text stringlengths 54 548k | label stringclasses 4 values | id_ stringlengths 32 32 |
|---|---|---|
The input parameters are taken as
{{formula:16650a28-4f42-4432-a7af-48caea761463}} and {{formula:f4e33755-cd0a-4241-b161-b513af1cfee4}} {{cite:9e6e6f8ceb883bb50f931f7f4d7187bc934ea5c3}}, {{cite:667039e3796cd381c4e557e07933cda114283ae7}}, {{cite:6c26b63f2199955a7a34138e581b3d16429bcbf2}},
and {{formula:9bfcbd62-c1a2-4e8d-84df-482ddc3cbe5e}} is set as the running charm mass
{{formula:13e1c203-5103-4c62-9b79-6205fe9b57ec}} {{cite:340c829cf158fba09a466c872c01733c8bde8637}} at first.
As one knows, the QCD sum rule method has made approximations
in the OPE of the correlation functions and
introduced a very complicated and largely unknown
structure of the hadronic dispersion integrals in the phenomenological side.
In this way, complying with the
criterion of sum rule analysis, one should find
appropriate work windows for both the continuum threshold {{formula:b9f5af1e-0070-4b20-a17a-e662633c9b0e}} and the Borel
parameter {{formula:8a21a4d6-d8d8-454f-9325-a0ad16dfea72}} in
which the two sides of QCD sum rules have a good overlap and information on the hadronic resonance can be reliably extracted.
In phenomenology, the threshold {{formula:bf23cb8a-24c6-48ec-837d-067836c8011f}} is the energy which
characterizes the beginning of the continuum state
and the gap {{formula:1060c21a-de5b-4c00-beea-4cb3d654739e}}
is typically evaluated to be about {{formula:b8bb408e-b558-4d12-be25-ae19946bb166}} {{cite:667039e3796cd381c4e557e07933cda114283ae7}}.
Meanwhile, the proper Borel window of {{formula:a8c84ae7-61db-433b-9eee-f367e53387fb}} can be found by analyzing the OPE convergence and the pole dominance:
the lower value of {{formula:47044e45-d52a-40f6-b169-ac5577f0197c}} is obtained by considering the OPE
convergence, and
the upper one of {{formula:0c0bedce-8238-45dd-8bb7-a0d1dbe239f9}} is get from
the condition that the pole contribution
should be bigger than the continuum contribution.
| d | 33d5ae03299b25b8bacd921bc12caa60 |
For all these SOTA methods, we use the public available official implementations to get the temporal grounding results. The results of the proposed test-iid and test-ood sets on two datasets come from the same model finetuned on the val set. For more fair comparisons, we have unified the feature representations of the videos and sentence queries. To cater for most of TSGV methods, we use I3D feature {{cite:2277dc87fb75f90a9812159e5c76e63595d070c2}} for the videos in dataset Charades-STA (Charades-CD), and C3D feature {{cite:76bca13dd5bea51df0b38f12d398d85d06a226c3}} for the videos in dataset ActivityNet Captions (Activity-CD). Each word in the query sentences is encoded by a pretrained GloVe {{cite:0dfca059c58d3b409577b9ab594192cf6d87c1d7}} word embedding.
{{table:26b147c6-46b5-49a8-b692-b02a16b77dc8}} | m | b14070c0ae83fc00af6d57f20f8850e3 |
The phase behavior for the vapor-liquid transition in this model in various space dimensions
have been studied {{cite:b7449be78aff1f3091dd13632d19ae44ff5869e4}}, {{cite:e64ce19e69e08e30d75aa4ebdd02e647eca44782}}, {{cite:cb50c90167f471f44729307665e0340c2552be54}} via Monte Carlo simulations {{cite:96ee47139565a279487fa60cbb65c42e2b138ebe}}, {{cite:55d4747b77764f48a3a06699bfbd16e7a438accc}}.
In this paper we are interested in {{formula:39852ed9-245e-4356-902e-6d2de7cfcf29}} . In this dimension the critical values of the temperature
({{formula:0e2d410b-b4db-47e6-bf3c-d9c95ca34847}} ) and the overall number density ({{formula:623a71cd-7f48-4c1e-83be-a2532b71a501}} ) are {{cite:e64ce19e69e08e30d75aa4ebdd02e647eca44782}}, {{cite:cb50c90167f471f44729307665e0340c2552be54}} respectively
{{formula:70d44e73-df95-4f24-86fc-3e16fd442551}} and {{formula:5d269da4-60d2-40ac-9fb7-5045235027af}} , where {{formula:74b08b19-317f-4636-b7a0-d44363a0ffe5}} is the Boltzmann constant.
For the study of aging dynamics related to the growth in droplet morphology we will consider {{formula:c369e7cf-c807-4ecd-bd31-1a3e660a8601}} ,
while for the percolating case simulations will be performed with {{formula:fb894fd8-45cf-4f98-9c02-46b82980c9b8}} . Of course,
both the densities fall inside the coexistence curve at the considered values of {{formula:c1de00bc-bec7-47ad-ac20-a5281d579b33}} , numbers for
which we will mention in appropriate places.
| m | 3df53b683bb35c2f6f166937f76a8625 |
Existing studies from relevant literature apply various lightweight heuristics {{cite:e580fada600d6aa2dd9a3ebfaee3a4c89d344be2}} and query reformulation strategies {{cite:d61e748d43da063f91dc179218041c9eac9c5e82}}, {{cite:0cca3c05733ce3253603e38190fec71ae2e6e669}}, {{cite:b31eec18a8c8cad8c5f1421a8ee6168183f661a1}}, {{cite:16f352b3f57e441cd759d430ce8771a1296b7934}}. They also perform different query quality analyses {{cite:71181764ed202af4897664e4aac0753dc6b42607}}, {{cite:61f6c6b96873f1eb683823be2e71249d999d2998}}, {{cite:e1497781b9a3972a578a68190bf803ab5d23f134}}, {{cite:0cca3c05733ce3253603e38190fec71ae2e6e669}} and data mining activities {{cite:c25b01b96b3f77fe5eba2f62cce186c50c906398}}, {{cite:4aad10cfeaae5b58c1d88645174ed49381d07efa}}, {{cite:16f352b3f57e441cd759d430ce8771a1296b7934}}.
However, most of these approaches expect a developer to provide the initial search query which they can improve upon.
Unfortunately, preparing such a query is often a non-trivial task for the developers as well {{cite:e580fada600d6aa2dd9a3ebfaee3a4c89d344be2}}, {{cite:769c2a83e2ced153491808fb107f0388c59038f5}}.
One can think of the whole change request as the initial query. However, this might lead the developers to major query reformulation tasks {{cite:d61e748d43da063f91dc179218041c9eac9c5e82}} or suboptimal solutions.
{{cite:e580fada600d6aa2dd9a3ebfaee3a4c89d344be2}} propose the only heuristic model for automatically identifying initial search terms for a change task where they consider different heuristics associated with frequency, location, parts of speech and notation of the terms from the change request. Although their model is found promising
according to the preliminary evaluation, it suffers from two major limitations.
First, their model is neither trained using a large dataset (
meaning:NTF . i.e catcode:NTF a i.e., i.e., uses only 20 change requests) nor cross-validated using the change requests from multiple software projects (
meaning:NTF . i.e catcode:NTF a i.e., i.e., uses only one project).
Thus, the model is yet to be matured and reliable.
Second, tf–idf is the most dominating feature of their model which is subject to the size of the change request dataset for inverse document frequency (idf) calculation.
Hence, their model is likely to be affected significantly by the size of the dataset. That is, it might require frequent re-training to keep itself useful for search term identification.
Thus, an approach that overcomes such limitations and yet can identify appropriate search terms from a given change request is warranted.
{{figure:bd3798a3-f261-4da3-bf0e-598999e33481}} | i | 244ea5b4062d9220b1a8a6b163713dd3 |
In Fig. REF , we observe that LiRE{{formula:739f2a05-6765-44fc-a3df-132c8846f889}} OMP always improves support recovery over LASSO (except for the white region where the support is never fully recovered, with and without LiRE) and is also significantly faster since LASSO solver's complexity is quadratic or cubic in {{formula:2e62c5dd-f36a-42f8-abf6-ccd76a1e9d17}} , depending on the sparsity level {{cite:b50833397048230b1bfa89e29e858efdaa4ddb3f}}. Outside the white region the improvement is significant, mostly by more than {{formula:fc6a7d8f-3eee-49f3-950d-d647d48d6e22}} .
| d | afa7b45d760ebb34e619c8931d731bbf |
blackGiven a scene point cloud, we first use the backbone network PointNet++ {{cite:b037705a82da54abcaf882d0a2e02e2c3b211547}} to encode features, then simultaneously attach three parallel decoders: instance segmentation, 6-DoF grasp pose, and collision detection. These three heads respectively output predicted point-wise instances, grasp configurations, and collisions. At inference phase, grasps from the same instance without collision are grouped, and a pose non-maximum suppression algorithm is proposed to form the final grasps.
| m | 3db72eb61ca6f8ca945461ec296ab9f7 |
In addition, because the precoding/combining matrices are computed only once per coherence block, the complexity difference between MF and ZF/MMSE is relatively small (the bulk of the complexity comes from FFTs and matrix-vector multiplications performed on a per-symbol basis {{cite:d32ee6a9074f644b81536db3fc802e0b6bd423b1}}). In addition, because the FD massive MIMO considered has many antennas ({{formula:c873a99b-da15-4c4b-9df0-b059432d8dbb}} ) and a high SNR environment, both ZF and MMSE have similar performance. Furthermore, ZF can provide a clear performance comparison with SI subtraction to satisfy our main contribution. For these reasons, the ZF is considered for downlink transmission. Future work will consider various precoding/combining matrices, including MRT/MRC, generalized MMSE and so on. From the antenna configuration of {{formula:c7fcb24e-c1ba-4b73-94cc-947af3283491}} , additional antennas can be utilized for spatial suppression {{cite:e67d7d48090a26dc060c76016bf988d5d1a6e300}}, {{cite:433a057dc06190a2b04b956fafcd59fbe5a2fa83}}, {{cite:4de2e301095832d831270395554757f30f43cd86}}. Spatial suppression-based algorithms for massive MIMO systems were proposed in {{cite:433a057dc06190a2b04b956fafcd59fbe5a2fa83}}, {{cite:4de2e301095832d831270395554757f30f43cd86}}.
