text stringlengths 54 548k | label stringclasses 4
values | id_ stringlengths 32 32 |
|---|---|---|
Weighted Data Sampling The long-tailed nature of the PBVS-MAVOC challenge dataset makes it very hard for the network to recognize and predict the tail class images. The network over-fits on the head classes as the number of images in the head classes are approximately more than 10 times the number of images in the tail classes. Long-tailed representation problems are therefore addressed mainly through data re-balancing, data re-sampling{{cite:1c3f217ff17b73c26f27c848cca4cd5f5432bc3e}}, re-weighting{{cite:29e8215ba713ffd001fe64acdec8d7e46c6dbec4}}, data augmentation{{cite:113d5899a643b6bafd70b606614f79e61385d461}} and two-stage training methods{{cite:3d195079e51dabb96ada9cb4d1bee7f4a50e6d50}}. Therefore in our work, several image augmentation techniques were used for the tail classes.
A training dataset for both the domains has been manually curated with around 7000 randomly selected images per class for head classes and 5500 images per class for the tail classes using augmentation techniques like random rotation, horizontal and vertical flips.
To account for the slight class imbalance in the manually curated dataset, a weighted random sampler was used with the weights(probability of an image being picked from a class) being equal to 1/{{formula:662a33f4-cb6b-4bb3-99df-5d1667519f36}} , where {{formula:cb7cc29e-2b68-4b45-90c1-bbaaa0ac92f9}} refers to the number of samples in the {{formula:59ea40e0-faed-4252-bb80-76aa89cc85f1}} class.
| m | 11179a17ce05cf29cf914789d7e4fc47 |
Considering the {{formula:3c8a31f5-73c2-452a-9fb5-3605ac81c1a3}} images, a texture analysis method based on entropy is proposed to quantify the patterns of different activities profiles. Entropy is a statistical measure of randomness and is formulated based on Shannon's equation {{cite:53b0e3d5fee87efd0f6138fafb8501b6e76ee660}} as follows:
{{formula:3e0db457-909c-48a4-bd3f-e2c918f7297e}}
| m | 52be1975316b827b73c3832268c31d1c |
For model-free methods, we can first observe that the overall performance of the SOTA methods is much worse than their performance on in-the-lab datasets like the CASIA-B {{cite:ecee290065a2e5ade7b215a29fb6fad643d28a54}} and OU-ISIR series {{cite:89b97d5e6b09b434e20e6af81c4cb580aa5a3a9f}}, {{cite:ab29a3657aed63ebe29d5ea0b4343bbc8b5cd17a}}.
This reflects that there is a huge gap between the in-the-lab research and the in-the-wild application that is much more challenging.
Meanwhile, the performance of the SOTA model-free methods varies significantly.
For example, the GEI-based method, i.e., GEINet obtains the worst results, which indicates that the GEIs discard too much useful information for gait recognition.
Moreover, the methods considering the order of the frames in sequences, i.e., GaitPart, GLN, GaitGL, and CSTL, obtain lower accuracy.
It means that the temporal information in the wild scene is hard to learn, because people may stop then continue to walk at varying speeds and routes in unconstrained scenarios.
On the contrary, the methods considering frames as an unordered set, i.e., GaitSet, obtain better results.
| r | 77c5aa3478d7c2747a76a0243de9cb90 |
Cell-free massive MIMO has been recently recognized as an alternative to co-located massive MIMO for future wireless networks owing to its substantial improvement of connectivity, spectral and energy efficiencies {{cite:24dcdb3146bbf8f232bfc96e4fc11143a3b13886}}. In cell-free massive MIMO, there are no cell boundaries and a large number of access points (APs) are distributed over a large geographic area and jointly serve many user equipments with different speed profiles. While a large body of research has delved into cell-free massive MIMO, they mostly consider flat-fading channels. A few recent works in the literature focus on the performance of cell-free massive MIMO over frequency-selective fading channels {{cite:2dddb6f641d1b10d27e64a446b0abad229795148}}, {{cite:31663409f2fbed9bf6a02642ab6306e03d1b89ee}}. In {{cite:2dddb6f641d1b10d27e64a446b0abad229795148}}, the authors analyzed the uplink achievable spectral efficiency of a frequency-selective cell-free massive MIMO system under the Wiener phase noise process and with single-carrier transmission. The work in {{cite:2dddb6f641d1b10d27e64a446b0abad229795148}} has been extended to multi-carrier transmission in {{cite:31663409f2fbed9bf6a02642ab6306e03d1b89ee}}, where a user-specific resource allocation method was proposed. However, the work of {{cite:2dddb6f641d1b10d27e64a446b0abad229795148}}, {{cite:31663409f2fbed9bf6a02642ab6306e03d1b89ee}} cannot accommodate high Doppler spread applications. In order to serve high-mobility users with time-variant channels, the integration of cell-free massive MIMO and OTFS modulation is expected to further improve the network performance. To the authors' best knowledge, the consolidation of OTFS modulation with cell-free massive MIMO has not been reported before. Thus, this paper will focus on the downlink achievable rate analysis of OTFS in cell-free massive MIMO. The main contributions of our work are as follows:
| i | 97f334861be7063e7f8196580751883b |
We compare our FASeg with state-of-the-art semantic segmentation methods on ADE20K val {{cite:adb34ffd531484d96fc676a075b8808c3b88376a}} (Table REF ) and Cityscapes val {{cite:dc25fa666df91dd28b915c13e5b8255ac5cbadf1}} (Table REF ). Specifically, we show FASeg with two {{formula:54a1d542-6490-4aad-925f-c38266c0d975}} implementations: learnable absolute positional encoding {{cite:74e3151f3e09d504a99c074817f07e559a0d3647}} and conditional positional encoding {{cite:a9d8c0d75a55e7fcacc30834ae3650e2c7f3a274}}.
| r | 28df7edbc098ba44f8dfb6afa726f968 |
To reduce variance while preserving the stability and convergence properties of on-policy Monte Carlo policy gradient algorithms, various variance reduction techniques for policy gradient methods have been investigated in the literature; with the goal to reduce the variance of the gradient without introducing bias. The subtraction of an appropriately chosen baseline, both state-dependent {{cite:609189e8b1fd21e79ef2e0de29825b38cec71760}}, {{cite:476f55f541f5f10461f405199a61e38922c2496e}} and action-dependent {{cite:f64ef7cde9ea5f87f982d097b0605b043a479b69}}, {{cite:4b743adac9a7e885623fee8b320cd46d18a0b789}}, {{cite:38799249b9d26de4a6f1f4940ad9587cd77f28e2}}, {{cite:77b935484cf7e82073bb263f115bfa3c8971ce3f}}, {{cite:3ec8e6f15f21de5408c1decd85b74decf88d7d91}} baselines, for variance reduction in policy gradient have been studied extensively over the last two decades. Use of action-dependent baselines and their effectiveness in reducing the variance over state-dependent baselines have been subject of much debate {{cite:c648725809539ee439a5f234fd02a6b504d6c72f}}, so we refrain from considering action-dependent baselines in this work. Although an optimal state-dependent baseline exist {{cite:535474594658bb07cfa33c59bac5ff7051e48cf9}}, it is typically hard to find.
| m | 4dd04dce6c2b0f5b39f73f9dc5c2dc39 |
Diagnosing entanglement in quantum materials out of equilibrium would be particularly impactful, as ultrafast lasers have led to the synthesis of nontrivial states of matter without equilibrium analogues {{cite:cce99f9895c09d2f526067369d0b538caf14d3b9}}, {{cite:70b78e55b006f1724402f521d65ece2ee05b229d}}, {{cite:9507dc7bfbbb548e75c2e5831008f4663495f2dd}}. Dynamical entanglement has been theoretically demonstrated in the wavefunction of solvable toy models {{cite:f18d7b42acc0cbd4321aa32b2ebd46f13ffcce35}}, {{cite:cc9ba91a2c9b1a1a5beeb12f0bb65a68fd0f341f}}, {{cite:3c429abe08676994fa974497e128b276cd442d5c}}, but so far it has been unclear how to experimentally characterize this phenomenon in pump-probe spectroscopic experiments. Inspired by quantum metrology, entanglement can be quantified by the fluctuations of local probes at a given distance or a finite momentum, and the recently developed technique of time-resolved resonant inelastic x-ray scattering (trRIXS) technique {{cite:d6ee22c523b35305c9b9a76059e60140cd6cfe9f}}, {{cite:f8a2ca574d8512cd3c77160b424547eebc99c6dd}}, {{cite:13924ce9867a7d5395881b2ed3d55de9066332bd}} — sensitive to collective charge, orbital, spin, and lattice excitations {{cite:d6ee22c523b35305c9b9a76059e60140cd6cfe9f}}, {{cite:af7779f7974b22afdb63a8813405fa6ee137e8e2}}, {{cite:a36af36afc58de2df97b87c2e34b77fed1e6a258}}, {{cite:c154653ebec2e4cda455f0048d5b555895a3236d}}, {{cite:0a4cc8ff68ffe8213866b45136f406f75e4d0ec5}} — opens new opportunities for diagnosing entanglement in light-driven systems.
| i | 67b91c504c90c94a0231974414c0952b |
Automatic Adjoint Differentiation (AAD) is a rapidly-growing field with a variety of applications ranging from biology {{cite:4d9e3012dc49ad1f80eefc0c0783825682d82f42}} to machine learning {{cite:4ed9786318432a44f317f8b112bbb4500fe5503f}}. It has recently become popular in financial applications for calculating sensitivities (first-order derivatives, the gradient) of functions automatically alongside the computation of the function, see e.g. {{cite:f9f81aaf1a5b55045320010062d7ad17bef353d2}}, {{cite:5ad9ec8750c70ebb6e3becf5b97de2cb3218f508}}, {{cite:54653469d9b253a986f63f5b92e7ba320c3e0dbd}}, {{cite:0d775428777c784c69b8cb09f4e4494892aad1ac}}, {{cite:168104f72e9d523b8966654d7a8d0250246347c3}}, {{cite:d8ee26c6662b85d72c35e9b251d6e21895518666}}.
| i | c29ce7b12ad6e52b56f2ec308819ea9a |
In our works, we use the Tensorflow 1.13.1 {{cite:30113313c602eb6b649950b03b5945dc653b8ce6}} for implementing our algorithms. We deploy the UAV at {{formula:0e31d8d1-2c7d-4085-8bcc-fc890793ad40}} , the RIS at {{formula:86026878-a976-4055-8cee-203da12bc4e7}} and assume {{formula:d4939e63-1750-4d44-a23e-4e386311e654}} for convenience. All other parameters are provided in Table REF . In order to compare our proposed model with other baseline schemes, in this paper, we consider the techniques as follows:
| r | 691ed285e1137afdb0f511141a348e67 |
The structure of rationale models guarantees that causal relationship between the rationale features and the model prediction, but this does not necessarily imply its usefulness to model understanding. Specifically, it could highlight {{formula:4385c5ad-6ffd-41da-98c1-9ea6a5d6a50c}} only barely, while including lots of non-correlating {{formula:06d13e77-b4fe-4fd4-a47c-6d106f80fedb}} (and, in particular, misleading words such as the non-correlating articles)There are additional concerns on the unfaithfulness of rationales as Trojan explanations {{cite:86f30b4dce6671953cc5f0fa546b651b073a7c59}}, {{cite:01b9b733db2c9ed8bbb95cac04270f3f814f984d}}, but they were not identified in our experiments.. Indeed, our results show that rationale methods are prone to selecting misleading non-correlating features, which obfuscates the model's reasoning process by giving more but unnecessary information to the human. The problem is more severe with RL training, possibly due to the known difficulty with REINFORCE {{cite:6d28d58c7df2b7a8313bd9783f5a8870287c5660}}. Post-processing methods could be developed to further prune rationales to mitigate this problem.
| d | 65a788d1574e8015318533cb19750710 |
In order to demonstrate the applicability of the results using NRF on a larger scale, we perform an experiment on the ImageNet-1K dataset {{cite:7ecd29e68d1ba3078c27d2c49e043f73acf517a7}}. The dataset contains around 1.2M training images from 1000 classes. A linear classifier on the original input images (150K features) achieves a 3.4% test top-1 accuracy. In contrast, a linear classifier trained on 4096 NRF, sampled from randomly initialized ResNet-18 models, achieves a 10.3% test accuracy. This result is fascinating given that NRF achieves a significantly higher accuracy with 37{{formula:f573cc6e-341f-4569-b16b-000aeec611f0}} fewer features than the original input features. Training a linear classifier and a two-layer MLP on 31568 NRF achieves 12.2% and 15.2% top-1 test accuracies, respectively.
| r | 818930676a44844e759196a9c2907adf |
In {{cite:6741d1c35dec7ad187f78c837e5298e8181edcca}}, the time-dependent Hamiltonian simulation problem (REF ) is addressed using the Dyson-series technique, giving query and gate complexity that scales with the {{formula:59b13896-6de4-4aae-aa84-a59999cc6cb8}} -norm of the interaction Hamiltonian {{formula:cf948f4c-1289-46c5-84be-3f3df7b29fbd}} {{cite:6741d1c35dec7ad187f78c837e5298e8181edcca}}. This can be improved to scale with the {{formula:bc79c92d-90b8-42bf-92bc-f4f0a0f85669}} -norm of {{formula:6ffc6b7e-4bdd-435b-a982-5e7b78bc9ffb}} using the rescaled Dyson-series algorithm {{cite:bff28df4cefa1891a6cd69536edef7c186601a0a}}.
| m | b9f17e5f80416ac774fca753abab255c |
Fig.REF reveals the influence of the pump-wave intensity ({{formula:4e3f2baf-7ba0-4e17-847c-a7a8c540c006}}{{formula:493027f4-a54a-4a94-8163-4836c76d677b}}{{formula:abf0c90e-389c-41a6-89f7-ca59d1d52949}} ) on the resonance parameters. The power broadening of the resonance is shown in fig.REF a. The zero power limit gives FWHM{{formula:879c7215-dc4e-4db2-be2b-0b8184e1f315}} mG (60 nT). The linewidth dependence does not show a significant difference between the regimes with {{formula:490ecb85-d5e7-4b5d-a032-0273ec86c114}}{{formula:c589d421-4a8b-41bb-b40f-ad5c675596e0}}{{formula:d3a77ae0-b3fd-45be-bb55-cdb5fa3d724d}} C and {{formula:9dad5648-9e2b-48b9-9acc-cc81269b8cf8}} C. However, further increase in the temperature results in additional broadening, because the medium acquires large optical density. The temperature increase, on the contrary, helps to observe an extremely high contrast of the resonance, reaching {{formula:baecdabb-6c50-4019-a78b-cc129613571b}} (see fig.REF b). The competition between the linewidth and the contrast dependencies leads to the formation of an extremum in the contrast-to-width ratio (fig.REF c). This extremum shows that the optimal pump-wave intensity is not high ({{formula:c8c7476b-ac2c-4a83-a67e-5f52d428f123}} mW/cm{{formula:e335c2da-358b-445f-92a9-d9d5ec428831}} ). We should note that the observed CWR{{formula:3e7f3cfd-475f-4983-9a5d-728d69608f61}} %/mG is relatively high, especially taking into account the cell size. For instance, in {{cite:60fd16047c692fe4e54241f4731b4284553883a8}}, the authors measured a CWR of the EIT resonance of {{formula:d15f1832-1ba3-4096-bb05-a7c6dc501289}} %/mG in a 5 cm long K buffer-gas cell. In {{cite:1e8ed5f6dd4dc62f6ce795ce36c4ed9c317f0c2b}}, the authors reported observation of an EIA LC resonance as narrow as {{formula:9bbb16d1-0a0b-428b-9b38-b448aa241eab}} mG in a 5 cm long Rb vapor cell with antirelaxation coating. In spite of the small linewidth, a contrast of only {{formula:49a46316-fbd0-45ef-aeb6-08aaaae1fe07}} % was achieved, yielding CWR{{formula:3231f63c-e099-427b-8627-d8db55c8e1fd}} %/mG.
| d | 8d776eab015bb372022fa20fd97a65c2 |
Due to computational restrictions, we were not able to run our method with RoBERTa-base and the same configuration proposed in {{cite:fbd481abf9d24f132d9130597f28b38fe4cb64ca}}.
