in reaction time by so doing might have a negligible effect on performance, and be justified by the substantial savings in sensing and computation.
| i | 1bbe19eec802c636989d7f1e3f6daf22 |
The qualitative results and quantitive results on the KITTI Eigen split are shown in Table REF and Figure REF . In Table REF , it can be seen that the proposed H-Net outperforms all existing state-of-the-art self-supervised methods by a significant margin. Compared with other approaches that applied direct supervision signals (supervised methods), the model was still competitive. As our H-Net takes stereo image pairs as the input, in contrast with {{cite:5de2838f4ded0999b64a2a5cf5fc164e38b7fbbf}} and {{cite:4205d50c642e78f48523442f6e05f1e6025da88b}}, we did not need to remove static frames. However, to make the comparison fair, we used both full Eigen and Eigen split dataset to make the dataset consistent with other methods. Among all the evaluation measures, the best and the second best ones were produced by our H-Net model, which indicates that our model can learn from the geometry constraints and benefits from the optimal transport solution, achieving state-of-the-art depth predictions. For the quantitative results, we can see that the depth maps generated by our model contained more details, i.e. the structural characteristics of buildings, protruding kerbs, bushes, and trees. Besides, our model could effectively distinguish different parts of every object, for example, the upper part of the tree is no longer uniform but is full of outlines and details.
| r | 1da38daa94cd5cc667014b8bae9c0fce |
Readout function. The readout functions are widely designed by statistics e.g. min/max/sum/average of nodes to represent the graphs {{cite:e615dc2679b0423bdd4aa6ca4e4684b69b922cf1}}, {{cite:9de1b8fe43285fdab5fe6218ea62339c1cfe8621}}. On the basis of global pooling schemes, SortPooling {{cite:345d15115d71889166ec8debb6e2818cf88423db}} chooses the top-{{formula:9f98e30d-a65f-4b83-bb15-d6d134f3ad70}} values from the sorted list of the node features to construct outputs. Instead, in this paper we propose to adopt a virtual node to help explicitly select the nodes to drop out. Interestingly, the term of virtual node is also called and used in VCN {{cite:cdd80c117c8b6bda33e2cabf6a8161a86b7296cf}}. In that work, the virtual node together with the real nodes are fed into an RNN for extracting the embedding of the whole graphWe are a bit in short of confidence to rephrase the exact structure of the network for graph feature extraction in that 5-page arxiv paper {{cite:cdd80c117c8b6bda33e2cabf6a8161a86b7296cf}}, whereby the details are not well described, and no modern GNN structure, but instead RNN is used..
| d | 18e98c7835a6ca038077cfde52e8ec5c |
NN model compression methods include not only techniques for pruning {{cite:20e365fc5478f4f3dbf7e54a2161980d30d8227d}}, {{cite:7973f8fa1f0d88532dc3b18647ff9ec808f9778d}}, {{cite:79bbb8c4d784fe89f52e3bd276548feca547eb2e}} or quantizing parameters {{cite:5759627c97201d84258db6f15198eccd45bd77b1}} of a trained model, but also designing models that are optimal in parameters {{cite:b64ef4bc5f06659accde6a6b9b3830cb48d4afa1}}, {{cite:dc650bf5389357c470fe4cb8a68938f97e2cb33c}}. We observe that the proliferation of deep CNN models has matured along two directions as, standard networks along improving accuracy of the model in general and edge networks along explicitly targetting inference in edge devices while maintaining acceptable accuracy. The standard networks {{cite:35a720af488e1b9917d227ecd0d6580490417c97}}, {{cite:ae0de0f36168ffafd262477cf324f1424ab54971}}, {{cite:41b345fa7de95a874bc36668b8721f849043c450}} employ standard convolutions in their basic modules to create deeper networks. While the edge networks {{cite:c1c250942e3b9a57c7c1bcb26d0ff6583ec210de}}, {{cite:dc650bf5389357c470fe4cb8a68938f97e2cb33c}} replace the expensive, standard convolutions with depthwise (DW) convolutions. Also, they sandwich DW convolutions between pointwise convolution ({{formula:8ddcde64-8a24-4057-9c7a-666fc5aefd47}} ), forming bottleneck blocks to control the number of channels fed into them, thereby maintaining the parameters in the network.
| i | 0ab5da26369bc6f6d78f40b6678599d6 |
With this definition, a maximal coupling {{cite:51f0cc47feb118d46195957a180e00ac8b8e26eb}}, {{cite:395fae925b8ddba9b4c4114c55f88418320bb4df}} is a diagonal coupling with maximal mass on the diagonal event {{formula:e6ed217e-7b94-4872-85b9-cb9b22f45095}} . The maximal coupling also has a connection with the total variation distance between the distributions {{formula:e20b8096-39f7-4643-84a2-74b8b88f47d7}} and {{formula:1e7821c4-6d4b-4583-a54c-2ef435f66c12}} . If we define total variation distance as {{formula:0da39642-5a3f-49dc-a34b-d0104a7180d8}} then for a maximal coupling {{formula:37e76bc1-3c3d-4d33-bf16-61fb816171f3}} , we have
{{formula:2c870dd8-8cd5-4d99-a916-5986a2e99bd0}}
| m | 9f75a450d47241a6425cb97b2b2c111d |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.