We adopt the extended ZF in {{cite:433a057dc06190a2b04b956fafcd59fbe5a2fa83}} for spatial suppression, which requires minimum computational complexity for nulling.
{{formula:7de911cf-7fa1-401e-80a2-32708b60480e}}
| m | def0dcc7fcfb0e008549ff8516b75c94 |
helps us to estimate critical exponent as {{formula:ee2e2217-fdbf-4834-b828-17e10cac741e}} . This value is again very close to {{formula:5c33feb8-2429-43f1-be55-2cf560df5c0b}} of 2D equilibrium Ising model within errors {{cite:9af7f80939380d103e4ffc5790cfc8dd7577e09b}}, {{cite:b8ed9a57b51dcd62a424c46f67dfb0cfffebff1f}}, {{cite:b00548ed24b57991502d7c4e368d6de6a4733ff1}}, {{cite:6e3607295ee6466c5553433d851de93808853327}}.
| r | bbfd95fa509a6540d6003a2d5d770d12 |
Nyström method {{cite:68e6be51ab3e1ddff3bd7c3a7913977faca400fb}} aims to calculate a low-rank approximation for a Gram matrix. For Transformers, the self-attention matrix can be viewed as a Gram matrix {{formula:82cfa8b7-6aff-4118-a23b-2b4ccbec4705}} with a Gaussian kernel {{formula:33c3999d-a015-4c1a-80c6-aaee0360ed51}} applied to the query {{formula:54265677-b898-464b-bd7b-6bde2dc22c26}} , with each element {{formula:78ef9b36-19b1-4ff9-87ea-239c822a9aa9}} expressed as:
{{formula:ff0e79d6-5e96-40f0-a2c2-72e801deab43}}
| m | 443b4bd0bcab8f3f98ee9551828b66a7 |
Are synthetic data rewards distributed to parties and their downstream ML task performances commensurate to their contributions?
We firstly assess whether our CGM framework can distribute synthetic data points {{formula:dcb8ad04-6e7b-4510-8c79-0687651779ee}} to each party {{formula:9ea867ef-0a8d-44fa-bf13-d2d6184cce7e}} as reward such that the closeness of the empirical distributions associated with {{formula:f04fe110-a868-4a84-bcc9-f6ce6982fa0e}} vs. reference dataset {{formula:20eb8ea0-86f9-455d-a291-5553cb8b9c11}} correlates well with its expected marginal contribution via the normalized Shapley value {{formula:f44dd3b0-f158-4a77-914f-8c360cc8d963}} . We quantify such a closeness using
4 metrics (which we take the negative of so that higher is better): (a) unbiased MMD estimate (REF ), (b) an estimate of reverse Kullback-Leibler divergence based on {{formula:c428dc72-e0c1-4894-ae53-a0557d8d8634}} -nearest neighbors {{cite:461ba7ae68c1f040d5e255c1dc14a48aa9ce781b}} averaged over {{formula:2157898f-f434-4276-93bd-83ff43859720}} , (c) Wasserstein-2 distance between multivariate Gaussians fit to {{formula:70ed827b-6146-4b5b-b49c-5771037d713a}} vs. {{formula:f44d7e83-b385-4aa6-8d95-8599f2ec5d63}} (i.e., how Fréchet Inception distance for evaluating GANs is computed {{cite:d81832f7d8fa583ae191413953dc1074cccb58f7}}), and (d) class imbalance {{formula:d9906c75-a4b5-40ad-b800-6786bf012ecf}} calculated with {{formula:6fed3f85-3051-48b9-9953-7e319005aa2b}} where {{formula:cf69034a-6d5a-4320-8a15-f2e2e6493192}} is the no. of classes and {{formula:4c5f4fec-f946-449d-81d6-773cc981151a}} is the proportion of data points in party {{formula:d0ea36c6-2975-4d64-b207-6a362677e33b}} 's combined dataset {{formula:d9bf9cf8-953d-43a5-8723-0596bc276a90}} belonging to class {{formula:8e86ffbe-9129-44fa-8000-286beaecd5e3}} . In all datasets, {{formula:cae07896-6cb1-43a4-a337-eb04b1aa578c}} is equally distributed among the classes and hence achieves a minimum for {{formula:5893fbfd-6858-4688-9cd4-afbeb2a2c72e}} . We also measure the correlation of the no. {{formula:0d68d449-141d-4016-9d90-db279614bc85}} of synthetic data points as reward to party {{formula:23c4ff68-7353-4465-b6b0-73e3f1c351bc}} with {{formula:8787f278-e62d-49d6-84ea-775122e2164a}} . Fig. REF shows results of the mean and standard error of the correlations over varying {{formula:1aefba19-9eca-4408-9f8c-10badf375926}} in the weighted sampling. It can be observed that across all splits, datasets, and metrics, the negative of the metrics and {{formula:4982c59d-568d-455b-8715-796987ca281c}} mostly display highly positive correlations with {{formula:734b9193-6874-471a-b27f-85ba57b7a578}} , as desired. We defer the discussion of the few negative correlations to Appendix REF .
| d | 437e8d81d1a9cb8edb985a5357202a8f |
To test the generation effect of our proposed model when cross-database, we conduct experiments on the LRW dataset {{cite:198e563bc36b506976d2a576384d4bf355724111}}. The example of generated frames can be seen in Fig. REF . Our proposed model and CRAN {{cite:bb06865f30dc6bc2f3ee431c36a27171d13c462f}} are trained on the GRID dataset, and DAVS {{cite:1e423a194e4bd22baee266bdca530f43353d65c6}} are trained on the LRW dataset. It can be seen that the images generated by DAVS are blurry, and the face texture of the images generated by CRAN are not well preserved. Our method can still generate high-quality talking face images when cross-database.
{{figure:8632ec59-c95d-4a75-854d-d29846a8c3b2}} | r | 8de6e292fbc3586c43d6de7a1e50ce11 |
We briefly outline the overall architecture of our model and then delve deeper into its individual components. Figure REF illustrates the VLC-BERT pipeline. Given an image with corresponding image regions {{formula:e7ecd9dc-35ba-4501-9726-597c903ef765}} precomputed using Fast RCNN {{cite:311f01a4714e79570b77920537b7e2ea13f3e2a8}} and a question {{formula:5fbc18cc-9ae4-4d6b-b382-5cb8bf3c303f}} related to the image, we generate commonsense inferences {{formula:2b49d15c-904f-4d6f-84ab-0320ed163f0b}} on the events and entities in the question phrase and two object tags {{formula:1ebcc8cd-895a-4575-a446-d126e5bce212}} , and select the set of commonsense inferences which is the most useful for answering the question, {{formula:7c07c2bb-3dfd-4656-bd2f-5c6f10440acb}} = {{{formula:58895a47-8d8b-4c56-b044-fe5c48f48949}} } (§REF ). Finally, we embed {{formula:b3395192-c5c4-40ab-a1cf-3f5c2c3a6255}} , {{formula:5946b2a7-a920-41ca-acfa-63878a52e74a}} and {{formula:6eb143b5-a7ae-4005-bc4e-4e32ab974318}} , as input to VLC-BERT and train it to predict an answer {{formula:6d27b86a-baca-4f64-a126-c373113c8b6d}} to {{formula:93eef2c8-c0ac-4a8f-a094-3d346c3d3882}} (§REF ).
| m | 66735fc2eded163a38fce169d1574cee |
This approach has been studied in various previous works {{cite:0fc5195f1380ba38cf30a0563332987382c65568}}, {{cite:52104f3e498318738374915ff4f46b75e2b845ce}}, {{cite:c2c6dab6565100d7bff871e158021076817bf1a2}}, {{cite:61ac21b717bb0b2c2ab4fa326e82296ab0895198}}, {{cite:89a524917b34cbd4dbee88e37a4bd13b80d98db1}}, {{cite:ef94c872468922345661a890d572d43fc215fba3}}, {{cite:69fae978c5a3eeba5a24eb44a03e9b7816d93adb}}, {{cite:5707393d3a89e9b819c8693150367472ce7c027d}}. See also the treatment by Townsend and collaborators for dynamical string tension {{cite:fe4a3eadb87a0407da8e4e3fb46d7e8f0b5554fe}}, {{cite:01986935d810025135383353e65d7f5d18ac54d1}}.
| i | decd9595707acb1504db28513e07c71b |
A physical object that can cause collisions. In adversarial machine learning, there are physical objects that can trick classifiers and object detectors into outputting attacker-desired decisions, such as a `Stop' sign with stickers that forces a classifier to output `Speed Limit' {{cite:b2c216364f3ce5d2363aedf258d0ee67dfdfc84b}}. One could hypothesize the existence of a similar object for perceptual hashing such that the resultant hash of the scene would always collide with a fixed pre-determined value. This can make the surveillance attack easier — the attacker has to fabricate an object and simply place it in the target location and only poison the hash database with a single value. This object could be as simple as an abstract piece of art. We have attempted to fabricate such an object by borrowing techniques from physical adversarial examples, but have not been successful. We suspect that this is because perceptual hashing functions like NeuralHash and PDQ are only invariant to small syntactic transformations and cannot generalize to larger semantic features of a scene. In contrast, machine learning image classifiers are trained to extract high level features of an image making them highly invariant to semantic transformations, which leads to the possibility of physical adversarial examples. A recent work which compares the success of expectation over transformation (the primary technique used to generate physical adversarial examples) with the semantic invariance of the model, corroborates the concept {{cite:2420e217e773dbef029c47afa7e65ff262cc53ab}}.
| d | 430639565ecc3014de05bb4ce77d20e5 |
The Stanford Online Products (SOP) dataset {{cite:4838599a6f1811f1ece31a4e1db76fd8b39b52a2}} does not contain ground-truth pose information for the object images. As detailed in Sec. REF , we obtain the pseudo ground-truth geometry information using LoFTR {{cite:d9f9fe0de0c6244f8c50369510f923b3651eaa02}} for matching and MAGSAC++ {{cite:a6437fa03904f1500cc2f723a7329a5246337d7c}} for robust estimation. We find that, even though these pseudo ground-truth poses are not entirely accurate, they are still useful for training with the Epipolar Loss.
Table REF shows a comprehensive comparison of our proposed methods with all the baselines.