However, we should expect some improvement when using RoBERTa-base (110M) compared to BERT-med ({{formula:4e6151a1-cb41-41a5-b3ed-c55542cd5dfb}} M), as bigger architectures, trained on larger datasets and batch sizes usually come with performance gains in supervised contrastive learning {{cite:1b45f677c52908ca7b448f977433a0aa5e5709cd}}.
Moreover, data augmentation in NLP is known to be hard, so we do not apply it given that using labels, i.e. product identifiers in our scenario, reduces the importance of data augmentation in contrastive learning {{cite:61d6aa0a250fcc304c35eb93e3f8f25d697bebe3}}.
| r | 84c8faa7a803f5733b5cd19fa32745d6 |
xg(x1)-xg(x2)Lgx1-x2, x1,x2X,
where {{formula:cb0a5a97-6756-4e06-826b-ff58b6f44ce2}} and {{formula:88bb3470-c0af-4af3-8b07-c9ce39a876f7}} represent {{formula:a115f746-9b76-4730-ba12-d29c7c3d1000}} and {{formula:35d7be56-d989-410c-af9f-e750d007d04c}} , respectively. Note that {{formula:e1baa983-ceff-40c6-b6b8-e62acd54e15e}} and {{formula:8faff4e0-4d81-45d8-b587-4bf49895c1e3}} (see {{cite:e49119cbfe382830d8dfaece220a11ee6efa3a1a}}). Furthermore, (REF )–(REF ) in Assumption REF imply that there exist constants {{formula:b2dbbe82-6a57-4f2b-8b18-ffab0043d01a}} and {{formula:53b11d5d-6f58-4364-b49c-ab6021fe08fa}} such that
x(x1)-x(x2)Mx1-x2, x1,x2X, 24c
| r | c8d65c04765e462e234b5b73ead02897 |
In the following, we show how to obtain a ready-to-be-implemented version incorporating ideas from the semi-naïve evaluation of Datalog programs.
Semi-naïve evaluation of Datalog, as described in {{cite:9dbe911605a33d1df322badc7acb16abb8eb5ed5}} introduces a number of ideas aiming at improving the efficiency of the naïve Datalog evaluation method;
we show how to leverage these tricks in our setting.
| m | 790be68c7c64105159bf2f0ad6ca5bcf |
The delay embedding dimension {{formula:16756fd3-14fe-43ef-b951-c65b82f7ae77}} and the time step {{formula:3d52a121-3c15-45f3-99c5-9fa7bfe2ff51}} are chosen to satisfy the equality {{formula:f0853a56-13a3-4f63-8cfc-9636dbd2f1d0}} , such that the delay embedding of {{formula:c877daaa-017a-43e8-b46c-c26c80057d51}} is most unfolded in the embedding phase space (see Appendix ); this choice is also consistent with that from HAVOK {{cite:10c9077c2fd708308b2ca2a964b687e64b6e6b6f}}.
Along with {{formula:058ae219-c49b-494b-a4aa-bcfd3e5d7201}} and {{formula:3c9c5c18-46bd-43b0-80a8-1e446ec62ef3}} , the hyperparameters over which we optimize {{formula:8fbd1ecd-af55-4924-a2a2-771a61e613c0}} , {{formula:03245003-54c2-4134-95d7-40aad83f547f}} , and {{formula:da3356ba-de15-4f02-afe8-87d88174f99d}} are the SVD rank dimension {{formula:3f8045dd-1663-4716-9498-edc6d974c5e3}} , the dimension of the latent variable {{formula:12a2a9a2-863b-4d08-a142-cce190aca859}} , the architecture of the neural network (number of nodes, number of layers, activation function etc.), and the SINDy library (polynomial order, trigonometric features, threshold value, etc.).
The high dimensionality of the hyperparameter space adds to the challenge of the optimization problem, which we address in the current study by running a hyperparameter search on parallel GPU clusters.
| r | 4495e0f93f21419f7e83954b117a9d37 |
These methods, which are visualisations of the spatial variation in importance of input images with respect to prediction, provide post-hoc explanations. These methods can be categorised into three types namely function, signal and attribution visualisation. These groups of methods present different information about model prediction that are complementary to each other{{cite:d9b86876013b02b5c31235dd9351b21b3546920d}}.
| m | 88d161e3205dea837305b036ec15b828 |
through an extra requirement that both {{formula:a6d48ef4-2381-4a7b-b881-2447c74a7218}} and {{formula:94a1f952-4ae7-4d8f-8965-b2a7c68a3eb0}} can not be too large.
This can be readily observed on the Petersen graph as shown in Fig. REF ,
and also implies that the anti-Cheeger cut problem may be harder than maxcut.
Actually, the NP-hardness for the anti-Cheeger cut can be reduced from that for maxcut {{cite:b2ccfac61cd67b3c1e1395aa1e8a7b40adea832f}}, {{cite:de930dab0dd0b112ca08069436fd5f615a40c11b}}.
According to
{{formula:61b8d209-ca7f-4e32-b98a-223aa5b43f69}} for any cut {{formula:6ebd2202-1825-4c64-b82e-14f84ca0c219}} of {{formula:7a64d008-f392-4cc0-b9f5-c4d7d930a696}} , we have
{{formula:add977d2-1d91-48e4-895e-a964949c2b3b}}
{{figure:7265b941-3dc3-4d69-98d0-2efde93a4462}} | i | 47e832df63086350e46c7e32189bbe15 |
In case of Question Answering (QA) task which is one of the promising areas in NLP, however, models outperforming human on SQuAD {{cite:32e5b8c4a98d3f6e13d0c65b7f1c2c9074556a70}} cannot generalize well to other datasets. Models rather overfit to a specific dataset and require additional training on other dataset to adapt to new domain {{cite:fcbf19e9863005a1ea252af17496b5b9cf1937a2}}.
| i | 6a457d98fb534f81d4733ae55f6114ea |
where {{formula:4bcb82dd-d2d8-446b-afa8-b633f9f94853}} the validation set, the subscript {{formula:93697de7-1398-420f-9061-5387dd3fae5d}} denotes the class index.
After {{formula:713234bd-8151-4a89-9a5b-fd16f7d8a955}} is trained, we close the gradient computation for {{formula:db8b3317-32db-4012-8ddf-8207ff86ead7}} . We then train the calibration network {{formula:a31bbf7f-f105-455a-be68-abc71d6510a4}} using the negative log likelihood loss that is commonly used for training post-hoc calibration networks {{cite:92390f13dcf7379dc97160c1d0a48ad0c5054073}}, {{cite:fa75f48fa5645e8b74151365ef0c83aae1a6cc21}}, {{cite:dcfb8861095a0a8bc76230fcc9c4208331147fe3}}:
{{formula:04089932-a91a-45b4-ad98-da203b0ed5b7}}
| m | 8a5fe64486ca667977ae08f361d97809 |
In {{cite:0933bc06707dce20fef77ff2c19e3eb5dccd8a2d}} the author proved the following result, which can be seen as the analog of Hamilton's identity for Ricci solitons. The latter plays a fundamental role to Ricci soliton's theory, as one can see for example in {{cite:4ef4b441b8e65598ea29a272325ef70d7b0a978b}}, {{cite:0d6384f8c4abd5af2a23000513ccbf1b7e8d2639}} and references therein.
| r | 53103e6a453a5f971662047c8ed5bcdc |
We present a framework called spatial prior attention (SPAN) that builds upon self-supervised Vision Transformers {{cite:1e864f01b4214691fa2a0d8b35d55c5891bc09fe}} and leads to more interpretable attention heads, higher downstream classification task performance as well as performance improvements when little annotated data are available – see Figure REF . SPAN incorporates prior knowledge into self-supervised training of Vision Transformers by regularizing a subset of attention heads in the multi-headed self-attention module to attend to objects of interest. Instead of relying on human annotations of these images to determine the boundaries of these objects, SPAN leverages the shared underlying structure to define a template of object boundaries (e.g. individual organs or parts). It then aligns and registers each image to this template, transforms the object boundaries accordingly, and then uses these obtained object boundaries to regularize the attention heads throughout the training procedure – Figure REF provides an overview of this process which is detailed in Section .
| i | a7a44b37bbbf208d599743fef168fb89 |
We present a scheme to completely distinguish 64 hyper-Bell states in spatial, polarization, and time-bin DOFs assisted by QD-cavity systems. In the present scheme, the time interval {{formula:aa16c902-96fb-4fad-831d-9c71ed0294a4}} between the two incident photons and the cavity photon lifetime ({{formula:071ede15-8b53-4d46-a922-f1f3b3670ada}} tens of picoseconds){{cite:a986728c0040126fab577aae9d808fcd39f7f4ad}}, {{cite:49067c7a69cb3b81de8903d04869198feee652c1}} is much shorter than the spin coherence time in QD ({{formula:c689d290-42a3-4cf0-975a-d1d81c06f943}} s).{{cite:d30da76c8add78e4c839e8f84c162d39de225e33}} Manipulation and readout of the electron spin in QD has been demonstrated with high precision.{{cite:a986728c0040126fab577aae9d808fcd39f7f4ad}}, {{cite:c7b3f2684940dc518614a8a1d7d2f89d4aad8ea6}}
Unity fidelity of our scheme can be achieved with efficiency
{{formula:34a5b686-8d23-4125-9b15-69502092d1c7}} . Here the fidelity and the efficiency are defined as {{formula:4e725de6-a2e3-42c0-a3ac-3d9de2f7c931}} and {{formula:7c574d88-b3c7-4816-8537-19c34f89809e}} , respectively. {{formula:08897a0c-096a-4195-be6d-d77eaaaed937}} ({{formula:9d65f7a4-e4a8-466b-a916-af4718646a36}} ) is the realism (idea) normal output states of the system. {{formula:a3fa6dea-9e58-49d8-a6c8-cb52d73fc3b8}} ({{formula:f0180085-50fb-4820-8bfc-2a84e0bd5d28}} ) is the number of the input (output) photons. From Figure REF , one can see that the high efficiency of our scheme can be achieved by optimizing the QD-cavity system, e.g., increasing the QD-cavity coupling strength and suppressing the cavity side leakage.
{{figure:8c3749e5-a016-45f3-ba39-1fd22fc309f3}} | d | c5005048ba129bb895be1b6f6bd18046 |
On the other hand, in a recent follow-up to their earlier study, {{cite:5cafb581adf131bb27e1e898fd852866b61cf9cd}}, argue that the learning curve for the case when very large models are trained to convergence (so that performance is bottlenecked by amount of data) poses a lower bound to the optimal-compute learning curve. They therefore suggest that beyond a certain compute budget, the data-limited learning curve determines the long term evolution of the system. If this hypothesis is true, it would mean that the kind of learning curves which we predict in this paper are the ones that asymptotically determine the fundamental limits of learning with DNN, even when optimal use of the compute budget is taken into account.
| d | 414dbfd1b830d56cf7601eb1fcbb67c6 |
We adapted the inspiring work
by {{cite:8eee5261b0238a6f64b55ca690f84cda47c8757f}} to work better in our setting and efficiently learned expert
models for non-IID client data. Sending all cluster models in each iteration introduces more
communication overhead. We addressed this by removing converged cluster models
from the set of selectable cluster models in alg:fl.server, although
this is not used in our main results. This only affects the
result to a minor degree, but has a larger effect on training time due to
wasting client updates on already converged models. Another improvement is
the reduces complexity in the cluster assignment step. A notable difference
between our work and IFCA is that we share the all weights, as
opposed to only the last layer in {{cite:8eee5261b0238a6f64b55ca690f84cda47c8757f}}. These
differences increase the communication overhead further, but this has
not been our priority and we leave this for future work.
| d | 4e8efa366d5517332e92969c601dba70 |
The starting point to this paper is the definitions of MSE and bias which differ between fields. Classical statistics define the metrics as a function of the unknown {{formula:3f677c10-0d14-4445-af6b-e0619406492b}} (see for example Eq. 2.25 in {{cite:0c4a8206f82a2fa533d925249ce86ed8d3d9ecfc}}, Chapter 2 in. {{cite:28d4ce7477470707d857550c84bf85918d2382b7}} or {{cite:8f7d3b11235699a227a8b46cff6eb598ba8f02dd}}):
{{formula:ebf1297b-1aec-423d-aad9-14a05ec22d4d}}
| i | b9e8f7f0a6e8439055e5d70ec97a1a98 |
Fig. REF shows style transfer when the content image has more objects than that of the style image. It is worth noting that merging lake object with sky or building would result in a high content mismatch. Neural style {{cite:1c4cc3e647ec084cc171d4e9fc3f341fe2f255fc}} spreads the style features disregarding the object boundaries. DPS {{cite:21702dd039853dc943ae2c5a06892935b4ec8ec4}} does not supervise features of the sky and the lake object well. WCT2 {{cite:b7933280980ba987620b1935ce9f99f68b8020a1}} does not fully utilize style features well, but it preserved better image features than STROTSS {{cite:e3544df5b1e52f0b9738f4936a4025abc4a28e7e}}. DeepObjStyle spreads the style features to all the objects in the output image while preserving the structure.
| r | a014e6b389e94c60d723ad342e4777b5 |
From section 2 in {{cite:e65c2ad7cda00001b626e940b810b13f8a3d31df}}: {{formula:5663061c-ab59-42b2-b5a7-39cfd49124e6}} . Using this we compute the cumulative functional dependence measure of {{formula:db536839-9b29-4f78-ae7e-2cd9bb6a41be}} as:
{{formula:188062b5-85f3-44d1-8604-16e63c1f3531}}
| r | 7b9df97891c2db846519752a81c7b0a7 |
A key limitation of pulse-shaping devices is their ability to generate time-continuous controls. They are therefore usually replaced by piecewise constant pulses in experiments with a minimum sampling period over which the controls remain constant. This hybrid situation is characterized by a digital control which evolves in a discrete way, while the state of the system varies continuously. This constraint was one of the motivations for the development of GRAPE, a gradient-based algorithm directly optimizing piecewise constant functions that can be implemented experimentally. Even if the efficiency of numerical algorithms has been demonstrated in a variety of domains {{cite:b7727123233e0bc1a577ffb5c01600dc7a41e2c1}}, {{cite:8e5d28b2120946d5c2d1f07a686d9d4dee839e72}}, their application is not completely satisfactory because only local optimal protocols are achieved, therefore with no certitude about the global optimality of the control process. This problem can be solved from the Pontryagin Maximim Principle (PMP), which has the decisive property of transforming the initial infinite-dimensional control problem into a generalized Hamiltonian system subject to a maximization condition and boundary constraints. The optimal solution corresponds then to a Hamiltonian trajectory both reaching the target state and minimizing a cost functional such as the control time. The low and finite dimension of this new dynamic space simplifies greatly the search for globally optimal controls {{cite:eb7b301a92966b671bd98a56ecb11c0ad2dcf921}}. This method has been applied recently to a series of quantum control problems {{cite:451cb1590a5a23f5d7c2b0e6c4319cc8a470e7b4}}, {{cite:07de44ceec166c7a15b1a3bd6a82a09d01015c60}}, {{cite:47943b6e4161541d26fd514e42694308d53a245c}}, {{cite:6059853babac7a0d8560ddbea2742891cb9c88cf}}, {{cite:e73faa4823db3efefed7850f5277d4547a81a4eb}}, {{cite:aa434ead99dc751eae722021da95e1cde017cc4a}}, {{cite:d08ab4add891c594f07c6c9cb3453cc9cf6c9af9}}, {{cite:ab06df169464ea223a7291a3a6be1e54566313ca}}. However, as with the original formulation of Pontryagin {{cite:0f69480f96e9fc8cf4002c0424c4f5f7ef9ec492}}, such studies consider that the control can be modified at any time, leading to continuous functions or possibly having discontinuities in the case of bang-bang protocols {{cite:eb7b301a92966b671bd98a56ecb11c0ad2dcf921}}. As this assumption is too strong for standard wave-form generators, the PMP only gives mathematical optima which cannot necessarily be physically achieved in practice. An interesting example in this setting is given by time-optimal control processes in which the goal is to perform a given task as quickly as possible. Such protocols have recently been extensively investigated from the concept of quantum speed limit {{cite:72a98bfb1656dfac02e9f053cc3a41aaed55fa34}}, {{cite:c191cca315ffad98b24b13e78084b0e9e889ba11}}, {{cite:aa8df50ff62d683e7ba2d590d49475bc3b8cfa56}}. Limiting the shaping to piecewise constant controls necessarily increases the minimum time. A key point for the experimental performance of the process is the quantitative impact of the control digitalization on this time.