We also include deep metric learning methods {{cite:c3666e0cfd2864df498d069639ce821f9d8f9d16}}, {{cite:81c9208be52da399e0d898001afad9a5b873f5a9}}, {{cite:edf54cf8d4703c9edbf19375f39a79059bb5f1d9}}, {{cite:cc0ae28a950efe877558d8d0435c2d0ed90774c4}} with reported numbers taken from {{cite:3c8a5bbdb17e44decd87b473559645ae078d062b}} into our comparisons. Our proposed implicit method outperforms all the baselines including the state-of-the-art Reranking Transformers {{cite:3c8a5bbdb17e44decd87b473559645ae078d062b}}.
| r | ee68dd4b5a639dea1e318a30ac2dc309 |
Relativistic quantum field theory is the adequate theoretical framework
to formulate the commonly accepted theory of the fundamental interactions,
the Standard Model of the strong and the
electroweak interactions {{cite:b33fe62cef9b38ab5e6c8db7d2d915348ab2f8e2}}, {{cite:a4e61ea4f14145285e6ec28b7432bcb172de9c83}}, {{cite:ef3945bb81052488575ec6dfb9432422caff2d9f}}, {{cite:89e4e106b1f722ed501922cdf7521ea873280199}}.
The Standard Model summarizes our present knowledge of the basic
constituents of matter and their interactions.
It is a gauge invariant quantum field theory based on
the symmetry group {{formula:42d852e2-f5a6-48b0-abcc-2a90f2fed476}} ,
with the colour group {{formula:3711d733-9eb3-4dbc-932b-a0cb4f1e4fa7}} for the strong interaction and
with {{formula:ae16b955-ac4b-482c-aea4-c0c66ed9ccec}} for the electroweak interaction
spontaneously broken by the Higgs mechanism.
The renormalizability of this class of theories
allows us to make precise predictions for measurable quantities
also in higher orders of the perturbative expansion, in terms of a
few input parameters.
The higher-order terms
contain the self-coupling of the vector bosons as well as their
interactions with the Higgs field and the top quark,
even for processes at lower energies
involving only light fermions.
Assuming the validity of the Standard Model, the presence of the
top quark and the Higgs boson in the loop contributions to electroweak
observables allows us to obtain indirect
significant bounds on their masses from
precision measurements of these observables.
The only unknown quantity at present is the Higgs boson.
Its mass is getting more and more
constrained by a comparison of the
Standard Model predictions with the experimental data,
preparing the ground for a crucial test at the LHC.
| i | 59b8d8f67e3082b504575cb9f3d6ab24 |
Representing filaments using regularized singularities is not a novel idea, as it was first employed for numerical purposes in the immersed boundary (IB) method of the early 1970s to model the flow patters inside the heart {{cite:f6358a90aaa5f123088c4c563c81785aea861c06}}. Later variations, including the force coupling method {{cite:34c639e9e038533fade5a3346c540c5d6cd93170}}, method of regularized Stokeslets {{cite:3771a2caf14fb67ef7e384ed02ab4f8cf083e382}}, {{cite:c9c6f6a8cef96930e25037ff1a7799081bdb7eb2}}, and Rotne-Prager-Yamakawa (RPY) tensor {{cite:88811456f07486c3fd714bc31c50fbfe98d53935}}, {{cite:238bfb4fb48f61d4e23a36d07e36effb09781146}}, focused on choosing the regularized delta function such that the Stokes equations are solvable analytically. The accuracy of using regularized singularities to approximate slender translating filaments has been studied numerically {{cite:d386715df944e92ab103b1cccd75b39e723012b0}} and analytically {{cite:9bd702044ffffe2c4608e82ca60e2d491b773e99}}, {{cite:4f3ffa3eaf30efb9820ecb7f2bd4f75b4668a760}}, {{cite:3bb70401945459dd1dc19e60b9c662ee79e5614b}}, {{cite:2a639ea9b40cdabbf34f80e126cbd9469d254d91}}. The conclusion of these studies is that regularization methods can effectively match slender body theory with a judicious choice of regularization radius {{formula:1aaed76b-f7a4-4c19-ac7c-aec60a635991}} . There has yet to be an analysis, however, of how regularized singularities perform in the context of rotation. For instance, do they reproduce the relationship between torque and rotational velocity for a straight cylinder? And what about the feedback of rotational velocity on translational motion for curved filaments?
| i | aa8e8e673b7d576666996f86e21517b3 |
Various oscillation modes of neutron stars are gaining much astrophysical interest because they can provide various information on the stellar properties {{cite:3f23b857acbfaeec2aff19a5a0df96b85ca18259}}, {{cite:bd597d51960ef06fdc4e7093731bacb54a70f7a9}}. For example, the {{formula:5e889142-5cdf-4584-a5d0-0420e51e52a6}} -mode oscillations of a neutron star are governed by the bulk properties, e.g., average density and radius {{cite:3be9d75961106b7a937205676a8bfad1356499d0}}, while the {{formula:ecd02ab5-955e-43ed-afdb-ac84d6064713}} -mode oscillations, with the restoring force being buoyancy, can reflect the interior structure {{cite:1702a280820d22874975aa23b802d73fb842e58c}}. As the gravitational waves (GWs) from the binary neutron star merger GW170817 have been detected {{cite:93f53f248eae1f796f82982b0690078010685541}} (which have already been used to put some constraints on the neutron star interior structure {{cite:5ad37449d08132d836b987fb22583177a0eaefb2}}), it is expected that people can directly detect the GWs from various neutron star oscillation modes and constrain neutron star properties {{cite:bea4aca8219e8fc15f57547c7f61dff32f3a4b03}}.
| i | 852657637d3414b45a5f2c1646d9abd9 |
Strength of augmentation: In most exemplar-based SSL methods, augmentation plays an important role since the main supervision signal is that the augmentation should not change the embedding much. Hence, recent methods, e.g., MoCo v2, SimCLR, and BYOL, use strong augmentations. We believe such aggressive augmentations on the target embeddings {{formula:01300ff2-1b11-4d43-a5ec-79615c15d2eb}} may add randomness to the learning process as some of those augmentation do not look natural, so the nearest neighbors will not be semantically close to the query image. Hence, in the “w/s” variation of our method, we use weaker augmentations for the target model to make {{formula:7edab944-ae2b-419f-8eda-0fc78018f374}} and {{formula:af970320-9691-4e61-acf4-ab8fb3f9e773}} less noisy while still using strong augmentations for the online model. This is inspired by {{cite:bf660900b1871f4a42ebccfa9e3809fd2b027237}} which uses weak augmentations in semi-supervised learning. This variation results in almost one point improvement over the regular variation where both encoders use strong augmentations. As shown in Fig. REF (c), the nearest neighbors are more pure in the “weak only” setting which is consistent with our above intuition. Our experiments show that BYOL also benefits from w/s augmentation to some extent. This is probably due to more robust target encoding.
| m | f5c0402ef680c5fbb604a991c0b13a39 |
We conduct several experiments to improve George's {{cite:992c81906ebc094e75db68579c101d5961817600}} architecture. We have found that the output channels in our architecture do not give much improvement. Firstly, we add GDN block in encoder and decoder modules, which substantially improves PSNR and SSIM. Secondly, we modify the hidden values in binarizer as {{formula:db86be98-7d48-4ae0-a0bd-41a5f19f2731}} to {{formula:5a5a92bd-bf3e-4d44-b884-f3fff00dc965}} , and in decoder hidden values as {{formula:40fd3ebc-4222-4200-bd65-a54108329612}} to {{formula:1194e9ef-66d5-4440-9a34-78e2068cafa0}} . Thirdly, we changed the encoder hidden values as {{formula:4515cdab-7164-47ea-815c-b0457ae4d51d}} to {{formula:3f7ccb00-fa0c-41b9-92aa-d363d86289bc}} combining experiments with GDN. Lastly, we set the encoder values as {{formula:bdb286d2-9eba-4d36-9638-dc40606002cf}} and remain changed in binarizer module, correspondingly. This work employed RNN cells with the implementation of GDN with the hyperparameters. The findings of our experiments are presented in Fig REF . Experiments have shown that the hidden values in binarizer module with GDN blocks {{formula:01d0690b-eb72-4f46-8f70-b4a6039b0a7e}} could be better than hidden values {{formula:a6c2a25b-0b10-4518-b6cb-1cfb19722ff5}} in terms of PSNR. This improvement remains the same in MS-SSIM. Finally, in the above-aforementioned experiments, the value of {{formula:c2d3b5a1-cd4e-4c09-aaa2-2cab67359312}} has been utilized in our network.
| d | 77e9b9b8c5a270ab3e53a1651dce136d |
Lemma 40 (Multiplicative matrix Chernoff – Theorem 1.1 of {{cite:066eb94e0b5e1ec327046bc6ee935a64be1acc1e}})
Let {{formula:5aa433af-c3d9-4628-96fa-552121d5aed5}} be independent, non-negative definite {{formula:6a536abe-0eb4-4acd-82e7-cbfd7b6ebc45}} matrices with {{formula:baa21ce9-8ce6-49e3-9ce8-f53acaa3d32b}} almost surely, for every {{formula:313eab21-dfb8-4d11-8ef4-e43bf5a49c01}} . Then, writing {{formula:64a22c70-e447-4370-9b93-3c61db827e7c}} , we have for every {{formula:f43e8b6b-3822-45da-889f-4de22774a584}} that
{{formula:5e1e3186-6c01-41db-9698-661b82cc72a9}}
| r | 34afa588388dffbe12ee932b6e74bfc4 |
In order to solve the anomalies one encounters in calculating the effective action, one can apply the so-called conformal dilaton gravity (CDG) model{{cite:70207467fa54358740525b7dc911b225f8a9d57d}}, {{cite:c3f02777b3fdf07e41f3e10a2b7c3a361cbbf96d}}, {{cite:b23c0f6838bf37f9aeaf142614c353195ea5e5d7}}, {{cite:57395a5e4544c74e94d94db83995e227a712bfcd}}, {{cite:6869fd2e88be9f4e70ed54a4564b3ae69eb10b22}}, {{cite:a87328fa3db3d150d36d4ce4af5258f0ce1dded3}}.
CDG is a promising route to tackle the problems arising in quantum gravity model, such as the loss of unitarity close to the horizon.
One assumes local conformal symmetry, which is spontaneously broken (for example by a quartic self-coupling of the Higgs field). Changing the symmetry of the action was also successful in the past, i.e., in the SM of particle physics.
A numerical investigation of a black hole solution of a non-vacuum CDG model, was performed{{cite:308391183fe50e3a3a4075d5500e670ca7c9a8fb}}.
The key feature in CDG, is the splitting of the metric tensor {{formula:298903f9-bf6d-4a9e-8f51-acf9e25b56da}} , with {{formula:e9c5ebe2-74e5-4a5b-9719-9dd62482eb17}} the dilaton field. Applying perturbation techniques (and renormalization/dimensional regularization), in order to find the effective action and its divergencies, one first integrate over {{formula:0dbae510-cba8-4dda-93a2-c151ce65c9d8}} (shifted to the complex contour), considered as a conventional renormalizable scalar field and afterwards over {{formula:8f43b542-ff43-41db-951c-38b628075630}} and matter fields.
The dilaton field is locally unobservable. It is fixed when we choose the global spacetime and coordinate system. If one applies this principle to a black hole spacetime, then the energy-momentum tensor of {{formula:25e5aa8e-ebe9-4bb5-8bf5-c005af876948}} influences the Hawking radiation.
When {{formula:d748103a-0454-4e48-953c-be22d306a0ef}} is flat, then the handling of the anomalies simplifies considerably. When {{formula:70900666-24a6-4985-b724-91ecf268670b}} is non-flat, the problems are more deep-seated{{cite:b23c0f6838bf37f9aeaf142614c353195ea5e5d7}}, {{cite:4d80f6fa0504b45b26667a054ed2e12610484847}}.
| i | c8b5f7bec5c760ac5522a64ba948e2d4 |
In this section, the Ref-Net is compared with the SOTA methods,
including unsupervised methods (CAC {{cite:1a3d16210b1586b55bad43c9ae0f64e9c6bb75f4}}, ReDO {{cite:f9b323a9123dc93c5be71a2a8aacc94dbc31c261}}),
few-shot methods (SG-One {{cite:5f7cd70023976d4d4c806e124884f54d71b69be7}}, PANet {{cite:2b29bcd4c1eada88e8cfdb0ba3049b9bcc43e565}}, SPNet {{cite:e62ab503cbf028c1c33ceaf73b3f5faac49b70ad}}, CANet {{cite:917361cff95f551ff5117fccc2e8eb27d4b022ee}}),
weakly-/semi-supervised methods (USSS {{cite:99e6575735e5bfa67ed7b9a042fb2da6bc03b58a}}, ALSSS {{cite:137a6d3689f7d745868d965e756697003742ec23}})
and fully supervised methods (boundary-aware methods:{ Gated-SCNN {{cite:0d53f26bedf0f23d7a8bc4f2a7851339eca99d5e}}, BFP {{cite:ae6a440ade9d52c90c3e2ac49b27fc2073de541a}}}, Unet {{cite:74aca1433e1adee551d8b29f5ed9356739439df3}}, FPN {{cite:6514199f59b2428cc0e98c2feef3bde8c014f640}}, LinkNet {{cite:442a5842f50596cba482417e8f79895f8a72ad65}}, PSPNet {{cite:8f7a0adba1dfde6c37b60df7a0c9e0fb5d9cd308}}, PAN {{cite:aaf591e7d7a07b5e14d1051e190b117d66c36dce}}
and DeeplabV3+ {{cite:13dc33f179a013ef9021686522a6ad8c544da37f}}) on six datasets.
For the semi-supervised methods, ten labeled samples are provided.
Except for the panoramic, the target dataset and open-source dataset have no overlapped object category.
For a fair comparison, the categories of each multiple category dataset (SBD and THUHR) are split into two non-overlapping parts.
The fully supervised methods are trained with both two parts.
The transfer learning based methods are initially trained on the half-category samples and then trained with specified labeled samples of the rest half-category.
Table REF shows the quantitative results,
where we can see that most scores of {{formula:249fb082-3901-4c39-812c-a9eaa41d5309}} achieve the state-of-the-art results on par with existing non-fully supervised methods.
Even with more labeled samples,
{{formula:8cd5c85f-bed0-4c47-bb9b-a5e0f57d5fbd}} only achieves higher scores than {{formula:9a553f3b-0ae1-499d-8d54-1089d26cb086}} on THUR and Bird dataset.
Moreover, with only 10 labeled samples, the Ref-Net can achieve better results than some fully supervised methods and close results on par with the best fully supervised method.
Meanwhile, with fully supervised samples, the Ref-Net achieves higher scores on the complex dataset (Cityscapes and SBD), which demonstrates the advantage of Ref-Net for handling datasets with more categories.
Note that the resolution of the Cityscapes dataset we adopted is {{formula:9c9e50d4-84d4-4c37-acbd-1c915df46249}} . The above two causes lead to the relatively low-scores of all the methods. However, it still validates the superior performance and wide application of the Ref-Net.
| m | 65e206a04be6a51dad030dff217fbcb1 |
Sections REF and Appendix REF highlight the predictive power of convolutional networks, which are capable of capturing spatial patterns in predictors, when they are applied to our data; we capitalise on their performance in our analyses in Section REF . A drawback of CNNs (and the recurrent neural networks we tried) is that they are only applicable to gridded data, which has been constructed as a sequence of images. Moreover, they can suffer from edge effects if the convolutional filters must pass outside of the spatial domain {{formula:01fa44d4-985c-4e27-bc81-4e358c8507e8}} ; recall that predictor values for locations that lie outside of {{formula:7f6d66b3-4c37-4ae0-b4fe-318046fcb868}} are set to zero, i.e., the mean of the predictor across all space-time locations. Hence, future studies may benefit from using alternatives to the CNN. For example, {{cite:134d2ec519e7210ab52d6398d70a616643b4a386}} propose graphical neural networks (GNNs), which are applicable to data with a graphical structure. Recent developments (see review by {{cite:fd738051a8d795effa1038e1079638f5e29275c9}}) of GNNs have seen the proposal of convolutional {{cite:34334351d7c7fbf97f420ddd27293d9a3c6291a6}} and recurrent (e.g., {{cite:6904735899a6a2b1a30b91717c14c16367e31331}}) extensions that can handle data which are irregularly-spaced throughout {{formula:a150ea04-d5df-4ae2-a074-59d02a2fb714}} and {{formula:3f11e7fd-917f-421e-b2d5-36772bf2955b}} , respectively; hence they could be used in our framework to model the extremes of processes observed at point locations or on irregularly-spaced grids.
| d | 2f3bf298e1720b90704cc16f20521266 |
To this end, we propose a novel pseudo supervision scheme, which
is leveraged to train the teacher-student network with distillation {{cite:8c9dbae092a16980a73610c39adae127df616856}}.
Specifically, the teacher network takes advantage of the effectiveness of unsupervised binocular depth estimation to produce accurate disparity maps.
The disparity maps are then used as the pseudo ground truth to train the student network for monocular depth estimation, which converts the problem of unsupervised learning to supervised learning.
This pseudo supervision mechanism enables us to exploit the benefits of both supervised learning and binocular processing for unsupervised monocular depth estimation. As a consequence, the aforementioned two weakness can be tackled to a certain extent.
| i | 07e06035de8fdc63c0f0ad96f97b36fb |
Deep learning models have been widely used in UDA. Earlier methods rely on minimizing the discrepancy between the source and target distributions by proposing different loss functions, such as Maximum Mean Discrepancy (MMD) {{cite:8ae8f8a87e0d35e1a9cc2a5dfa8241d65697e20b}}, CORrelation ALignment {{cite:99f3f448960aa59622963047de911a620e8e43c5}}, Kullback-Leibler divergence {{cite:e87fd737017d0f124682ebcb95094412898ca36a}}. To learn domain invariant features, adversarial domain adaptation methods aim to identify domain invariant features by playing a min-max game between domain discriminator and feature extractor {{cite:ff0491ee56af80f889cf5355d3993d2655f0d93e}}, {{cite:a4e61d0926c525404b6285d8e43aba6ee1c34b57}}. However, most of these UDA methods do not consider constructing separable domain-specific features ({{formula:2e010d79-3a12-4f2c-99f6-64ee4dc45dbc}} ) and domain-invariant features ({{formula:927aa124-a713-4f02-bc57-133be3dea2ea}} ) to learn more discriminative representations.
| i | 0d6263421b67009b116e980cb4202fef |
We compare to pre-trained LM based methods that leverage the output probabilities of the LM to make predictions given the sentence known to express the fact. Two methods are considered: (i) LAMA {{cite:000bba5ebb696eeed1f3661cf0ff4e1b94ee1f4e}} leverages the input sentence without the tail entity to query the LMs, and (ii) LAMA-Oracle {{cite:afffed550321b50ea56472501eab5f67e29a828d}} enriches the query with (at most) five gold sentences as additional context.
| m | 9ec7dfcbd332f8a7f575ed429a7f8a3f |
We found that the best-fit model for this source was an SED with a two-component blackbody excess, with temperatures {{formula:bf382f12-da7c-4362-b501-59440d3c6762}} K and {{formula:2906f656-b65f-458c-9653-da8638169e3b}} K and a total {{formula:5316915b-3dc1-4b21-a12e-3dc3beeec353}} and {{formula:8394d7a4-1032-49d3-813c-4b317906e35b}} .