| i | 12e660ad234fc50a7aa89b68cb7a4350 |
Figure shows learning curves of six individual runs, split into three shorter and three longer runsLocal evaluation performance is significantly lower than on competition evaluation servers for unknown reasons. Other competition participants have reported the same (private communication).. We include longer runs to confirm the agent does not improve with longer training. Most differences between consecutive evaluation points are not statistically significant according to the t-test (two-tailed, {{formula:2b151eca-03e4-431c-b080-a82ff046def0}} )We assume fixed sample mean variance, given the narrow range of means we focus on here. Larger sample means have larger variance, due to exponential spacing of individual rewards., as the variances of evaluations are large. There is also no consistent improvement after two million updates. That said, note the sudden spikes, especially in the beginning and at the four million updates. By stopping the training at the right time, we can jump from an average reward of {{formula:a953aa1e-12b5-40f4-b8db-fa131ddc000a}} to {{formula:cf3b40aa-2577-498c-8cbb-8388a66a70ea}} , a relative increase of {{formula:774f5ed6-d74a-46cb-8055-a44d7ed67287}} (statistically significant change with {{formula:796aed79-23d0-4096-be5b-a38d9d412234}} ). The same was observed with learning rate annealing to zero. This demonstrates the need to also study variances in results with BC methods, as done in e.g. {{cite:88b05c9505745fad9f39d5497cc6e9686cabbbf3}}.
| r | 2f61f341583f6cb0919dd56cace8b657 |
We adopt the definition of an MDP given in {{cite:117eb2463f35d7dcf6a06a21f671bc35b059db44}}. Let {{formula:876b331e-3bc8-46fe-b3b4-3d96f061bd58}} denote the source MDP, where {{formula:8aac5545-9fb5-4f73-a293-2a009165d4cc}} denotes the state space, {{formula:b269609d-2cf1-4cfb-b264-8baf21a82acf}} denotes the action space of the source domain, {{formula:ddd08fc3-4e30-444f-8a16-c7f7ce300ec4}} denotes the reward function, {{formula:69448fa5-ded0-4f29-a4fc-615c6b1864c9}} denotes the transition function, {{formula:94c5f55f-67ed-4dec-adbe-2e313425c442}} denotes the discount factor, {{formula:136d7cff-3678-4ddd-b8c9-18b6bc2c4cd4}} denotes the set of initial states, and {{formula:f149d7cd-958f-4e54-a1d1-e14bdf5010d8}} denotes the set of absorbing states. Let {{formula:44ebba60-7d6e-4c31-8356-bef31f78b756}} denote the target MDP, where {{formula:ad932390-a712-4d5e-839a-33226ce63024}} denotes the action space of the target domain. Reinforcement learning methods can choose to learn the unknown elements of each MDP explicitly to elicit an optimal policy; otherwise, reinforcement learning methods typically derive policies based on direct estimates of a value function. Here, we are interested in instantiations of the latter case involving the use of deep neural networks for the Q-functions. Let {{formula:16fe976c-333d-44e4-8d12-a65cd845d2e5}} denote the Q-function estimator for a source domain with a discrete action space. In this work, we assume that the neural network architecture for {{formula:27988961-92c2-4f26-8fdc-c7111dad3e0f}} employs a fully connected layer {{formula:70dec104-3f64-4f7c-9ec5-525505d5cd94}} immediately before the final output such that {{formula:3b7b62b0-06b3-4592-a8d8-509d822f7c30}} . Let {{formula:c10fc094-0ba0-4261-a7ea-e80962fb6474}} be the features of the last linear layer for input {{formula:24c7b18f-a1bd-4f7c-abb9-08fcd1735cc9}} . Without loss of generality, we define {{formula:bed5c3d1-a4fa-457e-805d-f20ec2cba049}} as the Q-function estimator for a source domain with a continuous action space for actor-critic methods {{cite:f4ce6873e5d1db3b1ac97d016aed8db7ed50ac62}}.
| m | 5878d269069df9b69a5c7a2ee550ae9e |
The performance of SAGAN has been studied with different CFA Patterns (i.e., Nona-Bayer and Bayer CFA patterns) and compared with state-of-the-art reconstruction methods. We included deep Bayer joint demosaicking and denoising methods like Deepjoint {{cite:7c7a86f3740415e8e9e844fe908d50b5de7147ee}}, Kokkinos {{cite:f80545231b4a79b0eb03830ffecf3cf3d10ac357}}, Non-Bayer JDD method like BJDD {{cite:4e429463e0fba838b67262c69d7cd5145d19a5a5}}, and Quad Bayer reconstruction method like DPN {{cite:88fcdc159b9279321b008166ee14dd3115dad028}} for the comparison. For a fair comparison, we trained and tested the reconstruction methods with the same datasets. The performance of the compared methods has cross-validated with three different noise levels, where the standard deviation of noise distribution was set as {{formula:4738a844-0210-41d8-a7ea-5893f3ecd797}} . Later, we summarized the performance of deep models with standard evaluation metrics like PSNR, SSIM, and DeltaE2000.
| m | 9e6b840fff89f28c1e5d1faee95dd4b3 |
In order to further assess the role of BAO + FS data, we also perform an MCMC analysis without considering them, and use the data set P18 + SN + {{formula:c9e3d41c-6292-44f3-95f5-82fcd027892e}}. The results are presented in Fig. REF and Table REF . As can be seen, removing BAO and FS data leads to a somewhat larger value of {{formula:141e2d95-c854-4ee4-b5ae-6bea86e78437}} km s{{formula:56f08ccd-b21e-4900-92c0-db325b278e5f}} Mpc{{formula:029c6d33-aa8f-44df-983f-ba3be5f05c4b}} for the EDE (and a much larger bestfit of {{formula:c8dde17c-13cc-494e-9567-f8825548e064}} km s{{formula:226d80b9-a94f-4e80-82dc-4428e7161a8a}} Mpc{{formula:b97e5590-0ecb-43bd-8215-ccc14242751b}} ), confirming that BAO + FS have the power to constrain these models, as shown inNote that our results slightly differ from the ones in Ref. {{cite:af618c6976556f13835dd427adc058099f4b620f}}. The main reason is that we use a different {{formula:f089b430-27cf-4f6b-8b01-29bd099e6e58}} prior which has a stronger impact on our MCMC analysis, whereas they used a prior obtained from earlier SH0ES results, i.e. {{formula:7ab934d6-738e-4281-8364-8a687ab66606}} km s{{formula:174a3a59-f036-41ea-8eaf-f25cbd770e8d}} Mpc{{formula:7ae39e7e-2a5d-459d-b3a8-c3df179addb6}} {{cite:1f15255bc1011495e09e6f587eb8e56d76575fba}}. Also, we fix {{formula:037c4435-8a19-4220-9f41-92c1a37a6418}} in our analysis which leads results slightly different from the ones obtained with the Planck assumption of one massive neutrino with {{formula:06523d8e-4bb5-4179-8bd4-8cb8eb78b9df}} eV (see main text). Finally, we do not use high redshift Lyman-{{formula:f5bbe4eb-a52e-40c4-80c3-d06e5b08c780}} forest data from eBOSS DR14 measurements {{cite:699962745626cfb665fed5299f3bc73ba0803210}}, {{cite:8bcf081b33dc365b6b3663c0fbecb2d8006b7f27}} . We have checked that we recover the results that are consistent with Ref. {{cite:af618c6976556f13835dd427adc058099f4b620f}} when using the EDE model with {{formula:e37e186b-e739-4eae-aa53-fb8ac1ef9007}} and their dataset and conventions. Refs. {{cite:0d6d657eb9141c7d15364264e8a3490769e2bdff}}, {{cite:26b838286b19543b78f7a4d7d72b972a78819f9c}}, {{cite:af618c6976556f13835dd427adc058099f4b620f}}. On the other hand, {{formula:b53049c3-463a-46cc-aa07-feac4a6046ed}} for the EMG model increases only a bit to {{formula:3fe932dc-e813-45c0-b4bd-c97ce2c432ad}} km s{{formula:d684fac4-0ab6-4704-aa88-711579aaa46c}} Mpc{{formula:258ce76d-e531-46d1-8cd3-27c0ce825ff2}} , since BAO + FS data constrain it less than they constrain EDE models. It is very interesting to note that the best-fit value for the coupling {{formula:350b4f13-78f9-42d1-9514-5bc92e180af1}} is very close to the one found including BAO + FS data. The EMG model fits all the data (except for Planck lensing) better than both the EDE and the {{formula:f8f93d8f-2c89-4f0e-ab0e-75f02d7c95cf}} CDM model, leading to a {{formula:115787ff-8ab7-4daf-b86a-a4b0cb31ed30}} . This time, however, the improvement in the fit does not warrant the increase in the model complexity compared to {{formula:e868bab5-bc98-4df6-9a7d-9930a5845e50}} CDM and we obtain a Bayes factor of {{formula:0836ac47-879a-48e0-8620-7e0c3c82dfde}} .
| r | 2ff2e5926ce9b30f226965d5207b045a |
where {{formula:54374ba4-d678-4042-8add-ee395da3ea9c}} and the sum is restricted over nearest neighboring sites of a two-dimensional square lattice. The critical temperature is given by {{formula:df9b16f5-0376-496d-b26c-5c61bb2c4c66}} {{cite:e4f5ced9b78a6dad72d74f6d1bd55826158e32ae}}. The correlation lengthLengths are measured in units of lattice spacing. in the low-temperature phase reads
{{formula:4f8c5bdc-8ee4-413a-8b46-07340c4eb66c}}
| r | 299464ca28025d40eea9faedfcef3a47 |
Theorem 7 (Maximum over a finite set, {{cite:365b7a9e1c224b2448fab3798e54a79d09affdf3}})
Let {{formula:2862923e-f2ea-4c77-874d-be093df4a26a}} be centered {{formula:f0234974-1a26-4100-963c-c63ba1fe2795}} -sub-Gaussian random variables. (i.e. {{formula:df9b60b1-f4f0-4133-803f-c2542140f796}} ). Then,
{{formula:48935bad-45e7-46c2-abc3-8e60102cbf4e}}
| r | eba15cb61c65ffc66f39a8838ab6a5a7 |
Mechanical systems composed of coupled bistable units have been explored in recent years for applications in soft robotics, shape memory, and information processing {{cite:0d7da45106f968230a4d44c8f35a84a46c91807e}}, {{cite:1f42ba596bfeee13e3d772124f052eaa97d4d157}}, {{cite:caa17323126d2ddaf5ece0d8b4f63f0cb0c3742d}}, {{cite:f43074b8507c2a317ff2f3ebb890c9401cf6494f}}, {{cite:af7689fdf4114acba65977e4f6fa304635040d1b}}, {{cite:41ce4bc3b4f33ab586684f405b64c5b78980dc73}}, {{cite:b0b67e7ab8e57218c23682dc76247cb093976c7e}}, {{cite:4dcebf59cfaefa2007765bc49819075e557316e5}}, {{cite:c109dcc8744c6c502795efbca4fe7e6bb9a5f37b}}, {{cite:2e0c1b6e4af049cfd5e73d30566c6279ebe55062}}, {{cite:a3e0fe47c9e4c76b538d94109217a90390f81bfa}}, {{cite:75e94b73427fc59b1306a9d52312e3f45c0f9f80}}, {{cite:bb22c5217afabe0f62ffac7b20bde890e16023c7}}, {{cite:a1c39e903f1796d311e5369bf61c2bc62e3b6d6b}}, {{cite:f7f05e93eea3d480146606f5b3ac84babb8f7c21}}, {{cite:4a1211c87ecc252f48d9392047f5489ad68ddbf6}}, {{cite:65608f5c514fe86e5a77715828fe61d2fcc2d9d0}}. An appealing feature of these metamaterials is their tunability—each of {{formula:af31469d-d7b4-4e62-a6ad-d1d7b4fdde55}} bistable units can be individually inverted, possibly leading to {{formula:9404c1e5-5803-4828-a762-11aa5024ba62}} metastable states and diverse macroscopic behaviors {{cite:e5a9d065a4bfc3329097a02a3207039c45ec81a5}}. Tunable materials are of interest for many technological applications, from optical filtering {{cite:1673a93ee1abed03ec293d21257980defcd5ced5}}, {{cite:831226a24807fd1d89f51830ad2924eb56942fb4}} to reconfigurable structures {{cite:f2425837e9f0738c2ad029ad1bcccd7e0d08e7af}}, {{cite:f31e0cac178148d063f71039bda7170c35798a4a}} in which it is desirable to have a single material serve multiple functions. A shared challenge of many tunable materials is determining how to easily and reversibly control microscopic configurations, enabling the desired macroscopic transformations.
| i | fea88f46698edb2b28098106fe018159 |
SW networks are generated using Newman-Watts-Strogatz small-world model (NWS) {{cite:23105e8534d08b4580066d37841c05b3030549bb}}. In each experiment, we chose the number of nodes {{formula:0f5ee0ef-8a97-46be-98f4-adf91f1f8ab7}} , {{formula:20a8a149-0ab0-45dd-92a6-c61976dd5345}} neighbors with which connect each node {{formula:fc891e0d-17b2-4858-9df3-cbb41569861a}} in the ring topology, and a probability {{formula:bae8a2b0-5ffd-4a87-ad44-f7feb3a6f96f}} of rewiring each edge.
| m | 7851f00de43150d8a50bf4fd25ef3a06 |
Bias Identification
Three SSD Mobilenet V1 models are iteratively implemented and trained on increasingly corrupted training data to study the effect of training image quality on precision of ODs. For all models, an SGD optimizer with a learning rate of 1 x 10−3 and a batch size 16 is used. Various experiments are carried out by tuning the hyperparameters namely, learning rate and the batch size. For identifying potential bias an investigative procedure is followed and results are reported after the carrying out the following steps:
{{table:8477b9da-1b2e-4e29-ace5-fe5aa37cc673}}
Implementation of Stage 1:
For the baseline determination, subsets of mixed Pascal 2007 {{cite:df0f83c67a0f74f3f866b0b2156451fed2cabcc2}} and Pascal VOC 2012 dataset {{cite:c912f437d6aae9e734777fc879dbc0ffe60f0df8}} are used for training and testing the SSD model. The training image set is the union of 2007trainval and 2012trainval and the validation image set is 2007test. This subset formed a collection of 16551 images for training and 4952 images for testing. As Pascal VOC supports 20 classes, however only 4 relevant classes to the DAWN dataset were selected for the analysis namely, car, bus, person and bicycle. The SSD model trained on uncorrupted data performs very well on uncorrupted data, as expected and reaches over 70 AP on all classes with 74.95 mAP as shown in Table REF . Therefore, the baseline was established and satisfactory results were obtained for OD with the MobileNet SSD model.