Using equation 3 from {{cite:09c22f7fc0eaf63e9878411af5d11ba292ac8829}} and our best-fit temperatures and stellar luminosity ({{formula:4acbfc3e-bcfd-42c7-a0bf-79d469bffe77}} ), the blackbody radius of the outer disk for J0925 is {{formula:23513e22-3f1a-4f2f-b106-f2778562695c}} AU, while the blackbody radius for the inner disk is {{formula:43fc89ba-bce7-41bf-8a1e-ce86ea6e4af2}} AU. We also derive a surface area for the warm disk of {{formula:a70c3292-ab45-4012-b57b-27eaeedd9ec2}} AU{{formula:90882d9e-8678-4040-8d0c-a3e44d418f6e}} , and for the hot disk of {{formula:29c0a9b3-6f1a-4c96-b42a-b621cd69898b}} AU{{formula:9f341703-fcfc-4114-ad9c-76c6e7fbc32a}} . Assuming 10-micron silicate grains, the dust mass of the warm disk is {{formula:6a78e176-3a57-4db5-9d75-88b11209e716}} lunar masses, while the dust mass of the hot disk is {{formula:7af65682-0d6e-4a95-9388-c1d31d8bbc9d}} lunar masses. This ignores larger bodies (e.g. comets) that could contain the bulk of the mass.
| d | 532b143e81123a883ccf6c584a507f1d |
Our approach bridges the limitations of both PINNs (cannot be used when the structure of the DE is not fully known) {{cite:534d07da1e6918389aff0451cb77eea4f225fd42}} and UDEs (not robust to noise and requires lots of data) {{cite:c297c9c3311c0144259bb1b34e83c13af32bf892}}. To address this, we replace the hard constraint of the UDE with that of PINN loss, which allows the approach to learn unknown components of the DE model from data. This approach is robust to noise and performs well in low-data regimes. Additionally, using the AI Feynman algorithm {{cite:ebe3e8949423c85e958c21958a3c3cacb1b71b5f}} yielded good results in identifying the underlying hidden terms of the DE model.
| i | c0e780193e0e5cac24af3a686bde263e |
Another nice aspect of our results is that the multipole expansion in
biadjoint theory also matches that in the gauge and gravity theories,
for the wide class of solutions we have considered. This adds a
powerful weight to the observations made in
refs. {{cite:720bd16cf9e3bbbd13022eeb40c28c6ccf41391a}}, {{cite:714c86d7f7afce40c4384f540acb14b21741c0f2}}, namely that the twistor
double copy allows us to understand the inverse zeroth copy from
biadjoint scalar theory to gauge theory. That is, we have seen
directly that the multipoles of vacuum type D gravity solutions and
their single copies are essentially inherited directly from a much
simpler scalar theory! It is interesting to ponder what other physical
quantities can be phrased in such an appealing manner.
| d | c9dd84d5fe8ef92ad45e41dfac5fea04 |
We have studied the role of nonlinear effects on tidally-excited gravity waves in the radiation zones of stars, primarily focussing on a new mechanism that could be important in stars possessing convective cores. Our work was partly motivated to study tides due to massive short-period hot Jupiters, which are observed to preferentially orbit stars with convective cores {{cite:839b66cf8c4b23c7cfcf117cfee64ea5ecf90c00}}. For these stars, the geometric focussing and consequent breaking (or nonlinear wave-wave interactions) of gravity waves in the stellar core (which we have revisited in § REF ), which can result in efficient tidal dissipation when this occurs {{cite:20e07c74387625898b1b3fa469e4fcf074e0b84c}}, {{cite:0daa2755a055c311f4b85a25ae64e818fe4ac443}}, {{cite:7bcc3493cb664d08477ab1b3a47c1ace936a5fba}}, {{cite:6ae793695fb81a2efc71c42d044ffde4a95acb72}}, {{cite:24c115c478057209bdd1558ccc44e553d1315d39}}, {{cite:b085a376577e25620e4bb4bd051bee1a1d93161d}}, {{cite:ecb78729a52fca25034ffca79a69ca3d235a909a}}, {{cite:915637e3957f5428ba4ac0fdee9cddee8b346a93}}, cannot take place for planetary mass companions, unlike in solar-type with radiative cores.
| d | 0a29f56f4f5262bceb1fb2f6bc005f95 |
Table REF shows the mean and standard deviation over
different training partitions for two simple baselines, ESZSL{{cite:b696d6f47c66d1c2ed4c25ce237f8dfcaf6d512b}} and
SJE{{cite:0243d37711db3bf602779a4b6d3071a4b0687882}}, on two fine-grained (SUN{{cite:131022583240d6ac619bfe2ec5124422cd3cde14}} and CUB{{cite:2012d764b123bba3901da18d77a3afbe16e17acf}}) and
two coarse-grained (AWA1{{cite:071b36dd465b17afa05ded91e8bd587036f5eb71}} and AWA2{{cite:10e44f6f1de2d76b178cc9d5a8c31d1cdf1cb4d0}})
datasets, using different class partitions sampled at random. We observe a
great deal of variability whether the sample per class imbalance is considered
(avg. acc.) or not (avg. per-class acc.). We observe that the difference in
performance (as reported in the literature) might bias the selection of one
method over the other even when their difference is not statistically
significant. The table also show p-values of a Wilcoxon signed-rank test
computed from 22 different partitions chosen at random. We see that while for
the fine-grained cases we can reject the null hypothesis for a fairly low
confidence level, this is not the case in the coarse-grained data regime.
Although the difference in mean values seems high (for the standards observed
in the literature), the variability observed in the experiments warns against
choosing one method over the other.
| d | bfca87eb30c29dcd9c8c579fe4f806a3 |
Initialization: {{formula:50e4652f-9d01-45cc-92dd-d5c1529f87b8}}
{{formula:6ca67212-fb29-42b2-8b2c-6f123122b5a3}}
{{formula:0ac89c28-4c54-4c1e-b9f6-d4b34e723051}}
Generate {{formula:2c89dbd6-8d10-4697-9236-a8502dba11af}} training data samples at random
Initialize weights {{formula:0a739f15-d05c-45e4-a9e8-3da2c39db23b}} , and {{formula:19df6cfd-2a02-4c54-8d05-59737048c64e}} according to the Xavier initialization method {{cite:887f21158c3e52eb40ebe78ce104cc59a08d28f1}}
epoch = {{formula:113a2fb4-0d06-45b4-8d0b-9cc45495fc4e}}
index = {{formula:d79f24aa-9ced-4783-a012-87b4d0f03bf0}}
Take {{formula:c11d15f5-9d6e-473d-ba2f-58d033669304}} one-hot vectors as training samples
Caculate transmitted signals: {{formula:525008b6-81e4-4bcc-b344-a45cc9c14831}}
Take {{formula:9ca0a8d3-fd9a-4053-bd17-aef36844e500}} channel samples:
{{formula:ed1da949-4204-44f7-8485-9c95447cbef2}}
Get real received signals: {{formula:a185d6ae-c8c0-49d1-9871-e095da79c5d5}}
Generate fake received signals: {{formula:8a85eef8-5d5a-4f4f-80fc-7b93432d74c3}}
{{formula:cee49f9e-9a50-4fe0-9e99-44417f11ef98}}
Caculate the loss function {{formula:49da1ee0-a0d5-4de9-b5c1-1a9d2d0ff354}} according to (REF )
Use {{formula:734ce6c5-e332-491c-ba26-2139eba3fdf6}} to update {{formula:5e54b539-6145-4612-8a0b-9c0aee170896}} according to the Adam method {{cite:b690a8949d016f5fb570915c1aca44f799177aa1}}
Return {{formula:e5658916-bf9f-4a17-b764-2fa2c575240d}} and {{formula:e1db0e38-95dc-4f8f-bfd8-26447f7896e9}} .
RA-GAN based E2E training scheme.
1em
| m | 6f123fbe44fbfc9fb818956a9dc18366 |
Hybrid vs. Conventional Attention. The hybrid attention module (HAM) in the proposed DA-DETR plays an important role in achieving effective cross-domain alignment, where the visualization of attention generated by HAM is illustrated in Fig.REF . To demonstrate how HAM helps to mitigate cross-domain gaps in an effective and unique way, we compare HAM with most popular attention methods by applying them to domain adaptation, where the experiments include direct alignment {{cite:3af2f7fc981bbb34dd7f117e3b4a83cc8bbb843c}} (without any attention operation), direct alignment with conventional spatial attention {{cite:8af0400aaed479b11ae5d6978b39f4ab569495d5}}, direct alignment with conventional channel attention {{cite:80273ad63b1bb9ce12a34e992cd00ec61b6a0659}} and direct alignment with conventional spatial and channel attention {{cite:78bcc0d9fc163203eba939681349177efc375ace}}. The feature alignment is performed with two types of features including latent features as produced by the transformer encoder and normal features as produced by the backbone feature extractor.
| d | 6d6da2861d18b2d29d87fad84d9fff6e |
where the intrinsic H{{formula:772afa82-72f2-4773-89d2-6ce6f4039383}} flux is in units of {{formula:915a03d1-4321-4f41-8ad9-da0f0c045ddd}} erg cm{{formula:62dfba8a-50c7-484c-b9c9-e6941c265501}} s{{formula:4dffbef8-6fd8-4ae8-b0cb-78681decbc25}} and {{formula:bd40ed91-d263-42b4-808e-f3eac35f5fb7}} in kpc.
The intrinsic H{{formula:04be4b58-5396-4f22-99a3-af1964e041c1}} flux of the jet is derived to be {{formula:f027e178-d00d-4027-8b49-ce5b304cb887}} erg cm{{formula:d67e1ddd-4b65-4e25-88a4-a1bb089e60c0}} s{{formula:6ab25bf2-71c0-43cb-b363-0c61b0a5d8b1}} from the observed H{{formula:9f6095c5-6402-4a8f-ac85-60aec8a9a4c4}} flux of {{formula:a159a8a6-1b23-40bf-ac84-18b199d8e17a}} erg cm{{formula:a21dc77a-3c51-4e5b-980b-82ce1bd5a6e6}} s{{formula:ce2c4586-2289-474d-8ab9-c0bffbddeb4c}} for a logarithmic extinction {{formula:e024d24b-4554-4212-8187-ee8789d33a91}} (H{{formula:0016d40d-0caa-4ec2-b30e-c559cea76857}} ) of 2.0 {{cite:34ef8d8e40d9f691c0f41403bf2add9e9c3c806b}} and case B recombination.
The density of the jet is {{formula:bff65158-8171-4a2b-b8f0-f94279d0086e}} 2,500 cm{{formula:988f9ab9-199d-48db-ba79-d44ae7f3703b}} , as derived from the density-sensitive [S ii] doublet (Fig. REF -right).
Assuming a value of 10,000 K for {{formula:84e681ac-6e31-452d-91da-bd95842b4fe7}} , an ionized mass of 0.011 M{{formula:6e4a28cc-c7a8-4e95-932a-a574ebd7b8da}} is derived.
Alternatively, the volume occupied by the jet and its density can be used to estimate the ionized mass assuming a conservative value of 0.1 for the “filling factor” {{formula:31f61697-a386-4937-920d-a1addcaf234e}} and a standard value of 1.4 for the mean molecular weight.
Adopting a single cylindrical geometry with radius 0.5{{formula:dabf577c-c5a3-4d31-afc3-3d0fe45f601e}} and height 5{{formula:7e005ffc-caf7-44d1-a240-ad08a6f49329}} tilted by {{formula:f61a6db4-ddd6-456a-bd3f-8006e58d5b36}} with the line of sight, a value of {{formula:5f092bdb-d936-4645-8176-ee315ff517a9}} 10{{formula:a1bc0b3f-7944-47d8-9b36-281fbfa00b5d}} M{{formula:4a7d1bd2-6bff-4d39-968e-f3103df6daaf}} is derived.
Given the uncertainties and assumptions of both methods, we shall adopt an intermediate value for the ionized mass of {{formula:e3a04d95-d13d-4ffa-b787-d6736ce4ae60}} 10{{formula:fee74026-abbc-4ae2-885c-099598b97c4c}} M{{formula:b91f71a2-05c6-4285-9038-f0645de12764}} .
| d | c33dc93a8f34d4b4c2a76652859b3573 |
Our proposed featurization for persistence modules is rather naive, but well
adapted to the use of lattice convolutions as a data processing method. The
lattice convolutional neural network shows promise as a method for classifying
features arising from a multiparameter persistence module. The algebraic
perspective on partially ordered sets, exemplified by lattices, may also offer
approaches to featurizing more complex invariants of persistence modules. In
particular, the incidence algebra may offer a natural way to represent the rank
invariant {{cite:825d786dd7fc645e5349450fa093fbbf0d1f72fe}} in a way amenable to convolution-like
operations.
We hope with these brief experiments to inspire further work on featurizing
multidimensional persistence for use in machine learning algorithms.
{{figure:bea38c0d-19bd-459a-baa1-ceb6380a6466}}{{figure:afeda8d3-4c41-4938-8965-a0658be9e466}} | d | 0834989b0073c945b0bb52ee3bcd21a6 |
Value-based methods learn a value function to estimate {{formula:fe74db94-8738-423d-8cea-927366b7d6b1}} given the current state, {{formula:bf7b2852-f997-4163-8b1a-ffb7a84b46af}} (state-value) {{cite:1393f70aa1ef9bd8a51ce8c22725a44512f90f60}}, or the current state and a potential action {{formula:7b588d71-c883-4415-b1c5-ad74aef12120}} (action-value) {{cite:2cb5ce172fd81c0682204fc087f57a01758ecce6}}. Value functions can be computed by neural networks that take the state representation {{formula:518db3ef-9210-4d5a-b795-0f094da2363d}} as inputs and output the state-value {{formula:4ba9d052-b8ce-4ab9-8030-18199c8a9672}} or the values of all potential actions {{formula:25b63836-5726-435b-a9f7-0fdc21df3e2a}} , assuming a discrete set of possible actions.
| m | 964828b35147f845ab4786a45cf127ed |
Cleaning up the training datasets is a necessary thing for the larger models to learn good generalization power instead of the labeling errors and clean validation sets can verify this power. However, the hidden benefits of mislabeled samples cannot be overlooked. Although these samples have wrong labels, they contribute to the feature learning in the deep learning models. When 10 % of labeling error candidates were removed from the ImageNet training set in {{cite:4b2771369208cd722a78250322501644fb932133}}, a degradation in the validation performance was observed. The author found similar effects during the initial experimentation when 5-15 % wrong sample candidates were removed from the training set, the result was 0.5-1.5 % accuracy loss on the validation set. These results suggested for the author that the labeling fixes and the sample removal are necessary for a better dataset, but these methods must be balanced. Too many fixes can change false positive candidates to a wrong label more likely while too many samples cannot be removed from the training set to avoid negative effects of losing generalization power.
| d | 33fba247220b15b573b182351971a13e |
Another branch of explanation methods for CNN models is gradient-based. Those methods can also
be used to explain ViT models. As baselines to be compared with ViT-CX, we choose three popular methods, namely Grad-CAM {{cite:2657aad718d4f222ed8c09b44231b56dc2c8ce16}}, Integrated-Grad {{cite:f38e7b9ef3fbbfa91272177aabde17b1a517d657}} and Smooth-Grad {{cite:d1c992fe45f4136b15e10d5054043e2fc36e4d55}}. Gradient-based methods are not causal and have known drawbacks. For instance, they can suffer from gradient saturation, which can lead to poor explanation results {{cite:f38e7b9ef3fbbfa91272177aabde17b1a517d657}}, {{cite:14e0a27285ee86ca365b48c98052e2bbc59f946c}}.
| m | f107a3728cb7cf1a3d3c4f87a94bfc31 |
paragraph40em-1emA note on augmentations.
In this paper we have solely used the hyperparameters that come with the methods, and as such, have been optimised by the computer vision community for the last few years on IN-1k.
Augmentations play an important role for self-supervised learning {{cite:7a205240ccec8ee632ec14b4297768f7efb73d94}}, {{cite:eb77c715d22e1dc867123da5399555febccf12b7}}, {{cite:ab84d5c97cc9c2eb227c2322adeedf40f42d830d}}, and strongly influence the final performance.
We cannot replicate several years of augmentation tuning for our PASS dataset but can expect that with time, better settings can be found.
To support this claim, we show one example of such adaption in tab:aug that we were able to find.
By changing the minimum-size parameter used in random-resized crops from {{formula:519b4a43-bfca-4213-be75-38a501daf6a9}} to {{formula:f5584c26-c65c-4adf-add9-84e499b7fbfc}} , we observe an improvement in performance in almost every benchmark.
While small, the improvement might be explained by the fact that our dataset is less object-centric than ImageNet, and so stronger crops can be used, as there is no need for the crop to cover an object.
| r | 8676fd6812f96ebac5a492a4e15b8e9c |
Finally, as noted in section vision transformers (ViTs) have emerged as promising architectures for visual representation learning. Figure REF compares recent ViT-based methods against ReLICv2 using a variety of larger ResNet architectures. Notably, ReLICv2 outperforms DINO {{cite:960272a3c012a1f77677d260481a7ff860f56c87}} and MoCo v3 {{cite:b1d4d399d6a389d3b5b8e90de65175d092df839f}} and exhibits similar performance to EsViT {{cite:8db21b1602923e527c56301baba656d4faf6daa4}} for comparable parameter counts despite these methods using more powerful architectures and more involved training procedures.
Our results suggest that combining the insights we have developed with ReLICv2 alongside recent architectural innovations could lead to further improvements in representation learning and more powerful foundation models.
| d | 994faf5d42f8a268419ffe9f5d0cee47 |
A systematic comparison of fusion excitation functions have been carried out for
systems involving the target nucleus {{formula:151d3d63-348d-463a-8922-df41b2acd490}} Ho and different projectiles and is
shown in Fig. 8. The systems involving {{formula:4f38c641-a501-463d-9fa6-0bf2c0c29b5b}} Ho target nucleus, for which sub-barrier
fusion cross sections have been reported in literature are {{formula:f9d9ce73-9bbc-422e-88ae-bb88447b1d9b}} F+{{formula:21f02aa3-6b53-405d-9655-82f70938c441}} Ho {{cite:cd509ee457ef82756c26478b5db2ba99f6c17c1e}} and
{{formula:5ef876c8-59c4-4650-97ca-b46d0b5b3fab}} Li+{{formula:45233870-3ba0-4a7d-a01e-f243d7dc5c9c}} Ho {{cite:c4a88efa8fe7707cd2d58923811b8befa4010783}}. The projectile {{formula:6de7b028-8b6f-4a4a-a858-2be3e8a01781}} O in the present system is a stiff
nucleus, while {{formula:4901efc1-8d19-4ee2-961d-1c28c3975198}} F is relatively heavier and less bound than {{formula:cc3c9505-1a16-4c9d-860a-1e93231ff9f9}} O, and {{formula:8eddc5e4-0900-4c95-acda-b16f72abfa34}} Li is a
well known weakly bound stable nucleus. For comparison of different projectile-target systems,
the fusion excitation functions have been plotted in a reduced scale. Barrier radii {{formula:7640e8e2-8dc8-47bd-bf79-7ab0e137bc7d}} ={{formula:d02c4c2d-7886-41fd-8939-b598c5bb263b}} and Coulomb barrier {{formula:aa5887e2-9857-48df-81ab-7bc5ad1739b7}} for each system was obtained by performing 1DBPM model calculation, using the Akyüz-Winther parametrization of Woods-Saxon potential {{cite:0bdb8f185c087958d32433760b8155ee41c5a720}}.
{{table:7b2b3283-837d-4495-86b9-0507b1a6eb99}} | d | e99d55b98cb6fd98d95b8ab558f95e87 |
TransUNet {{cite:f95d4201e4c82feeffd24b3ada0bba524e3fc7e4}} incorporated the CNN-Transformer network. The CNN layers capture spatial information, and Transformer responsible for global feature. It has u-shape, where extracted self-attention features maps upsampled to be merged with varying CNN features skipped from the encoder. The implementation is available at https://github.com/Beckschen/TransUNet.
| m | 4a9344853971ad6a380d01ee68f80452 |
We will derive integrals as indicated in the abstract in terms of special functions. Some special cases of these integrals have been reported in Gradshteyn and Ryzhik {{cite:e5a59a0532fcc7f3121662f58939b770f04af1fc}}. In 1867 David Bierens de Haan {{cite:e236f1746b8ecedb73878258fd5d85f74cc56310}} derived hyperbolic integrals of the form
{{formula:515200db-f161-42ff-89c2-2f56dc37dc24}}
| i | ccbcf6d8e6f634f1eae0400e40d46224 |
The NPR is a significant diagnostic tool to characterize the localization transition. The localized (extended) phases are characterized by {{formula:8ff0f679-3f7b-4efc-bf1c-de7fb6ded6e0}} ({{formula:1553b136-75a9-4711-a2e8-40502459c6b3}} ) in the large {{formula:755ca8aa-4402-4633-b74b-daceb954331f}} limit {{cite:2a6a5b8b802c39910056de1eb29d022df8c15a08}}, {{cite:ed0a28614fa1b3b53ab939ab1c584799327f4cc3}}, {{cite:4044dd8e04fc34d12a661d3ae6412685f6947299}}.
| m | d438ef4c860beba9b652aa0e5f879a69 |
This work studies to what extent voice can hint face geometry motivated by recent studies on voice-face matching and cross-modal learning {{cite:595ef0793a6cc897873d03ce39e499b83ace8ac0}}, {{cite:60f2505e8b7aa53a0fb2baee2bef419ec679a7ff}}, {{cite:3f24578ddd74fcdfba5d39b3effeb8b9053dc911}}.