Implementation of Stage 2:
For measuring the potential bias, the model was trained on the same training data as Step 1 and tested the performance on the 1000 adverse weather condition images of the DAWN dataset {{cite:45f701ad1d9533d9ca37eb91aeb6d65bd4c37074}}. The model fails to perform satisfactorily on the corrupted images and reaches 3.75 mAP with nearly 7 AP on most classes as shown in Table REF . It fails to detect buses and bicycles completely and shows low performance across all classes. In spite of the similarity in object classes, the inability of the model to detect objects in the DAWN dataset, leads to potential bias hindering the OD.
Implementation of Stage 3:
Finally the model was trained on a mixture of bad and good weather images from PASCAL VOC 2012, 2007 and DAWN data and tested the performance on the 1000 adverse weather condition images of the DAWN dataset {{cite:45f701ad1d9533d9ca37eb91aeb6d65bd4c37074}}. The model fails to perform satisfactorily on the corrupted images and reaches 5.5 mAP with nearly 7 AP on most classes as seen in Table REF . It fails to detect bicycles completely, but shows improved performance (7 AP) on the bus class. Mixing a small percentage of images is able to counter the bias minutely, however this method is insufficient for boosting overall performance.
In Step 1-3, we established that there was a measurable bias in OD while testing models on DAWN dataset. However, to prevent another form of bias from creeping into this experiment due to the use of only one perfect weather dataset (Pascal VOC), the same experimentation was carried out using the Microsoft COCO dataset {{cite:e7f914340606fe92b6c0929fad6fd5facc02192e}}.
For the baseline determination, subsets of Microsoft COCO 2017 dataset {{cite:e7f914340606fe92b6c0929fad6fd5facc02192e}} are used. The training imageset is train2017 and validation imageset is val2017. This subset formed a collection of 118000 images for training and 5000 images for testing. As COCO supports 80 classes, however only 4 relevant classes to the DAWN dataset were selected for the analysis namely, car, bus, person and bicycle. The SSD model trained on uncorrupted data performs very well on uncorrupted data, as expected and reaches over 47 AP on bus class with 28.37 mAP as shown in Table REF . Therefore, the baseline was established and the satisfactory OD results were obtained with the MobileNet SSD model.
In measuring the potential bias, the model was trained on the same training data as Step 4 part a and tested the performance on the 1000 adverse weather condition images of the DAWN dataset {{cite:45f701ad1d9533d9ca37eb91aeb6d65bd4c37074}}. The model fails to perform satisfactorily on the corrupted images and reaches 9.5 mAP with 29 AP on person classes as shown in Table REF . It fails to detect bicycles completely and shows low performance across all classes. In Spite of the similarity in object classes, the inability of the model to detect objects in the DAWN dataset, leads to potential bias hindering the OD.
Finally the model was trained on a mixture of bad and good weather images from Microsoft COCO 2017 and DAWN data and tested the performance on the 1000 adverse weather condition images of the DAWN dataset {{cite:45f701ad1d9533d9ca37eb91aeb6d65bd4c37074}}. The model fails to perform satisfactorily on the corrupted images and reaches 15.25 mAP with nearly 29 AP on person classes as shown in Table REF . It fails to detect bicycles completely, but shows improved performance (26 AP) on the bus class. Mixing a small percentage of images is able to counter the bias minutely, however this method is insufficient for boosting overall performance.
Bias Mitigation
As observed in the above procedure, the models are clearly biased against adverse weather and are unable to perform satisfactorily. In this section, two methods of bias mitigation are presented. First, using images of the target dataset and the second in the absence of target images.
Knowledge Transfer:
The model is first trained on the Pascal VOC 2017 dataset and then learned by fine tuning on the DAWN dataset. As displayed in Table REF , there is a rise by 1.75 mAP using this method as the mAP reaches 5.5 and the model is able to detect buses with increased 7 AP.
Synthetic image corruption:
Here a Double Gaussian Blurring technique {{cite:b6a9e8333c3afcb00e5b70cdc858ae06f96bf3c6}} is employed to corrupt the images of the Pascal VOC 2017 dataset. The model is first trained on the corrupted Pascal VOC 2017 dataset and then tested its performance on the DAWN dataset. As seen in Table REF , there is a sharp rise by 22.5 mAP using this method as the mAP reaches 26.25 mAP and the model is able to detect bicycles with 92 AP.
| r | 7b02f82233e858caa972404fb6cf4329 |
Sometimes, when compared to more traditional SARL, cooperative MARL can be characterised by the additional challenge of the environment being partially observable. Formally, this corresponds to agents interacting in a decentralised partially observable Markov decision process (Dec-POMDP) {{cite:a9727b8a671d9bf54d35249678d6dc0d6fce0499}}, which results in agents that only perceive a subset of the state space and that select their actions independently (more details are given in the background section).
It is well known that partial observability can be particularly challenging and prevents a centralised controller to be used to solve these kinds of MARL problems.
As a result, the MARL community first considered SARL approaches to independently control each agent with a fully decentralised (FD) controller. Later it exploited a method called centralised training with decentralised execution (CTDE), which allows exploiting additional information of the environment.
| i | 0420788145b5f18f3bcc318cf26a147d |
To this end, we introduce a novel technique called uncertainty-aware mixup (UMix), by reweighting the mixed samples according to uncertainty within the mini-batch while mitigating overfitting.
Specifically, we employ the well-known mixup technique to produce “mixed” augmented samples. Then we train the model on these mixed samples to make sure it can always see “novel” samples thus the effects of IW will not dissipate even at the end of the training epoch.
To enforce the model to perform fairly well on all subpopulations, we further efficiently reweight the mixed samples according to uncertainty of the original samples.
The weighted mixup loss function is induced by combining the weighted losses of the corresponding two original samples. At a high level, this approach augments training samples in an uncertainty-aware manner, i.e., putting more focus on samples with higher prediction uncertainties that belong to minority subpopulations with high probabilities. We also show UMix can provide additional theoretical benefit which achieves a tighter generalization bound than weighted ERM {{cite:8154663969ba3daf487ae4d115ee00f2eb2d4d38}}, {{cite:992b580f3eeae63bb02a54f084b1bdc65ee6bc67}}, {{cite:ac3aefe0fbd8e13a836934884a5bfa433248944b}}, {{cite:7c701ddc033dfd20e3560a2263ec163d411ddd92}}.
The contributions of this paper are:
| i | 6e8ca13ef3b57487e753aece6163ec0b |
There has been extensive computational research about controversy in online discussions.
To quantify controversy, {{cite:990b97fef0387a11076fc534f9a99790b53eb22a}} (2010)
used the ratio of positive and negative sentiment words;
{{cite:8891eb699d5d6b82f3a9e227449ba01d375ced4e}} ({{cite:8891eb699d5d6b82f3a9e227449ba01d375ced4e}}) and {{cite:a757d3d5eb5178e3f2ec2c250a2cbf2497a89dd0}} ({{cite:a757d3d5eb5178e3f2ec2c250a2cbf2497a89dd0}}) define controversial content as that which attracts a mix of positive and negative feedback.
Other lines of research used predefined lexicons to collect controversial discussions {{cite:3ad227ad17502a7d79401719086b5344223234be}}.
Prior work on predicting controversy or distinguishing controversial discussions from non-controversial ones rely heavily on dictionary based approaches {{cite:1988fed9985c417cd88da09d593216c253f6e9eb}}, {{cite:0c6ea20a90e1647d456faf5099e9833b6546a711}}. Such methods benefit from being explainable but have limited predictive power {{cite:f870205c887b36b90228e39f179b4418270b8d70}}.
Supervised neural models such as CNNs have also been used recently for predicting controversy {{cite:8891eb699d5d6b82f3a9e227449ba01d375ced4e}}. Since it is difficult to get ground-truth annotations of controversy, these approaches typically use distant supervision, such as upvote-downvote ratio, to mark examples as either controversial or non-controversial.
The growing popularity of Reddit has attracted research attention to predict and model controversy on its platform {{cite:fdd604c1651a0f90806e34c5870d2a268d5a2843}}. {{cite:a757d3d5eb5178e3f2ec2c250a2cbf2497a89dd0}} (2019) uses BERT to predict controversial conversations on Reddit. {{cite:3f972b9578ccd5f2cf9102b4826a764ad5290412}} (2019) refines controversy into disputes, disruptions, and discrepancies via user actions and connects those user actions to sentiment and topic analysis to explain controversy specifically within political discussion.
| d | 6df9d6efa3aee998dd5b3cfaac8557b9 |
Discriminative correlation filter (DCF) is widely used in object tracking due to its competitive performance and computational efficiency enabled by fast Fourier transform (FFT). DCF produces filters by minimizing the output sum of squared error {{cite:bd1ffd05c09f3fcb1e88a0cda9cd08ff4cf533d3}} for all circular shifts of a training sample. The periodic assumption on training samples causes unwanted boundary effects, which can be alleviated by adding spatial regularization {{cite:f0ab8113d5d37c4a14a33e4ab44d4ecb3f0b56cc}}, {{cite:5143c8a7064701c0f2460c2b4e7be244ad2f96eb}}, {{cite:d0fbd35f2155e5c41015b06a57d1eae8c50e762f}}, {{cite:580a7fc86855834f13acfe110f53716186e6417b}}. Concretely, background-aware correlation filter (BACF) considers all background patches as negative samples by using a rectangular mask covered central part of the circular samples {{cite:5143c8a7064701c0f2460c2b4e7be244ad2f96eb}}. Benefit from alternating direction method of multipliers (ADMM) and FFT, BACF is also computationally efficient.
| m | 6a0d87afd50a35ffd9e064aa49f5608c |
The resulting gamma-ray line fluxes on a timescale of {{formula:11f4eaa0-ed4f-405e-99cd-d03bd2a6f1a7}} days are {{formula:57d35b7a-b811-4812-9a60-86d2963fbd98}} ph cm{{formula:13bdc53a-e884-4e33-aa71-4c0252a12a4e}} s{{formula:249dd189-5124-4c82-8c7e-bfccc4a92961}} in the photon energy range of {{formula:db593388-37c1-4304-afa0-46001ef0d189}} MeV at a distance of 40 Mpc. The sensitivity of the current MeV gamma-ray missions, such as INTEGRAL {{cite:38a906aa43123e76dd0736c89ca3f2d3aecee791}},
is {{formula:53294b7f-3af4-48be-8f1e-fa5d2dea39b6}} ph cm{{formula:e99b2562-7899-46e4-a7cf-41e8397ef6e1}} s{{formula:63d2b6fb-8916-484d-ae56-40434a20a54f}} in the MeV band, being much lower than the line fluxes derived from our analysis. The sensitivities of the proposed next-generation missions, such as AMEGO (All-sky Medium Energy Gamma-ray Observatory, {{cite:eb97b4e712e698531eb710077619e58b601337dc}}, {{cite:f6dd554e57ccb42ec762bb561f1a454a423d4c0f}}), the e-ASTROGAM space mission {{cite:9985ea2d7383af92c39ffe55e528ac8bdcb2a329}}, ETCC (Eletron Tracking Compton Camera, {{cite:06c5e4010fdd0b63bb28e149890504e5c631a114}}, {{cite:93857ecaf347560281dd1bdc628ffc69131808ef}}) and LOX (Lunar Occultation Explorer, {{cite:d59b609f7ba6411dee9c2183b540e9d3772cf146}}), are {{formula:e5f5cfae-ede9-418e-bc30-d378fafbc4ae}} ph cm{{formula:19eaaf44-afbc-4c87-b70d-ddd82b7853b7}} s{{formula:14baf189-5fbf-4e94-85c1-9163d6f9d432}} . The gamma-ray lines would be marginally detectable with these missions.
| d | 09b3b1a12723697b28a74a57def62b76 |
Studies of the temporal variations in solar differential rotation (DR) and solar activity are important for understanding how the solar magnetic cycle is generated. Sunspots and sunspot groups have very often been used as tracers to investigate temporal variations in solar rotation using the Greenwich Photoheliographic Results (GPR) {{cite:9db170951085ad402724dab3e747cf5f55a0f550}}, {{cite:d152646e0c823ca15d57f133df04a852c10fadce}}, {{cite:f3d618ca306a570707c13fb745581ccdc299abf6}}, {{cite:3f93370cde3c1b90a56e3ef5266b5ec8d81f1b8c}}, the data set from the Solar Observing Optical Network/United States Air Force/National
Oceanic and Atmospheric Administration (SOON/USAF/NOAA, abbreviated SOON/NOAA) {{cite:a1f6c8fa56a4df131987abd44238fc76a8cde8e0}}, {{cite:e7b0cfc49247636215f8ebf2525b255d8cb4de43}}, {{cite:7e520455ac877d5549fb69a06abb01d40aa685e8}}, {{cite:6e6856360967bbc39fcdd3ed47d6991dd24d96b8}}, {{cite:ace1f51c70ff55d4d6a1f0b97f09b64ceb8b0de9}}, the Mount Wilson data set {{cite:5134bf822148ae55997bbec37449c71afbb0c432}}, {{cite:28a03127788729d2d547c6d11e316131f06a065e}}, {{cite:a4cf986c681cfd71494734eec0022227d75e57b6}}, and data set from the Kanzelhöhe Observatory for Solar and Environmental Research (KSO) {{cite:16a65eebb9f7d0c19bc22859322de9de84b1679d}}. Long-term studies are necessary to verify the results of temporal changes in DR, as well as to clear ambiguities related to the relation between the rotation and activity.
| i | 230a3846aaf18a5898aafd6098e30616 |
Decades of research has explored geometric models of image structure that can be used to regularize solutions to this inverse problem, including {{cite:bb785596a1eebc14b6a890b7aab8b14458ec2392}}, {{cite:198a3c71077f2b2da71996be4082d764f9096c7e}}, {{cite:aaa4ea3cb52056f1c1ee5233b004a253755c5f68}} and many others. More recent efforts have focused instead on using large collections of training images, {{formula:f23c7c5a-c422-4a2b-baf1-c9edaa090b6c}} , to learn effective regularizers.
| i | 78f5c6803ea091e4eac0ad3dd6888836 |
For continuous waves, thermalization is suppressed in a many-body-localized system, which is necessary to realize a discrete time crystal.{{cite:aa782aa8a751eda847cf4fd02c727516879c618c}}, {{cite:444d750fb26a7169c4a305a21cd1a97497252631}} The time evolution operators for the two-site one-fermion model that is referred to for the discussion of the frequency of the strong-field-induced charge oscillation in Sect. are similar to local unitary operators discussed for a discrete time crystal. This similarity may be helpful when considering the possibility of emergent charge oscillations in different situations.
| d | d8d758a325e62ca18af5196a742f5cdf |
We addressed the task of goal-oriented navigation in a two-fold way for each scenario and each approach. In the first one, the vehicle was set to start in the air and diving to an underwater target, while in the second one it was placed underwater and should navigate upwards to an aerial target. We set a fixed starting position and a target location to test our agents and compare their performance with Lee et al. {{cite:fcc746ae24f29c1b592a25dca5a45f65daac296a}}. In the first scenario, the initial position was (0.0, 0.0, 2.5) in the Gazebo cartesian coordinates, while the target position was (2.0, 3.0, -1.0) at the bottom of the water tank, representing a transition from air to water.
| r | fb28043cb5beaa219aaed30d4ef418f7 |
To address the safety issue in real-world RL training, we present the Intervention Aided Reinforcement Learning (IARL) framework. Intervention is commonly used in many automatic control systems in real world for safety insurance. It is also regarded as an important evaluation criteria for autonomous navigation systems, e.g. the disengagement ratio in autonomous drivingFor details, see: Autonomous Vehicle Disengagement Reports. In this work, instead of using disruptive events such as collisions as feedbacks, we try to optimize the policy by avoiding human intervention. Though utilizing human intervention to avoid fatal mistakes in RL training has been applied to Atari games recently {{cite:7bbda45fca954c1b7bed3f7c030fdb4b99b589b9}}, our work is featured in real world applications. Besides, we propose a more general framework to take human intervention into account: We define the expression of the intervention as a combination of a unknown classifier deciding when to seize control and a reference control policy; We redefine the behavior policy as blending of the policy to be optimized and the intervention; Then we try to reduce the probability of intervention and learn from the reference policy at the same time.
| i | 246cde70a4bc783d2541ebc2dc51ed2a |
Our ASWL training pipeline was implemented in TensorFlow {{cite:0acf2ca181b58ff93ee69f8b6f864eb30b311a01}}. All models are trained on a computer with Intel i7 8700K CPU, 16GB RAM, and two NVIDIA RTX 2080 Ti graphic cards, each of which has 11GB of GDDR SDRAM. The source code of this work is available in the supplemental materials with detailed comments and will be made publicly available after the review period of AAAI 2022.
{{figure:28e3609f-0ff4-4a5d-a902-993659ffe5e7}}{{figure:f3297920-43f6-42bd-a140-b436e7c0b1c3}} | r | dc04aa01ad09396afe2e5d8ee766cdf2 |
Our analysis stops at a normalization batch size of 2.
The relation to InstanceNorm {{cite:f7e2b923490a2a6037da69ecb938c783326669f0}} will be discussed separately in Appendix REF .
| d | 15f50153844bee0eb25790a2377e4329 |
Asymptotically AdS black holes have been extensively studied during the past 25 years
or so, mostly in connection with AdS/CFT correspondence and its various extensions
such as AdS/CMT (condensed matter theory)
{{cite:abeb6ad715f07d70b5be32c0a2a2c17a931cdcd1}}, {{cite:a56021623a1c4e9810c3b65acf13c38f0fbfc80b}}, {{cite:71da58ad34282a25a0c4c9c66f809834e003c236}},
AdS/QCD {{cite:87f58965d7710405ad709ea7a2458b943a9de6be}}, {{cite:e1fb192222072e6e5575acde988e03ccf0f94e36}}, {{cite:cbb330c7ce7d1be5f3d736830a98b678d4bf9080}},
etc. Although a countless number of theoretical results have been obtained
in these fields, most of the correspondences remain
qualitative. The present work adds some extra contribution to the field of AdS/CMT
correspondence with a precise quantitative match between the two sides.
Let us remark that the AdS/phonon gas correspondence described in this work
holds in generic spacetime dimension {{formula:cb166e28-fba7-44c4-91d8-db376c3ca86c}} , where {{formula:95a4792c-228b-4f62-9fb2-015bb85925b9}} is the dimension
of the bifurcation horizon of the black hole and {{formula:43bc5bf9-2412-4549-aa24-31db210413b1}} is the dimension of the
boundary of the black hole spacetime in which the {{formula:f18d6dfd-f617-461e-b4e0-b0fb2611636b}} -dimensional
non-metallic crystal resides. Let us stress that the AdS/phonon gas correspondence
that we report here is established purely on the level of thermodynamics. No attempts
on the microscopic description were made here, and there is no need to assume
broken translational symmetries from the beginning, as did in previous works
on similar topics in the context of AdS/CFT correspondence {{cite:85db5ab96a10a8f89ecd0543c2ebd7e76f8bc849}}, {{cite:4fc8c122be4f577b4c41f66def96aebc7d70d757}}, {{cite:e590e37afbd6a052a8de82c89e1677c940b942ff}}, {{cite:803890f59544be1f76c5592304356f495ec38806}}.
On the level of thermodynamics, the correspondence is complete, because, according to the
famous Mathieu's theorem, the thermodynamic properties of
any macroscopic system is completely determined by a single thermodynamic potential
with adapted independent variables. In the present context, the Helmholtz free
energy {{formula:1e793f09-2355-478d-9c60-25fbc609c78c}} plays the role of the thermodynamic potential on both sides.
| d | 1ba4c542fa4a655f0ce7e7df7d022170 |
CRF First proposed by {{cite:3eb3b716a91f8e75dfab0578508bfe10faf52ddd}}, conditional random field (CRF) is a type of probabilistic graphical model to model sequential data such as labels and words in sentences. During training, CRF will determine the weights of hand-crafted feature functions to predict the labels. We adopt CRF as a type of non-neural network method.
LSTM Long Short-Term Memory (LSTM) is a redesign of traditional recurrent neural network architecture around its memory cell {{cite:5c236b04e5b41c53a6431536b3065a9e3df45c73}}. LSTM has been shown capable of storing and accessing information over very long timespans in varied sequence labeling tasks such as POS tagging. We train a LSTM model as a non-pretrained neural network model.
CNN Convolutional neural network is successful in some NLP tasks using convolution and max-pooling layer to summarize inputs {{cite:900670df1eaab47e6e3b8ce323de87154af318e6}}. CNN runs in parallel to extract extracting local spatial and temporal dependence features. Therefore, it is faster than LSTM. We add CNN layer to our model to further enhance ability in feature extraction.
BERT {{cite:d42b5c254006ea54102a4aceac1497c5aba80ebd}} develop BERT, a pre-trained transformer-based language model that achieves excellent performance through simple fine-tuning on various natural language understanding tasks.
Compared with previous methods, BERT has stronger generalization ability, which can effectively extract semantic and grammatical features considering long-distance dependencies, but also involves massive parameters that occupy a lot of storage space and take a long time to run once.
| m | a3a5ed5b923e177ca22aa013e79ee9f0 |
Fig. REF displays the non-collinear ground-state magnetic spin configurations calculated self-consistently for the nanostructures described above.
The ground state spin moments of the NW
exhibits an (ortho) helical-spin-spiral along the [{{formula:8647efa7-e367-4293-97dd-a6ea1ce175aa}} ] direction, with a periodicity of {{formula:0e78a5c9-5e58-481a-8dcb-af5e975f8073}} 1.1 nm. The angles between spin magnetic moments of nearest neighbors (NN) Mn atoms is {{formula:f96f5d01-e1f1-4063-8ad2-a0227836dbf8}} 90{{formula:773ddbab-8a0e-4798-b252-3df2907debd0}} and the spin chirality is clock-wise.
The DNM system presents a complex non-collinear configuration in which the spin magnetic moments of NN Mn atoms in different rows are almost parallel, whereas among those located on the same row it is {{formula:92e20b36-1666-4512-b688-6e51442fc408}} 135{{formula:08f0b9dd-ed93-420f-96d0-51d9508de69b}} .
The first-principles calculations for the nanostripes reveal cycloidal spin spiral ground states, with long-range modulations along the [{{formula:0f58e180-db98-44a5-ac3b-c3d69908985d}} ] direction. Stripes 1 and 2 display half-periods of {{formula:9dd8f52d-58f2-418c-b7ab-bb2ba6a3f599}} 3.6 nm and {{formula:340812c4-9f47-40dd-8d07-62d161ce5a65}} 7.24 nm, respectively, and inter-row antiferromagnetic ordering. These spin configurations resemble the ones observed by spin-polarized scanning tunneling microscopy measurements in a monolayer of Mn/W(110), which display a spin-spiral period of {{formula:e887f20d-afc0-4d95-aacd-0d9ee23e8a4c}} 6 nm propagating along the [{{formula:3499b096-ede8-4dc0-af62-afdee5dd5a0c}} ] direction {{cite:689a696fc7d7d3ef43b32f21d6548e5b3eab0dd3}}.
{{figure:3db58e85-51ba-4a40-b406-4e291bd03c82}} | r | 15e5ddeaa958de88fe8500f0c1cd4aaf |
Third, the planner cannot scale up to a large number of states. A planner cannot work well on a large-scale scenario such as the game GO {{cite:69ae2ed3b556a936431290473ec76f5263faf252}}.
| d | dec32aebf377766b289d00345e0b80a7 |
While DeepSTI shares similarities to the ideas proposed in {{cite:f2c9ec05fe243eeae6272eafd1a6a848f1e9fcd5}}, there are several differences. First, instead of training on estimated ground-truth images from COSMOS, we developed a phantom-based training scheme that addresses the issue of lacking ground-truth samples from real measurements in STI. Moreover, a new network architecture based on a Residual Symmetric U-Net {{cite:8127b5743b88f2b8826ac02a7d2d5f5a065ad7b9}}, {{cite:4db3b41bb3e9c66e4c55205cc7c98f9a82b5eab5}} was used to learn the proximal operator instead of the Wide ResNet {{cite:ece9ba5d0b91789374b0c08651c3c578676d633d}}. This led to faster and more stable training as well as superior results in the reconstruction of the tensor image and estimation of white matter fiber directions, especially in the case of very few (e.g. single) head orientations. Finally, DeepSTI leverages a new multi-resolution training strategy that allows the model to handle measurements at different resolutions without the need for any ad-hoc re-sampling or re-training. Being able to handle data at different resolutions and different numbers of orientation sampling is central to STI reconstructions, where acquisition protocols between subjects and studies often differ.
| d | 4be3726df5b581590dcab6190d9b6243 |
We start by illustrating the proposed general framework for text-to-image generation. Our framework is shown in Figure REF , which consists of three key components: (1) a pre-trained image encoder that maps images to their embeddings; (2) a decoder that generates images from the corresponding embeddings; and (3) a prior model that generates image embeddings from the corresponding text captions. In our implementation, we use the pre-trained CLIP image encoder because its output space is a multi-modal embedding space that has been demonstrated to benefit the text-to-image generation task {{cite:d02eb0a402965b2dd520d0692adc1939c46a6abf}}. The decoder can be either a diffusion model or a generative adversarial network (GAN). Note that if one chooses the decoder as a hierarchical diffusion model and makes it conditioned on both image embedding and text, our final
structure will be similar to DALL-E 2 {{cite:443438f55327abe4a4dbad64940d15930bf147de}}.
| m | 3a24c6f97fbac65682ccd2adc94c93fb |
In particle physics, the fundamental interactions of the standard model, the electroweak and the QCD interactions, are described by non-abelian gauge theories. The presence of a gauge symmetry implies the appearance of unphysical degrees of freedom in the Lagrangian, which stand in the way of the usual quantization methods. Typically, the redundant degrees of freedom are removed through a gauge-fixing procedure. Ghosts, i.e., fields with unphysical statistics, are introduced to compensate for effects of the gauge degrees of freedom and preserve unitarity. The gauge-fixed action retains a nilpotent, odd, global symmetry involving transformations of both fields and ghosts, the Becchi–Rouet–Stora–Tyutin (BRST) symmetry {{cite:b2eeb4732ef1d8f6569882bbf45adb91fdfe4e57}}, {{cite:2e7b3f6d07c96947f71f62134e9b25633b3809e1}}, {{cite:effcf44a1f669c751fcdc7089d5d5be63e1822ec}}, {{cite:537dd801c43bd5b885eac5c387f3b195f99b8922}}. The BRST symmetry has played an important role in quantization, renormalization, unitarity, and other aspects of gauge theories. The Batalin–Vilkovisky formalism of antibrackets and antifields {{cite:fbf4a183db92809356d7f653a9830a8cb4ae8bfb}}, {{cite:46d763e5a621e9cdfc1b5294de54f1872a80d3a6}}, {{cite:68809526f8a5ab928517aefc6addf4005a86cda3}}, {{cite:09e378a9227aca85ae403ca5f3f1b8b24e34a851}} retains BRST symmetry as fundamental principle while dealing with very general gauge theories, included those with open or reducible gauge symmetry algebras. The antibracket formalism covers a broad spectrum of applications, ranging from supergravity to string and topological field theories. From the mathematical point of view, probably the most convenient approach to Batalin–Vilkovisky formalism is based upon a certain hierarchy of (super)symmetric multilinear maps introduced by Koszul {{cite:5a2610cd67df735186ad642590786b4989b1bfb9}} in the framework of differential operators, Calabi–Yau manifolds and symplectic geometry.
| i | 8e77a34e9cf0b2538701f7d343d6ba6f |
In our proof, at no point do we need to approximate a continuous map by smooth maps.
Instead, the proof relies on a result about Čech nerves of open coverings from {{cite:733f254c60aeceb665b515fdfe5d07a578c3a9fd}}, a version of the Nerve Theorem, see e.g. {{cite:562985e04db02fbe1452fc742ba954cf719c8bb9}}, {{cite:f932b3dc2bb8dfcf7fb52fcbb249779391cfc5d2}}, {{cite:019dcb2de17cc38ac540dcc9dccc7e5e4a8d55e6}}, {{cite:9e2ee51be4abaf00f2d0699e6eb009d71c648e58}} (concretely, we use {{cite:798bb10f01758714a338aaa2fd311824efeda994}}, which is sometimes also called Lurie's Seifert van Kampen Theorem), and a modified two-sided simplicial bar construction for simplicial presheaves, which we introduce in Appendix .
This illustrates that the model categories {{formula:ee0201b1-7eef-4163-a3f6-e03dbcdc8c4e}} are of interest beyond their abstract properties: they provide useful tools for doing smooth homotopy theory.
| i | 331fcd0b53dc5746fbdaff26e0fc6252 |
LIME. {{cite:176cafc1f4571e04e162b8b830dd0c4b99b40c9f}} propose a system to explain why a classifier makes a prediction by identifying useful tokens of the input. They use a linear model {{formula:d9228e1c-7597-463c-a723-97ffe888ab33}} as the interpretation model to approximate the evaluated model locally. And they use the weighted square loss and a set of perturbed samples which contains {{formula:e55f849b-9963-4655-9d23-e57dae9e1f3c}} tokens of the original input to optimize the selection of useful tokens. We set sample size to 5000 and set {{formula:b69c9402-5b9c-4ee6-9ca9-b4f7b3c699a8}} to 10 in our experiments.
{{table:3472aa51-6dd3-4c98-b0fb-71e85e289b40}} | m | 25df702cac2ac7067f227cfcaaf0a00e |
To fit the parameters for quadratic potential, we first took the
spin averaging over {{formula:f465e3af-adff-4e9e-9022-43e00b6a7823}} to obtain the value of {{formula:cb5f8389-6c14-4079-8462-f8441885775a}} and {{formula:b680ba40-20f1-456f-9560-c0940b615c73}}
for the set of mesons {{formula:53fc5be0-d5fc-4772-8dfd-40d49d555c74}} , {{formula:ae90558e-06c9-4cd9-97d6-b5e31882d48e}} , {{formula:1075f500-6e7a-4052-9c85-a0f3f95f9be8}} and
{{formula:3aead395-dbf4-47af-a582-bd73ba1e436f}} . Here {{formula:afbea0c9-8bf9-465b-bc1c-46164959c754}} {{cite:c4ac14082dfa95cd8a96d4ece967ad29506b0a09}}. By
using the fitted value of {{formula:48898a4c-54c0-452f-8093-1ff956e43440}} GeV and {{formula:86ffd66e-5b0a-4290-b7f4-3fd6b0a9a3e6}} GeV the constant
{{formula:528b12c8-2ced-4a66-bb39-75a5440cd272}} , mesons sizes {{formula:5d5d502f-1f7e-4438-908b-a8215c485a04}} and
{{formula:4fc71ad5-196f-4a47-a1b5-d494f8ec06bc}} are obtained. The constituent quark mass is taken from
ref.{{cite:c4c2cf60c5fd323b4f1b47e16c9241d60a13f305}} and the meson mass is obtained from
{{cite:26018f0c1d9661ae9790af290a32fbc474bbcf66}}. Hence the required parameters are {{formula:60d8951a-3442-40de-8254-b9a821c74bdb}} = 1.49 GeV
{{formula:bc6664d3-3c22-4b66-9a55-5026f4fb3388}} , quark mass m = 1.48 GeV, {{formula:da45d595-5420-4fea-af41-9f3dd4dcc0fb}} =
3.10 GeV, {{formula:20284031-b88e-443c-9ccf-9ffa84919479}} = -0.0255 GeV{{formula:8c51a5a8-19a6-45d5-978d-3de023a57919}} and {{formula:da840c1c-7df6-439f-9a55-054081854a0c}} = 0.259 GeV.