Many physiological attributes are embedded in voices. For example, speech is produced by articulatory structures, such as vocal folds, facial muscles, and facial skeletons, which are all densely connected.
Such a fact intuitively indicates potential correlations between voices and face shapes {{cite:f10c9f26217fac4635042781ae0fcdcb9d007324}}.
Experiments in cognitive science point out that audio cues are associated with visual cues in human perception– especially in recognizing a person's identity {{cite:8ef5ae04e7abbce3195efbe506a5fe171704a863}}.
Recent neuroscience research further shows that two parallel processing of low-level auditory and visual cues are integrated in the cortex, where voice processing affects facial structural analysis for the perception purpose {{cite:a399aba51450c1f69745f34c567fa42e3172b0eb}}.
| i | 4b8eba88c978f360ad747b075410df68 |
Our luminosity functions can be compared directly with the best fit
models in the same redshift ranges derived by {{cite:6e631beb328acfb7846c13fa98b38194280f160d}} and
{{cite:1a81cbbbc087976edb7fdc7c03e7df237c744e66}} in Fig. REF . As {{cite:6e631beb328acfb7846c13fa98b38194280f160d}} and
{{cite:1a81cbbbc087976edb7fdc7c03e7df237c744e66}} carried out their studies in different regions of the
sky to us, and to each other, differences in the normalisation, and to
a lesser extent {{formula:5c48f997-01da-4beb-a276-e614a9011e0c}} , between the three studies are expected
because of cosmic variance {{cite:16bd0acba69ee97bb7e7bfd1112ee938306a7dec}}, {{cite:468b74adac8e5acd8f1d142f13a8d3229cca79bf}}. Ideally, cosmic
variance would be overcome with the use of a statistical sample of
independent UV survey fields, but at present there are only three. The
study of {{cite:6e631beb328acfb7846c13fa98b38194280f160d}} is based on a larger area of sky than our
study, or that of {{cite:1a81cbbbc087976edb7fdc7c03e7df237c744e66}}, and so we might expect it to probe
the most representative range of large-scale-structure environment. A
simple estimate of the relative richness of our survey region compared
to that of {{cite:6e631beb328acfb7846c13fa98b38194280f160d}} can be obtained by comparing the models
around the faint limit of our survey, where we measure the largest
space density of galaxies. In the redshift range {{formula:2e1b2ce9-7bb4-468d-9c47-569426e0982e}} , at
{{formula:d68b19c0-5af6-4126-a673-f404b6a721ef}} , our model for log {{formula:bc512554-bdeb-4ff7-87b2-e9f8e277a2be}} is higher than that of
{{cite:6e631beb328acfb7846c13fa98b38194280f160d}} by 0.2. If we were to assume that this difference
represents an overdensity in the 13{{formula:5f5a01f8-895e-4bae-a805-14c7d73ecc10}} field due to cosmic
variance, comparison with Fig. 7 of {{cite:16bd0acba69ee97bb7e7bfd1112ee938306a7dec}} suggests that we
might expect our measurement of {{formula:a7dfaa0e-64d1-46c0-a4e5-7cb0531fa17b}} to be biased toward brighter
absolute magnitudes by about 0.2 mag. We consider a potential bias of
this size to be benign, because it is only half as large as the
1 {{formula:e7b88e6f-c536-49d9-b34c-4508bffa9efc}} statistical uncertainty on {{formula:6477f8e5-fda5-45a4-8875-645387b696a4}} . In the redshift range
{{formula:d642e796-69e1-4a7b-8a03-ff69f05e3275}} , at {{formula:6058e23b-0287-4d50-9120-e5bc61b53145}} our model for log {{formula:b161fdeb-f52e-44e5-af50-706dc14c1c5c}} differs from
that of {{cite:6e631beb328acfb7846c13fa98b38194280f160d}} by only 0.05, so we have no reason to expect
any significant bias in our determination of {{formula:f0456ff5-f585-4bef-9b32-0578a35cb8ec}} .
| d | 65013ac911c50a4f1e1d1ca2aa0162f7 |
For an RW starting from an initial node {{formula:83be59ce-f417-438d-aa0b-153b0f44e7e3}} , the
first return (FR) time {{formula:569b289d-858f-45c1-a2bc-47351e992a89}}
is the first time at which the RW
returns to {{formula:8a6a6a0c-6358-40e3-9519-b5880009b3a2}}
{{cite:3d7518f7fa2977bf9fbfb61d531fd4d8c0eec99f}}.
The first return time varies between different instances of the RW
trajectory and its properties can be captured by a suitable distribution.
The distribution of first return times may depend on the
specific realization of the random network and on the
choice of the initial node {{formula:88bda612-d2ba-4649-8ef7-7d96e309eb44}} .
The distribution of first return times for a given ensemble of
random networks is denoted by {{formula:55a5f377-d9c4-44d4-8b56-4853908e40e4}} .
This distribution is calculated by averaging over many
network instances drawn from the ensemble.
For each network instance one needs to sample many RW trajectories
starting from random initial nodes.
| i | 2f09b65291623634f2c693e753c5ac10 |
Besides, we compare the inference speed in terms of inference FPS in Tab. REF . Note that this metric has a bias. It is more favorable for methods that use a high input frame rate (, 25 in R-C3D {{cite:9b9db5b110cd087a59c67fdd95a8a7a10d98166c}}). Therefore we also report the speedup ratio, the ratio of inference FPS to the input frame rate. Our detector runs at 5076 FPS and has a speedup ratio of 508, which is much faster than the other end-to-end methods.
{{table:05c38dc1-99bc-4380-ba47-a931ec735add}} | m | facc83a9ec6a4c32ef0817c51f66d1c2 |
In three dimensions, we perform similar measurements to obtain a lower bound for the fidelity of our state with respect to {{formula:6261e888-1d15-4956-a70f-6ae57f579fa3}} . By measuring the correlations in the standard and DFT bases, we certify genuine three-dimensional entanglement by obtaining {{formula:3243ac56-36e2-40e1-b430-f540df60dc98}} (fig. REF d,e), which is above the upper bound of {{formula:0a764ab0-a4e6-41f6-95c8-3a56f536146c}} obtainable with two-dimensional entanglement{{cite:60f8eaf57d59899cf1dcd48c7b4cc4539764b732}}, {{cite:c6def48859d6b0b4249297feeaca8d1103999ab6}}. In three dimensions, two relative phases between the different terms in the quantum state, {{formula:ff6c5858-b86c-4153-87ea-f90a91708e25}} , can affect the quantum interference between the photons in the DFT basis. By varying the phases of two input spots and measuring the coincidence rate between different output modes we observe quantum interference with an average visibility {{formula:483960d3-6847-49a1-b6c0-0214d84220bc}} (fig. REF f-h), and with good agreement with the theoretical prediction (fig. REF i-k).
{{figure:c816a2ac-b6dd-4dec-9446-310be396875f}} | r | df0f6f4e20474edeb21ff333954ecfcb |
There are caveats to our study of SMM-Cs.
We have only demonstrated SMM-Cs on robotic domains with smooth dynamics, while BBNNs have been demonstrated as useful model parameterizations on simulated contact-rich domains {{cite:e8a8dfcdd037a4910c7372d944569382e1776447}}, {{cite:037383ab70418520643f8849dcc7b42e1ddcf2ea}}, {{cite:94d9a75794915ab114511160ef91d22948c5fa55}}.
However, SMMs can be used in contact-rich domains with the use of variational integrators {{cite:2e0c60a901153bae734389223ce90e37b1f45501}}.
To do so, kinematic constraints resulting in contact are explicitly satisfied when making state predictions by solving a constrained optimization problem, where contact and non-holonomic constraints are enforced using Lagrange multipliers.
Furthermore, constraints can be introduced and removed after the model is trained.
| d | dc6cb6e3d815fb54e11eed6a4fde7b79 |
Recently, orthogonal time frequency space (OTFS) modulation has been proposed in {{cite:ae5dfad3af053742c5da5f2bfca5dfdf94c28a28}}, showing significant advantages over OFDM in high-mobility environments. OTFS places information symbols in the delay-Doppler (DD) domain to capture the channel geometry that models mobile terminals and reflectors in a high mobility scene. Leveraging on this representation, the OTFS modulator multiplexes each information symbol over 2D orthogonal basis functions (IFFT along Doppler and FFT along delay), which span across the entire time–frequency domain required to transmit a frame. The set of basis functions is designed to combat the dynamics of the time-varying multipath channelA similar scheme to OTFS was independently proposed in {{cite:b27ce53b04b4ed6c06f48dfb20d963e0602a70d0}} for underwater acoustic communications..
Further, it was shown in {{cite:8e112e6a2dc6aa7e64e0a201097b370d42b473f4}} that any 2-D orthogonal transformation (precoding) with constant modulus basis functions operating on the time-frequency domain enables the receiver to exploit maximum time-frequency diversity. Since the Fourier basis are constant modulus, OTFS guarantees that the information symbols experience the same signal-to-noise ratio (SNR). In the recent few years, there has been a number of efforts dedicated to the development of OTFS (e.g. {{cite:98b254aa0d789e6ce6ad27568c9986e6c66a2d20}}, {{cite:b2e06d28429e0787c90e6aa5d58a99e14e6a2d6e}}, {{cite:5476346c651ab16a64dbae67e71d1b2da171a5ad}}, {{cite:99fd72461fc40f80e1af146ddb3a2c2cd891391c}}, {{cite:ef24e4773abe70040e3cff76d90725f77d5d706f}}, {{cite:8e112e6a2dc6aa7e64e0a201097b370d42b473f4}}, {{cite:edc003f519c63c639156a8d35de890736de26b24}}, {{cite:6a24765cb26a69eb94cc7448c4ce24a634f9d32e}}, {{cite:d5df7a3c9df941c0c072ed083c475aee5e8eb018}}, {{cite:0832f630617203cf7f96867d169010cdb4fd8aff}}, {{cite:51a8ba87cb7a67470714200ec98a4ef8d81d9bf3}}, {{cite:21256c6c5cc45d021544658fe60dcf316148caf1}}, {{cite:b9be23ef232a24c3ffbae655858ebbcd1a4a56ff}}, {{cite:cf5d083f54f2fbaa26c72203cf6a31af3e479fed}}, {{cite:46996e59017ed45f455d6cfbc8e8cc4f6fc9383f}}, {{cite:8a8e5a9a95954c5aff2481b5c4299b15147cddf3}}, {{cite:5abdec65057b1f6b16c2e8f03f3668dfa6fe0c7b}} and references therein).
| i | 18fa4c67fe23b7d02027d32ceeb61518 |
The results depend far more on specifications for covariance matrices, and how a large
covariance matrix is estimated from a relatively small ensemble of model runs.