The obtained results are as under:
{{table:c8cae913-cf8d-4f65-9698-d25aabcdfb51}} | r | a636058e5e1afcc384a74463ee12bb98 |
Through theoretical analysis and experiments, we rigorously investigate why MLE with SRLMC could converge to an EBM with wrong density estimates,
and reveal that it is caused by a combination of two heuristic modifications to LMC introduced by previous works: (a) early termination of LMC in short-run LMC and (b) using incorrect learning rate and noise scale ratio in LMC.
To avoid the pitfalls of MLE with SRLMC, we propose a novel technique, USP, to solve MLE for EBMs. USP solves an optimization problem to find a set of points which uniformly partition the support of EBMs. Then, it uses the points to approximate the MLE objective through numerical integration. We also introduce a practical version of USP for training deep EBMs.
We demonstrate on a toy example that USP is capable of accurately learning a distribution with multiple separated modes. We also show on the Fashion-MNIST dataset {{cite:74dde046f4db56e80b4b3f653c9bcd653a3d6936}} that deep EBMs trained with USP attain significantly better OOD data detection performance than deep EBMs trained with SRLMC.
| i | 31003b875b6aab3d4d16440e6d7054b3 |
1. When evaluated in two (or better three) dimensions in the Bertsch-Pratt
system, a small elongation of the emission region (better region of
homogeneity {{cite:0522d585a2cf494907896af326396fec71d7c109}} is observed along the event axis in all types of
collisions (hadron-hadron {{cite:4fbc3f9e4f7e2f3688191372eedb42b14df7d2f7}}, all four LEP experiments {{cite:8311b4b172730fec68fd2018e96d5d7742606c5a}},
ZEUS {{cite:dfcf7deb1304c33f0b02beb0c79a9e6c12aa6d8c}}, RHIC {{cite:d4beb53c6b9ef3e3cf517d3f58d8045b2fd05e4e}}).
However, it is important to note that the longitudinal radius of
homogeneity is much shorter than the length of the sting (of order 1%).
| r | 5574090cb1294a438ef19bd561c40ef0 |
Deep Neural Networks (DNNs) have shown outstanding results in solving linear inverse imaging problems. On the one hand, end-to-end approaches provide extremely fast reconstruction. They are widespread in other imaging communities {{cite:6ca8680e7581c3f803fde3e9a9554705957052f6}}, {{cite:2fbb9bc10f6a5a6166e06fdd205d88d187568691}}, {{cite:7ab0b21955d52e5e5e161b146553e33f021dd7f8}}, but their use remains limited in RI imaging {{cite:037f23b20b151967b185906925b0b9b02b967f09}}, {{cite:9efb495a868076484571cde14c1b4e90df369b64}}, {{cite:59c39c41b68914c625e8e3bc568e93c960231c2b}}. This is mainly due to the lack of ground-truth datasets, combined with a wide intrinsic variability of the RI observation model, leading to generalisation issues. While unfolded architectures {{cite:726e654cc687c53047299b3aa045df202a460b59}}, {{cite:f22894865a57035943717c760c9427daa7371896}} provide necessary robustness to variations of the measurement setting, embedding large-scale measurement operators in DNN architectures is impractical, both for training and inference. On the other hand, Plug-and-Play (PnP) algorithms {{cite:0cdd5f4372a9a87f364b2051754903a0334e26b6}}, {{cite:effae4f1fd2083de2a19dc335910319e1e902d7c}}, {{cite:4b6765f9643d87dc5ec3863d62876fc7719345c3}}, substituting learned DNN denoisers in lieu of proximal regularisation operators in optimisation algorithms,
have shown outstanding performance and robustness, including for high-dynamic range imaging {{cite:037f23b20b151967b185906925b0b9b02b967f09}}, {{cite:019379bddc222aed72e4c483bc5026f450f8c8a8}}. However, PnP approaches remain highly iterative and will still struggle to scale to the image sizes and data volumes of interest in applications such as RI imaging.
| i | c7ecd45275f34a6372b0957711ced0e2 |
gives Eq. (11) as a hypergeometric type equation {{cite:20d67d6eaed0d36bda7b9bed723ac5de4927e02c}}
{{formula:b1bffff4-0255-4405-9400-b8990742e7b7}}
| m | 17b093a725f1a9744612bed38a012619 |
Learning with noisy labels has imposed additional challenges. Sometimes the data quality is known a priori {{cite:c02f48cd1283e3620d69dfdb2e84b024a33e71d2}}, {{cite:86b61f3581b97722d99600de4abc4bdb9679fb27}}, {{cite:c3689be1ca407659d6c3b40d3968ccc1ba679e2c}}, but a more common scenario is that, the data available is a mixture of samples with both clean and noisy labels and one does not know, or only has partial knowledge of the underlying distribution of the noise {{cite:edbc8943145c2d9c9dc2c22c0c2a2ac61d0b77e3}}, {{cite:d7006a1e8fd1bf962f967e08bc451f7914137b77}}, {{cite:b6e9902f7f3752800d5602b6781ffb01004c5288}}, {{cite:3ea84825960bc49966ef1676d7c46c7fc4726046}}. In this problem setting, a learning process that is aware of noise in the labels and actively mitigates the negative impacts from the noisy labels, is the key to improving the generalization of learned models.
| i | a65d510b902b5bbb6aae0ac8132a82ae |
Regularization.
Overfitting is considered to be one of the major culprits of membership exposure {{cite:da0f5c71c3e275412ff4c83bb56e742bdc659757}}, {{cite:a1e465ae8a5f5434ecc74c92c5456f87ee69aa73}}, {{cite:f03e4f8c8501b21b9775c707a218306faa60f592}}.
Therefore, the regularization technique may be feasible to defend against our attacks.
In our experiment, we introduce an {{formula:df389980-f0ef-4eb7-b277-faaa3d88de29}} -norm regularizer with a penalty of 0.05 during the model training process.
Early Stopping.
Early stopping is another common practice to prevent overfitting {{cite:826f72f66d2e01f31236ff4c877dfad1b12192ae}}.
During the training process, for each epoch, we randomly sample out 10% of the training data as validation data and use 90% of other data as training data.
We monitor the validation loss, and if it does not decrease in three epochs, we stop the training process.
DP-SGD.
Differential privacy (DP) provides a rigorous guarantee to limit privacy leakage {{cite:13e5e53eaa2a3b984589c37650893728f88d2338}}, {{cite:0bb8f8180644ca016a81ce9e8a75f073a27ec590}}.
Recently, privacy-preserving machine learning algorithms under the differential privacy have been proposed {{cite:ee043169271e72f0af283dcec954c1bf80b24a39}}, {{cite:ecf0090317a7779c9df32ba8d6d9212121bc9d14}}, {{cite:cf120797cc801eb2deb8bf902400b7393069b2db}}, among which differentially private stochastic descent (DP-SGD) {{cite:15e0b077b049ac4d473682db599df68f0e18ab18}} receives the most attention {{cite:040e765ca226af6805abc2966a8187e3c58cf481}}, {{cite:a4cd761fd13e9999e7869df680dd26adb49d87fa}}, {{cite:2f6228f940295711ac1352e7650bf9239a696534}}.
In our experiment, we utilize the DP-SGD optimizer provided by the TensorFlow Privacy packagehttps://github.com/tensorflow/privacy to implement differential privacy training.
The hyperparameters used in our implementation are summarized in table:dphyperparameter.
As reported by the analysis tool provided by TensorFlow Privacy, we achieve (3.25, {{formula:95cc8538-e6d8-4576-8263-f1a291e3803a}} )-differential privacy on the clean model ({{formula:8bb47d1e-ebcf-4920-b114-957636aff497}} ={{formula:213cf4d8-2523-4e3b-ba16-5c7fe03f697c}} =10,000),
while we achieve (3.10, {{formula:1edd9425-7338-4b54-9fd7-5433e984265d}} )-differential privacy on poisoned models ({{formula:ae76d1f8-61ad-420b-a02c-75f0e0608e9f}} , {{formula:064eafbb-a085-4d06-b00e-379caa7e686e}} ={{formula:7081fb08-d8cb-4d63-b520-1e8b8765c346}} =11,000).
{{table:25a3bb75-8d8c-42e4-ae4b-f2b88d6a4b19}} | d | 6b0c36d1afcfd4f5fb234342a955d4d3 |
Kinetic transport equations have been proposed to model the run-and-tumble motion of chemotactic bacteria {{cite:7ec4a62d4038fb1112d94dcc3a446f1e3543e422}}, {{cite:474bc7e27549d1e1a0471f3bc80c1eb90b358c00}}, {{cite:aa5c8e361eabe60605a4fc997aa612a144debc14}}, {{cite:d801d4101c019bc1ebb5924cf1bf69ab21a25efe}}, {{cite:ae878f5606a5e206f213ec18912881d53d38b988}} and successfully utilized to elucidate the mathematics and physics behind the complicated collective dynamics {{cite:a839edf59f95c83110fdb7cbd178ebd0d3a7307f}}, {{cite:57b2bef466eeea6dbac8a2af2440b1468211c8ae}}, {{cite:c52171cf39f0f94e4cde92de8e21c21f5daaf369}}, {{cite:0fd84d2c217f7c2275b261c80fd08c61e85fcc2c}}.
Since the duration of tumbling is much shorter than the running duration, the tumbling duration is usually ignored, and instead, the instantaneous velocity jump process is considered in the kinetic transport equations.
However, it is fundamental and important to know how that small but finite duration of tumbling affects the collective dynamics.
| i | e57653712f75a6fe897c7e385e99b251 |
The currently most popular and well-known method used for simulating samples from known distributions is the inverse transformation method (also known as inverse transform sampling) {{cite:5e387f213f796783262f1c6921245788c24e7ef4}}).
| m | e4e56639612666c5139163f96dca6047 |
However, this solution strategy is known to be extremely slow for quantum chemical systems {{cite:b41452d7feca8a9e4a8828c0a49a039f2131c177}}, {{cite:6b98c1cd7d3eb2b411411ef2b514603a87540fd6}}. In order to apply the Hamiltonian, a time evolution operator of the form {{formula:fa85c77a-c54f-49e3-93ae-59dac544b931}} must be applied to the wavefunction. The Trotter-Suzuki decomposition of the time evolution operator must be decomposed into many terms, {{formula:2ee81947-0a9a-4877-a969-4bb11f7410dd}} to capture the full electron-electron interaction term, although this can be reduced as {{formula:cc8eacab-8a84-4b1d-90ca-156da5ea0fd8}} to {{formula:f9597efd-2deb-4384-99e7-a32fd85ba0f3}} for the case of local basis functions {{cite:c9aa72d8f03bdad20d3333fe95c6e048f762d272}}. However, since the time step {{formula:0d268496-9e5f-484c-8f01-c910fbfb91c2}} must be very small depending on the strength of correlation in the system, the resulting number of operations makes the time necessary to solve for even small molecules extremely long. This is true of other classical solution techniques such as Hartree-Fock {{cite:3eac163b7cec5293b5728facafbd8d3fc50ea9a8}}. This is expected based on the complexity of solving quantum chemistry systems {{cite:877da5464c96d23eda4d2254e03d6fee198386f3}}.
| i | d6b9f0f3f227e569d83611409a2d91dd |
The baselines used can be grouped into several distinct categories: human evaluations – traditional machine learning approaches (SVM) – classical deep learning approaches (CNN {{cite:16bf62d9030ae6df487007c8416e3588aee30591}}, BiGRU {{cite:578147b5643fe2cba0fb5e9bd81fab69e0240630}} , and HAN {{cite:2e316d1e7736ee32af5599af70e75347dfa51b8b}}) – and various transformer approaches with/without pruning strategies. These transformer-based models are generally considered to provide the current state-of-the-art in text classification. We reproduce these baselines based on the Transformers.https://github.com/huggingface/pytorch-transformers
| m | d84a53090bd3c700bff434661f2979e0 |
As shown in Tabel REF , we have compared our proposed method with those SOTAs semi-supervised semantic segmentation counterparts, including Re-Seg {{cite:554c7f34bfae788e4d6aa76ba07d82ab947fedeb}}, EC-Seg {{cite:49c22ad1e16345a4c80cc3249ed09e98951d6941}}, High-Low-Cons {{cite:e20f338de264bc0b36918217409ea5314f8bf77f}}, and CC-Seg {{cite:9305c7e19ddbf345594eeb8830000b4d5286ab05}}. We have used the threshold (denoted as Thres) that can offer the best final segmentation map for each method. It can be demonstrated that our proposed approach achieves significant improvement based on the previous SOTAs methods. For example, in the circumstances of merely 3.5% labeled data, our proposed method outperforms previous best method Re-Seg {{cite:554c7f34bfae788e4d6aa76ba07d82ab947fedeb}} by a large margin of 2.0% for the task of crack semantic segmentation, and by a margin of 1.9% for the task of road extraction. It can be also seen that our proposed approach has great performance in the circumstances that the training data is extremely limited (e.g. the 3.5% and the 5% labeled case). And our proposed method has nearly equal performance with those semi-supervised counterparts when the is fully labeled (e.g. the labeling percentage is 100%). It demonstrates that our proposed method has superior performance when the labeled data is extremely limited.
| m | 24d8e6e7d5302d13ba74c8af2ce45c19 |
Although the results are promising, the required processing time used for iterative MAP inference is currently too high for real-time application. Although not the focus of this work, one avenue to explore is the use of deep unfolding {{cite:d4d05c66d41db5214260798b69e3a55a02a85b87}}, in which the iterations of the algorithm are unfolded as a feed-forward deep network {{cite:9d7710cb7b636dd5008919fd02fb403110735ab6}}. In other applications this has yielded major speedups (as high as a factor 100) {{cite:4e13adf3e70524ddc65bb87368b66379c4df26ab}}, {{cite:700290e153ceb5a3f2ea81fc1a5031989058a9ae}}, potentially enabling real-time inference. Another means of acceleration would be to exploit the temporal structure/persistence of ultrasound sequences, initializing the MAP solver for the next frame with the with the solution for the previous frame. This way the algorithm “tracks” the statistics. We leave this for future work.
| d | 97ed9a5feea8e2d4d321e35238b06707 |
We showed that a model of counting based on the ESBN architecture, and trained with reinforcement learning, exhibited a developmental trajectory qualitatively similar to the one observed in humans learning to count, as well as the capacity for systematic extrapolation. A model that implemented only the retrieval operation required for the counting task (the dot-product similarity operation) displayed good performance and extrapolation, but not a clear inflection in its developmental trajectory. Baseline models using either an LSTM or transformer as a controller, but without the external memory component, displayed much slower learning overall, no inflection in the learning trajectory, and no capacity for extrapolation. The transformer model did particularly poorly, possibly because standard transformers are ill-suited to processing adjacent time-steps {{cite:ce20fac63ed52eac873a506cd8ec5e9e1047a604}}.
| d | 11fe85920c0218fd9084f7020a794c2f |
Diffusion models {{cite:490667a20861f5118e8f93fd26a7d30811c926bb}}, {{cite:b49e5283dea37345ef00c9049bcff8c04bd1a981}}, a recent family of generative models, have achieved remarkable image generation performance. Diffusion models have been rapidly studied, as they offer several desirable properties for image synthesis, including stable training, easy model scaling, and good distribution coverage {{cite:20b30fd787c303de9c7b97fd5b8511a09209c8a9}}. Starting from Ho et al. {{cite:b49e5283dea37345ef00c9049bcff8c04bd1a981}}, recent works {{cite:20b30fd787c303de9c7b97fd5b8511a09209c8a9}}, {{cite:72c6cf984fa6e1de0e9bc74e1c505896b9b1c023}}, {{cite:bc203ef1d9d6590717df078ee7252561ff48da53}} have shown that the diffusion models can render high-fidelity images comparable to those generated by generative adversarial networks (GANs) {{cite:649fddd6413c9d0236ba28aa2b6e87b190d0e41d}}, especially in class-conditional settings, by relying on additional efforts such as classifier guidance {{cite:72c6cf984fa6e1de0e9bc74e1c505896b9b1c023}} and cascaded models {{cite:4c94e9765613b9541024adae170cc5e4ec6408c3}}. However, the unconditional generation of single models still has considerable room for improvement, and performance has not been explored for various high-resolution datasets (e.g., FFHQ {{cite:d35fa941b125eb91ee37fac4f72a592114d97842}}, MetFaces {{cite:8d307c7d46e37ee1223be6414d03aba79bdca615}}) where other families of generative models {{cite:d35fa941b125eb91ee37fac4f72a592114d97842}}, {{cite:5c6b7b9a31dd8a0d228b2e170044f51d407bee96}}, {{cite:b976bafca5f1a2a2bfd87cb4912723825075fbdb}}, {{cite:24e958843723f9463c3a5434a8237e32d9db42fb}}, {{cite:9e27992a9148dc28fe3b55b29eaef78bec4b97a0}} mainly compete.
| i | 1d4e3b47d473093c76a15a1bcdf4632d |
According to the model, CAG has several parameters that can be investigated to assess their effect on the agent population, with particular interest for the density of miners.
Starting with the temperature {{formula:d15aa1ea-c9ee-49ce-9673-4313ce559656}} , we consider a population with 2500 agents, to analyse the outcomes at different degrees of rationality.
Such analysis is performed with special attention to two ranges, the first one for {{formula:308f8b8a-bb13-4904-9ad7-04f7c936d6de}} and the second range for {{formula:18614499-2a67-42a5-82db-c90a9cbd0753}} .
Previous investigations on dilemma games, as the Prisoners' dilemma and the Public Goods Game, reported that for the range of low temperatures the density of cooperators increases as the temperature increases from 0 to {{formula:f687ba24-804e-4697-971e-fee14ff83e57}} and then the beneficial effect of the temperature on supporting cooperation reduces {{cite:7530542bfe2dd128f7062d98b0a379d1e5752111}}. Then, as reported in {{cite:aa6cdfdb0551b91ba5bbaaa326e13ba84e267c9b}}, increasing the temperature the dynamics of an evolutionary game resembles that of the voter model {{cite:a3055043b250c46beb83f5c0167c61dfb4704d56}}. Namely, too high temperatures entail agents change strategies by a process equivalent to a coin flip.
In this case, we cannot map cooperation to the use of a token and defection to mining, because as described before, proof-of-work-based blockchains need miners, while populations are perfectly healthy without defectors.
The result of this analysis is shown in figure REF .
{{figure:a92226a9-b064-4b14-b873-f655e28f38a3}} | r | 18d074641b05be64667a8c1edbb7798e |
In this section, an additional discussion regarding the results of the case studies is added. In the case studies, theour proposed method (GP-MBDoE) was compared with two alternative methods, a method that deals with parametric only uncertainty (MC-MBDoE) and a disturbance based estimation method (DE-MBDoE). The GP-MBDoE has the main advantage: it can deal with potential model-mismatch and have as good performance as the standard MBDoE techniques in terms of information metrics.
The non-parametric nature of the GPs offers the opportunity to approximate the physical system with a small number of data-points locally with the help of trust regions. This nice property comes with additional experiments' expense due to the conservative nature of the trust regions. This method's main computational burden is the training of the second Gaussian process for the objective function. An arbitrary large amount of data-points can be used for this surrogate, as only in-silico data points are needed. This can be an issue as the number of the design variables and model parameters increases; however, approximated methods have been proposed, including sparse GPs and generally variational methods for the training {{cite:920bb746129e6101c3b19f7ce12ca6722d3ffc80}}. In this work, the computational times needed are on average 42 and 121 CPU-sec with 10 multi-starts for each GP and 10 multi-starts for the optimization for case study 1 and case study 2, respectively. The computational times for the GP-MBDoE, MC-MBDoE and DE-MBDoE are given in Table REF .
Please note that a preliminary set of experiments is required, particularly we need {{formula:12c789dd-3334-493a-b696-b9a680b3206a}} data points to compute the initial Gaussian process as the covariance of the input data points needs to be invertible.
| d | abc822c59d039e6504f4445a069f4438 |
The appropriate classical 3-point amplitude has been argued to be fixed to all orders in the BH spin at tree level {{cite:09726a228e671094fde9f93a7cd8971bc34f55ca}}, {{cite:99d0987db9a4fe539509b850d8fbbb5300c20f30}}, {{cite:c7405f072552c3d13f7b661f9cf3243c0bc8bdad}}, {{cite:ebe74f0f033ad516dc1c40c1ce5054fcdc431b7b}} by its correspondence with a linearized Kerr solution {{cite:5410d0a7c847f1b8c7da3bf45d03a83861bf3e36}}, {{cite:aa38a3b1fa40f9224eef99cd3ea3b012bbb71eb9}}, {{cite:44f1f6f57e2bc46a6b770adb96a73d84b61f2bbc}}, {{cite:c75692e8fbc1ba91499a760698e4c65dc91f574a}}. Remarkably, the same 3-point amplitude {{formula:85a4aa6e-e6bf-42c0-8cab-515444d3b514}} arose in {{cite:218f5618bfa926acd8e3a90027437f895b37378b}}, independently of any consideration of BHs, (in the large-{{formula:9ad3c71a-b612-44ab-952e-bf1aa390f888}} limit for massive spin-{{formula:2ebd57c8-e723-406d-aea0-6c6157785245}} particles meeting a graviton) as the unique amplitude with a well behaved high-energy limit smoothly connecting to the corresponding massless amplitude, the latter fixed from kinematics considerations alone {{cite:098e1d52406ae64f262d28709b29dc0aa572a48e}}. This link has sparked significant interest in exploring the space of amplitudes, chiefly 4-point Compton amplitudes {{formula:6eb0eba2-72da-46f0-974c-dd3eb6f8dae7}} , which may effectively describe classical gravitational interactions of spinning black holes, in search of certain strongly constraining properties of the amplitudes which may single out the appropriate black-hole solutions.
| i | d9b2e9692d71bc245eb23def5b088ffb |
Based on our previous result, the statement can be traced back to existing results in
the literature. Since {{formula:7564c06b-5a53-4c48-9468-20e1c2bf15af}} is a KKT point of REF , we know
that the bi-active set {{formula:a3608ee5-d390-41dd-8356-6424b5078209}} is empty. Therefore, it
follows from assumption {{formula:4a308662-5da5-406a-ae33-957c3ef94b7f}} and Theorem REF
that ordinary LICQ holds for REF at {{formula:936c21ef-d9a2-4ccb-bef7-46feeb0413de}} . Similarly, assumption {{formula:fde72e77-ffc5-4620-b074-ec8aa595f37a}}
and Theorem REF imply that the strong second-order sufficiency conditions holds
for REF at {{formula:35672f76-2535-44ff-b4e3-fa1c9df72740}} . Standard results on the local convergence of nonsmooth
Newton methods then imply that all elements {{formula:b87a777d-bd38-4f79-87cb-078e35f95857}} are
nonsingular, see, e.g., {{cite:52e97e84d79fccf3ff101c78781b80cf164800ed}}, {{cite:9b24ceed8b83c48a04d2407dbf3252eeedcb9252}}, {{cite:46f294d19684db180984cef32d06364ca395946c}}.
| m | 574781bcbb8bd6277afd8618f17743d8 |
All models used the PaiNN architecture {{cite:41faa3b40f207c9273b52b8cc9a8c7becf78346b}} and were implemented in PyTorch {{cite:9f3d313c43fa65f16b7c3e1bd0ef0947be39c7f0}}. As in our previous work, we used five convolutions instead of the three used originally, as this substantially improved model performance {{cite:85a5a6e2a0f6a196bc9bcf6b8c38c1e38ff4d966}}. We also allowed the {{formula:6edf04ff-c548-4a53-8f31-2ea36476bfc2}} values in the radial basis functions to be updated during training. Remaining hyperparameters can be found in Ref. {{cite:85a5a6e2a0f6a196bc9bcf6b8c38c1e38ff4d966}}, and an in-depth explanation of their meaning can be found in Ref. {{cite:41faa3b40f207c9273b52b8cc9a8c7becf78346b}}. Further hyperparameter optimization is likely possible {{cite:ea76607612a6fc628dbc24d0ba6d34468e85af3f}}.
| m | a7c8612cc8d04c23eff28c01c4e3460d |
As this method does not require any clean counterpart of the real-noisy image, it is more suitable to compare with the denoising algorithms, which also do not require noisy-clean pair and can only operate on a single real-noisy observation of an image. Hence, 4 such widely-used denoising algorithms, namely: BM3D {{cite:09e6a3b8f1845c6ce2dc28e456cb9305bc6a52af}}, non-local means (NLM){{cite:183105dead03b58cde98cb0b966522479b4c67a3}} , wavelet-based denoising {{cite:9aeb2269dabd36b04f5509cb48419841583177a3}}, and total-variation (TV) based denoising {{cite:5d12b62e79cb51e4d99152ee83a616ce197d853b}}, were compared with our model's output. Table 1 contains the quantitative results in terms of mean and standard deviation of all different methods in tabulated form. The proposed method outperformed other existing approaches in terms of widely used image quality metrics PSNR and SSIM {{cite:f43c810b4d09e392af6cd519dedf88a54844f21a}}. For comparison and as an upper limit baseline, the results corresponding to noisy-clean paired image supervised training are also shown in the table. In Fig. 2 results of the methods are shown for visual comparison between the proposed method and other methods. Fig. 2 (a) is the original noisy image, (b) is the ground truth, (c-f) and (h) are the outputs given by traditional algorithms BM3D, NLM, TV-based denoising,and wavelet-denoising, paired training respectively. It can be observed that these methods may result in blurriness, low-level artifacts, and residual noises compared to the output of our proposed method, which is significantly cleaner in comparison. The comparative baseline results with noisy-clean paired images are also shown in the same figure. Some more example results are shown in the attached supplementary file.
| r | f51dfa6afb58c407b8be13f4b090d9d6 |
DeepTaylor {{cite:d0e2b97374523061376d72c2dc30e9208c5fc7d1}} has provided an approach to generating specific positive evidence for a given prediction. The deepTaylor approach of XAI is useful for justifying CNN-based classification. It explains without changing underlying architecture, this property makes it an effective XAI tool. DeepExplain provides a unified framework using gradient and perturbation-based attribution methods {{cite:1a8896760378f2205170f74885ff994fd82f789b}} {{cite:dfd4b14e01d51aa0c6452805df8e615e283b89d5}}.
| m | dfe0351eff4c501a8ddae8e133abcdd5 |
where {{formula:42df64b0-e25d-4dd5-93f5-507637db544c}} denotes the Hilbert space dimension
corresponding to states residing in {{formula:cf6ebbe7-f7c6-4d98-9c58-2d9a9d558c0f}} with open boundary condition
and the states {{formula:415cd7b6-fbd7-4e99-bb94-09b2e7b3914a}} and {{formula:7645bcdc-e03a-4e13-bb54-2a12cbf259eb}} have weights in
region {{formula:8997e505-65cb-492a-a492-2f7da0819b3d}} {{cite:7eb228a78ddf9965d5feb3827a428595683da23e}}. While carrying out this procedure, one
has to be careful in excluding states where the right end of {{formula:900e4012-f957-4350-867b-142a3ab7752b}} and
the left end of {{formula:96173da5-3425-4ff1-b52e-16b61f1b5be8}} both has spin-up (or dipoles) since these states
were not part of the Hilbert space of the full chain owing to the
constraint. The half-chain entanglement {{formula:5196c4d8-aaac-4da7-b58e-6a24ba3dc0f4}} can then be
obtained numerically using
{{formula:a12252b8-e4b4-4e39-9359-96bf8d2501dd}}
| m | c65d9af3b351fd22b6c1bc703d4bc84a |
We selected 100 images, 10 from each of 10 fruit categories including Banana, Custard Apple, Fig, Granny Smith, Jackfruit, Lemon, Orange, Pineapple, Pomegranate, and Strawberry (images were randomly picked using web search). All of these categories exist in the ImageNet dataset {{cite:71bd693d9157dc2dff1c50260c5e6382bb832809}}, therefore pretrained models on ImageNet should be able to recognize images from these categories. Samples from the dataset are shown in Fig. REF .
| r | 49dcbb1a34eae97ce0d4a35952895379 |
Previously, Soloduchin & Shamir investigated rhythmogenesis using the framework of two neuronal populations with reciprocal inhibition and short term adaptation in the form of firing rate adaptation {{cite:bb2c48225d5957ae0f66f6cda65b6f3be80c897a}}, {{cite:1d3253f7137b881a56bab39446849e719aa2b4f5}}. The network motif of reciprocal inhibition has been widely reported in the central nervous system {{cite:a2bacc4eb2e945c03154992b7319e5412ed2a434}}, {{cite:67022f0d5b57cad08168d470576aa15bd0a464e9}}, {{cite:17b4dbb8e2e70599253ef4d94b58a4a67965dbc5}}. However, it is mainly associated with winner-take-all like competition {{cite:68fd09c30c3aa61954c3c71e309e305532f87e69}}, {{cite:ac4949d311374442443bf8c2db0d79fe37c5bc00}}, {{cite:9ce14f97e57613d8c62b5634eace7e6b7d1cadfb}}, {{cite:22a229b08ef8c10d329ed04b0dacae8faf6ef564}}, {{cite:1275c2d3d489bfc1767ce9c3758e4517e7021a76}}, {{cite:704e2e22595e9fa6d246ffb29bd19db55296ff7a}}, {{cite:fb575d3bc68b766f9000df9ab44bdd2ffb14105a}} rather than generating rhythmic activity (but see {{cite:de9050e1ddc7601f01e0243921ed304512394488}}, {{cite:fc86dd17b2b11f20414fc705fdbe953ea2705abc}} in the spinal cord). Here, rhythmogenesis was studied in the framework of a network that is considered a valid hypothesis for generating gamma rhythm in the brain {{cite:54b6cf6ac3b2f50019af2940f6a5859acf7ab5c1}}, {{cite:251fea5794ab0f77400b6e3134446eafbf72b22d}}, {{cite:4ab67375ce2e6536e1462e4069407a9fff46d6a0}}.
| d | 6be47f6e0d7d302a662ff0bf06f05759 |
In order to allow for a general construction of the auxiliary cross section, the aforementioned color and spin correlations are implemented into the factorization formula by realizing the splitting functions {{formula:231b5142-0170-4d7c-82b9-fa56d897ec8f}} as operators that act on matrix elements which are defined as abstract objects in color and spin space. For this purpose we make use of the of conventions and the notation established in Refs. {{cite:697bc828ed6db616f711f6d73c2e9dbae620ddcc}}, {{cite:4c1fe50303861da20f947d2710a61b9e17748140}} which we introduce in the following. That is, colored particles in the initial state are labelled by {{formula:43897f36-2d42-4c25-8851-799fa06d12f2}} and those in the final state by {{formula:80ce4353-aa57-4fe2-a4c1-c106b34d23d2}} . Since non-colored particles are irrelevant for the subtraction procedure, they are suppressed in the notation. Scattering amplitudes are considered as objects in an abstract vector space spanned by the spins {{formula:b071bd35-15f9-4622-bc5e-d0dd634c71c1}} and colors {{formula:b01dd760-fade-488d-bac9-c3a0821d70bc}} of all colored particles involved in the process
{{formula:aa7c9a57-997e-4224-b0ff-a8b3cf3773d5}}
| m | 2cc21b56f0d47655de7989290832f5a6 |
For {{formula:80730ed1-a28d-43aa-859b-a8879a5335c5}} , the system (REF ) becomes the Hunter-Saxton (HS) equation {{cite:338b527f9af8a134fb64c58a87f8eea464f0daf6}},
which models the propagation of weakly nonlinear orientation waves in a massive
nematic liquid crystal director field.
Here, {{formula:60cfa02e-323d-4f1f-9f35-f3345b6651dc}} stands for the director field of a nematic liquid crystal,
{{formula:13d3820c-264d-4bb9-a2d1-bdad985dc6d2}} is a
space variable in a reference frame moving with the linearized wave velocity,
and {{formula:89570a40-8c39-4bab-b5ad-7696497ac184}} is a slow time variable.
The field of unit vectors {{formula:8dadffc3-4070-4474-930e-9016c991233f}} describes
the orientation of the molecules {{cite:0ed400a26dafe8c6f535fa50a575c88200d93be7}}, {{cite:338b527f9af8a134fb64c58a87f8eea464f0daf6}}, {{cite:679650efeb87b6936e825cd7fdbb446ee41de136}}.
| i | e90be5fb9100c6628382ccaec939eebc |
Branching fractions are determined relative to {{formula:fb698e43-df54-4b26-8557-7d89743aca1f}}
and the known value {{cite:7cffa256bc39b95f8ab5b623dbf1695c7d451708}} is used to determine the absolute branching fractions.
Efficiency corrections and corrections due to the fragmentation fraction ({{formula:b6118fa9-8687-4ac9-be58-7d8029a9fc6d}} ) are applied.
The results are quoted for {{formula:12677003-451c-48e2-95a4-a65f662e156e}} or {{formula:dfc963f8-245b-4dba-b6cd-95925886200e}} , according to
the expectation for each decay
{{formula:d42b1a52-b3c4-40e6-8a3e-53147f8537b6}}
| r | 58a64f0d13f0f93da9988b85a29a19fb |
Thirdly, Yan et al. proposes a end-to-end offline trained RGB-D tracker DeT {{cite:fbc8b65274b5ead5d110e26b2321b73fb458f316}} based on the framework of the RGB-only trackers ATOM {{cite:8ecf609506dc3ff2fb676d551ed39e5d55edefef}} and DiMP {{cite:32265959567b9ad5742fcffbb11313ad593ff6b2}}, using a additional feature fusion module.
As far as we know, it is the first offline trained RGB-D tracking method.
Considering the fact that the offline trained algorithms require the support of large-scale data, they generate a considerable number of RGB-D videos by conducting monocular depth prediction on existing RGB tracking training data.
With a small amount of real RGB-D videos of the training set of DepthTrack and a large number of generated RGB-D videos, the offline trained RGB-D tracker DeT achieves remarkable performance on the testing set of DepthTrack.
| m | be754a61da947fd9f5046918291fd0af |
Following the standard evaluation protocol {{cite:b5cff9fa58ba6951c0c98ea08635dd27486ace60}}, {{cite:74fc6deba21d6e5298df948921acfc7fdc32ca8c}} for incremental object detection, we group classes from Pascal VOC 2007 {{cite:9acc9ad699c1cbffefbc97d13b1983cf1267ab7f}} into two tasks. Three different task combinations are considered here. We initially learn 10, 15 or 19 classes, and then introduce 10, 5 or one class as the second task, respectively. Table REF shows the results of this experiment. The first two rows in each section give the upper-bound and the accuracy after learning the first task.
The `Std Training' row shows how the performance on previous classes deteriorate when simply finetuning the model on the new class instances. The next three rows titled Shmelkov {{cite:74fc6deba21d6e5298df948921acfc7fdc32ca8c}}, Faster ILOD {{cite:080910085f2cbeb581bb889b2352a7bd351adb54}} and ORE {{cite:cba60ca5e2152f90e73757abdce06f22e78fa9c8}} show how existing methods help to address catastrophic forgetting. We add ELI to iOD {{cite:b5cff9fa58ba6951c0c98ea08635dd27486ace60}}, the current state-of-the-art method, to improve its mAP by {{formula:ed98e1fa-086f-4978-8993-dfcd7f11ad2b}} , {{formula:0048cd7a-a42c-46e2-89af-30b0796e072d}} and {{formula:b4c62201-ca37-4f34-9053-5533321b5e6e}} while adding 10, 5 and one class respectively, to a detector trained on the rest. This improvement can be attributed to the effectiveness of ELI in aligning the latent representations to reduce forgetting.
These results also demonstrate that ELI is an effective plug-and-play method to reduce forgetting, across classification and detection tasks. Fig. REF shows our qualitative results.
{{table:81940608-e894-47c1-9113-deb2239891a3}} | r | 7d7a1450a157d1a1380a3b4879dc88a9 |
This work generalizes one of the most important archetype of quantum interferometry – namely, the single MZI with coherent{{formula:10958e77-cd62-45f9-b9d9-cd3685b48ee2}} squeezed-vacuum light {{cite:484928206899a56bfc4f2ec6ef689bfb783b29bc}}, {{cite:590d20e0b59cca712c6e424b6308896bbcff1a89}}, {{cite:9df78077aeb7a2cf2961e44f2ea9263d774e43bd}}, {{cite:198334b4913498636bb106aec0e1f3ca5e8f6e47}}, {{cite:1984c72413cfc0be83a53436d9fb41506efe80df}}, {{cite:9dc39839aaade23315adbe4bf6528ff5325b63f1}}, {{cite:54efafa0e015496aaa1dc52f4e4359f35137da0a}} – to a distributed sensor network composed by {{formula:c625a245-bfd7-4362-9796-50ca8e3ce9b2}} MZIs, see Fig. REF (a).
The multiphase estimation analysis is based on a method of moments requiring local and independent photocounting at the output of each MZI.
This avoids to recombine the phase-shifted modes using a second multimode beam splitter {{cite:448dff1c0bfe04b020cd9b4c4225c714a9f66731}}, {{cite:45373a7b5de12534861f21660cb7e3b69da0ef9f}}, {{cite:fcef3b470481dcba743556fc1647258f8168f41f}}.
The scheme is thus optimal to realize a highly spatially-separated sensor using a multimode entangled state of a large number of particles.
In particular, our MZI sensor network is characterized by different regimes reaching {{formula:10b17b2a-06b3-4a21-a103-07c9d42fe6a7}} and the Heisenberg scaling {{formula:9a9c6a56-b513-498c-82eb-84f04190ae75}} .
| d | e4f4c284a893b006727070b16fd975fe |
For large baths, as defined by eqs. (REF ) and (REF ), the minimal late-time RT surfaces connect the boundary points by crossing the brane. From the brane perspective, this corresponds to the formation of a quantum extremal island in the gravitating region, as described in {{cite:78098cb492e5284d4e22db2209282878a65d2997}}. In the high tension regime (i.e., large {{formula:18003d0b-8795-494b-8f9c-9658ae8b1fe3}} or small {{formula:a3877af6-0d17-45ee-9021-8c58f8fceacf}} ), the late-time entropy is approximately twice the horizon entropy of the two-dimensional black hole – see eqs. (REF ) and (REF ). Thus this scenario evades the the information loss paradox and instead recoveres the expected Page curve.
| d | 062dec0e5ee663967ff1090e922cc710 |
MuRP {{cite:4355d93232b99307d5c9f753c438c55b57c0cb35}}: By establishing a comparison with word analogies through hyperbolic distances {{cite:679bc75c53257df2a6fdf4e8374a9a4fa41c2454}}, the authors propose a scoring function based on relation-specific Möbius multiplication on the head entity, and Möbius addition {{cite:a7a9f784735982171d545f24d8d34c057831fed3}} on the tail entity:
{{formula:c0cc7a71-7e49-45a6-b167-b91f0086044d}}
| m | 87903c9532f5ce525e2cef7653f334ac |
The effectiveness of our proposed model is evaluated by comparing with several existing baselines such as Textrank {{cite:f6413248136c66f26a03d5abac94320f9afb1385}}, LexRank {{cite:f6b65e65e166617afae31479d08cb440b277c980}}, SumBasic {{cite:c82c19f0bd647fc76c77c6ada8ff829609a7c10a}}, and KL Greedy {{cite:d13053e652cc3434a408fad7e77ea42ac363756b}}.
In particular, we analyze our GAE-ISumm method on TELSUM as well as on other existing Indian language datasets.
We also describe the performance of each component of our proposed method in the ablation study.
| r | 217e31b22485be884a0adf98ae3d6d1f |
This result is essentially due to Tanguy {{cite:fe4eaf37c72ffca5499884ba29c2ba7f8d04d318}}, although he proved it under the additional assumption that the covariance kernel {{formula:52574191-785a-43b1-9b96-7a318f619b91}} is non-increasing, and hence only for processes which are positively-correlated. Since we wish to work with non-positively-correlated processes, in Section we show how to adapt the proof of {{cite:fe4eaf37c72ffca5499884ba29c2ba7f8d04d318}} to lift this assumption. The bound in (REF ) is sharp in the sense that it captures the exponential (right-)tail of the limiting Gumbel random variable in (REF ) on the correct scale {{formula:9dc1670d-72b7-4f45-8112-902823ff56c3}} ; indeed the classical statement of Gaussian concentration of the supremum (see, e.g., {{cite:1b898ebe7813182ec4b8b685f51b4725466f2180}})
{{formula:a272aae8-0662-4991-8df3-b70a0393e78c}}
| r | 215da755a52c3e78f1741fd3c92486be |
On the other hand, MONAH significantly outperformed Jefferson in Kinesics (body movements or postures, H3B). We refer the reader to {{cite:c349ba5aa2cc074765a1077754477340cd0ff9d7}} for the method of automatic extraction of body movement and posture. Some participants made suggestions of additional automatic extraction of body movement such as “(look down)". Using a combination of gaze / pose extraction and timestamps, such information can be computed with the help of open-source libraries like OpenFace (gaze) {{cite:beabd69e50bab6b76c4b86b4ecc0add047deccc0}} and OpenPose (pose) {{cite:596cf290a236594a7339eb7e12b07c4bba5a0ff1}}. As for manual methods, there were other systems that included manually supplemented drawings or photographs {{cite:052db9a2fa5551638682c03e1a9add5d75151d9e}}, {{cite:6fb88224342dd8670a5109ad176774532e8dfb13}}, {{cite:aea5470c0e1bf9db0b9b15988b841924bcfeda23}}, these systems were out-of-scope of this study because they are not text-based.
| r | 623208f864f9e2b5e089feb07026afc1 |
In addition, we find that the tracking performance is easily affected by the challenge of sudden camera motion, which frequently occurs in RGBT tracking task.
The major reason is that under such challenge search windows are hardly cover target objects, which would lead to tracking failure.
Common attempts are to expand search region {{cite:653461c6121e0ec44fc81651d9d1e66a7712a4be}} and perform global search {{cite:77363083210766b30b18a72bdd8051967eb18daa}}, but these methods bring more background information and thus increase the risk of model drift.
Meanwhile, the computational cost is usually greatly increased.
To deal with this problem, we develop a simple yet effective resampling method based on a fast optical flow algorithm, DisFlow {{cite:7ed8d8667aa51829d97ea79f132a7bde519fca68}}.
By comparing with a predefined threshold, we can judge whether the sudden camera motion occurs or not.
If occurs, we resample candidate target regions along the direction and magnitude of camera motion.
Note that our resampling method does not increase computational cost much since we execute it only when the tracking failure caused by sudden camera motion is detected and the optical flow computation is only performed on the local regions around target objects.
| i | 1a4f679f1b8bc3d19864b52331a65935 |
Let us mention that currently QAOA and Quantum Annealing are the two prominent and competing methods for solving combinatorial optimization with quantum hardware. While quantum annealers have currently a remarkable number of qubits (over 5600 by D-Wave) compared to the largest gate-based quantum computer (289 qubits by QuEra Computing Inc. {{cite:d13c0afdef322484f4535e3e11b12b16f07f43dd}}), there is also a significant difference in the numbers of qubits required by the two methods. As an example, let us again consider TSP. As it was shown, for the gate-based model, it is enough to use {{formula:3f88199c-32c5-46c8-98cb-f4c02cc700c8}} qubits {{cite:0e2f81cb3dcaaa04e9831149acd4f1ffbe0283bb}}, {{cite:37ea84886b62a973566e463ef3dd33539cb2565c}}, or even {{formula:b2f3c25e-fda5-4755-946b-5dec9afad776}} qubits as presented in {{cite:83e437e80eb9e65a62683f3f44eb6f1690a4e246}} or in this paper. However, in the case of quantum annealers, one should also take into account the limitation of only 2-local interactions on a limited 2D graph topology. Assuming that the interactions for future quantum annealers will still follow 2D or even 3D topology, the physical interaction graph will have a bounded degree. This means that the number of interactions available on the machine will grow proportionally to the number of physical qubits. However, since for the TSP we need {{formula:3bebf2ee-3ef2-454c-8191-fa833bd67906}} interactions for a QUBO {{cite:37ea84886b62a973566e463ef3dd33539cb2565c}}, {{cite:83e437e80eb9e65a62683f3f44eb6f1690a4e246}}, the embedding will require at least {{formula:14deaa9b-98dc-4b8a-8593-1be1a77dd07b}} physical qubits, and thus the quantum annealers need to grow much faster than the gate-based quantum computers. Note that the same problem will occur for any optimization problem defined over permutations.
| d | ee89d350875b1ce6ece22583f7309e76 |
As shown in Tab. REF -Tab. REF , we provide the results of four benchmark datasets (Office + Caltech-10, Office-31, Office-Home and VisDA-2017). In this experiment, C {{formula:21786f13-1e9e-48cc-8404-a781860bbed8}} A means learning from existing domain C, and transferring knowledge to classify domain A. These results indicate that deep learning-based methods usually achieve better performance than traditional methods. However, some traditional methods ({{cite:5520c4faf4fd0ba7c2ca21b83a20d07f8f367789}}, {{cite:9c5d1f617c5de1fcc45a240d2f83dd4a79dbb03b}}) observe higher accuracy than some deep learning-based methods. This is mainly because the extracted features are from pre-trained deep neural networks. Therefore, there is a trend of combining traditional based methods with deep learning features. Also, deep learning models with pseudo-labeling techniques achieve promising results.
| r | 5e4cc8d39f89f22cc3e207d26d4d75f2 |
Multiwavelength observations by both IACTs and Fermi satellite may provide some special cases. For example, GRB 190829A was clearly detected by HESS but not detected by Fermi-LAT {{cite:391c72e40aa4cb44af293a114339cadc6bc9d0eb}}. {{cite:afe2d388545fc1fd4101dcffca78cc00d9128637}} and {{cite:7a7b14c018a3172ad02b2553099f20b62a3a5629}} suggested that SSC or IC mechanism produces a detectable flux density for HESS in 80 GeV but a very low flux density in 100 MeV that cannot be observed by Fermi-LAT. Thus, this peculiar phenomenon of GRB 190829A can be explained. In our model, the jitter radiation with the turbulent feature can reproduce the TeV emissions of GRB 190114C and GRB 180720B.
Here, we note that the turbulent energy dissipation has the intermittent feature both in space and in time {{cite:811a047dcd5676e48fa696439af4d470697b969a}}. If the energy spectrum of the turbulence has the intermittent feature, the corresponding intermittent feature in the jitter radiation spectrum can be also shown. The intermittent feature in the turbulent flow of GRB 190829A may induce the intermittent feature of the high-energy emissions in the different energy bands. Thus, the detections by Fermi-GBM and HESS and the non-detection by Fermi-LAT can be explained by the turbulent intermittency. The intermittent current sheets in the kinetic turbulence were investigated, and it may have an important application in high-energy astrophysical objects {{cite:356d59f5b0a9ef0756559e820ad759a707963e57}}. However, the intermittent property is not universal in turbulent flows, and it is strongly dependent on the
space dimension, coherent structure and statistical fluctuations in a certain turbulent flow. How to link turbulent intermittency to radiation phenomena in high-energy objects is a challenge. We shall further explore this interesting issue in the future.
| d | cae25630c35d7c93a8030c3c4ef53296 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.