We used likelihood and Bayesian approaches for estimation of covariance
parameters; a variety of alternative approaches exist in the literature
{{cite:169a402893e73a58952291c93832b35ecba06573}}, {{cite:8e1905360f45dfc00f6503709c748bf6b9cde713}}, {{cite:3561c888e7e9e68e33311394a44d6ae84f6083b8}}, {{cite:d20ece7caae2ad8868808fbb6a64cd31a780f095}}, {{cite:907e184850751842c4426ce71b56a94e16a4e835}}.
| d | 9bb3fdfa4c24f201476f786e0d49ff82 |
According to BIMA observations {{cite:cb61f2a9c91ce1238d8dc38929c9cd06f050d498}}, glycolaldehyde is greatly extended in comparison to the ethyl cyanide and dimethyl ether, which are largely confined to the Large Molecule Heimat source (LMH) {{cite:aec109eb89aacae8d4f0efb6eacdaed0b604a2ee}}, {{cite:24d0ae6b4d85c6589da6ddf07d6c1e5f647430ff}}, so these molecules were classified as either “extended" or “compact" here based on the half-power radius of the spatial distributions in Figures REF and REF .
Figure REF shows variation of the normalized integrated intensity of observed transitions with the distance to Sgr B2(N) for sampling points with {{formula:e850df8d-8261-46de-ae4a-5bef85acc0cb}} RA=0. The normalized intensity is obtained through dividing the integrated intensity by the maximum integrated intensity of molecule. We change the declination and take a measurement at constant RA ({{formula:cf5750ab-44d6-451e-9312-59e50cac8931}} RA=0). FWHM, the full width at half maximum, is always used to evaluate the line width of spectral line. Thus we selected 50% as a metric to evaluate the distribution of molecules. We could see from Figure REF that the normalized intensity of CH{{formula:20a15b5e-c9e2-4a65-bd7e-577adb0f0528}} OCH{{formula:a8d598e1-071d-4c66-8fa8-ba5334f485f4}} , C{{formula:101cc822-bd9d-4cf7-bb62-d15b88ae70c8}} H{{formula:9116a8db-171c-425f-93f7-a0760a5412b9}} CN and H{{formula:1f602aa5-bbc6-4423-a8ab-bf956d527bf7}} NCH{{formula:6acf23ea-bc50-4ce2-8ced-b16ee75223a0}} CN quickly decrease below 50% for sampling points away from Sgr B2(N). The half-power radius of these molecules are smaller than the sampling interval (1). Thus these three molecules are classified as “compact” molecules. H{{formula:c8cecd85-c7a9-4a0e-8b2b-081538fe8d7e}} NCH{{formula:c355e11c-681d-4553-a0e2-ff8a09160df0}} CN was only detected in Sgr B2(N) in our observations, thus it is also classified as “compact" molecules. On the other hand, the normalized intensity of CH{{formula:81c9a2f1-1a65-4eff-a070-fd163075a881}} OHCHO, CH{{formula:2b95526e-31c6-4930-bcf7-ff677947145f}} OCHO, t-HCOOH, C{{formula:0286bb81-feac-4b48-9b3a-9eedc28d353e}} H{{formula:81f826bc-d978-43fa-a464-0903e5e64c40}} OH and CH{{formula:1990abe2-e97f-4faf-946d-3d00871e97cf}} NH{{formula:8ca561e0-6ef6-45e0-901d-aac66f42dee0}} vary slowly as the distance to Sgr B2(N), thus these molecules are classified as “extended" molecules. In order to investigate whether only the strongest lines are extended, the “extended" line profiles were colored with red, while the “compact" line profiles were colored with blue in spectra of Sgr B2N (Figure 1 upper panel). As is shown in Figure 1 upper panel, the strongest emissions come from C{{formula:0f735179-e2c1-42f0-ba09-de308c82bdf6}} H{{formula:e556427d-4f35-4017-b264-43d4e18a919e}} CN, C{{formula:c475ac28-1f51-4026-8549-5845b54c856c}} H{{formula:4c9c6bda-9a4b-4742-b908-88da14d20fb3}} OH, t-HCOOH and CH{{formula:08bdb980-8546-4573-88ce-66fafe3bfb76}} OCH{{formula:d27860f2-c34b-416f-8516-85ae879d5025}} . Among them, C{{formula:9c094a86-7df5-4421-9d5e-dbfbedc457ab}} H{{formula:d7b77f35-5d14-49ba-b71f-0d4e3b80ed90}} CN and CH{{formula:698e0c2f-a739-44a0-b5d7-7c8483886963}} OCH{{formula:d86e517c-4dd9-4ea2-933e-e0c21e46b32d}} emission are confined to Sgr B2N and Sgr B2M. Thus it is sure that not all the strongest lines are the most extended. It is noted that CH{{formula:71018da6-9e57-4f0c-aadb-852a93bb12a0}} OHCHO peaks away from Sgr B2N, which seems to differ from most COMs presented in this paper. The line blending makes it difficult to obtain the intensity of CH{{formula:d1e77955-5ea2-4977-be52-30aef1371a74}} OHCHO accurately. We have mapped CH{{formula:815632c3-9cf4-43db-b36e-e30c0e9b5d99}} OHCHO around Sgr B2 with IRAM 30m telescope with better spatial and spectral resolution to further investigate this issue. The preliminary result is consistent with ARO 12m result present here. Detailed chemical model is needed to explain the spatial distribution of glycolaldehyde in this region.
{{figure:f7917fd3-4d4d-495b-93fa-18fd74c0a256}} | r | 6104a7eba4ac576a8d6309ad8a4ec44a |
Our quantum algorithm prepares a pure state {{formula:63c9bf1c-a585-494e-acff-1b9195f207d4}} , a purification of {{formula:ba48510f-fd26-4965-a708-835b965b4d5b}} , and {{formula:7ac6c02f-6b65-419e-aa8b-0666b12d6bb9}} is obtained by discarding or tracing out {{formula:12397b80-d60a-4bd4-b912-96b458aff812}} .
Working with purifications is advantageous for several reasons, including the efficient computation of thermal properties, where quantum-metrology techniques can be applied {{cite:4ee33ce3ff8f7ce355eb5ce08a4147ed24c50891}}.
The quantum algorithm implements certain unitaries {{formula:51005431-4e0a-4f14-8ce8-9fbd849d1adb}} , {{formula:c26b40a5-11cd-4ea1-9c3b-38f412c7d8ab}} , {{formula:da44cfab-9314-407c-b8ad-be1c24f843d8}} , their inverses, and the controlled versions thereof. We work in a model where the implementation details of these unitaries are not essential, so the main complexity results presented in Thm. REF can be regarded as the quantum query complexity of the quantum algorithm.
One could obtain the full gate complexity
from explicit constructions for those unitaries, but a detailed analysis of these constructions, which also depend on the presentation of the Hamiltonians, is outside the scope of this paper. Nevertheless, for completeness, in Appendix we give one possible construction for {{formula:9ca57b6c-f09d-4027-a2c7-81a540f80cbc}} using quantum signal processing and qubitization, when the Hamiltonians are presented as linear combination of unitaries {{cite:39cc69ded9ce90d5ec1e3dba59fa8564310e0dec}}, {{cite:b69e693f4ea9f091ce1109ec1effbb9070913d5a}}. Additionally, finding {{formula:da01e14c-6f5a-4e90-9243-6f36864b003f}} is
a difficult problem in general.
Our quantum algorithm is then expected to be mostly useful
in instances where {{formula:1859008e-f098-4c10-9b35-b6bf90f8c804}} can be accessed or efficiently
constructed, and we provide an example of this in Sec. REF , where {{formula:a5b31db1-3005-4d45-8330-bf26320de6d8}} is a non-interacting spin system.
| r | 581fa14fcbfa5dabd7582e3905f3b18a |
Our work focuses on policy improvement methods, an algorithmic framework that was first introduced by {{cite:cf998e383dab879908746479dedb5d7a50f8206c}} in Conservative Policy Iteration (CPI). CPI guarantees policy improvement by considering a mixture between the current and greedy policies at every update. The policy improvement bounds proposed by {{cite:cf998e383dab879908746479dedb5d7a50f8206c}} were later refined by {{cite:cae9db375d1ead1a0937b6ce1d2356fc45d9a262}} and {{cite:d920131a06d21eb4ea61c28ba116b9d7a9066547}}, making them compatible with the deep reinforcement learning setting. These advances motivated the development of many popular on-policy policy improvement methods, including TRPO {{cite:cae9db375d1ead1a0937b6ce1d2356fc45d9a262}} and PPO {{cite:9268f9e396a36d306cc5430143492162523717ba}}. The strong empirical performance of TRPO and PPO has been studied in detail {{cite:45bdd7b0f8fc4c5f9bdcb0a96d84630bc567d1d9}}, {{cite:98af9f4f9a28360956522f4626529427c98b66d8}}, {{cite:d0bb7d774584d69e2b69e8a08a4232cea1cb3ba3}}. In addition, these algorithms have served as the foundation for many on-policy extensions. {{cite:d920131a06d21eb4ea61c28ba116b9d7a9066547}} extended TRPO to the constrained setting, {{cite:e52318c696ef6b706e80ee00887d77f30655b923}} proposed a robust version of TRPO from limited data, and {{cite:8a69325c0b0ed61868b3d7f7f7026ba4a1992568}}, {{cite:ee1989db3b98c5697763f2351b2e33334c6bacf2}} considered modifications to the clipping mechanism in PPO. Given the popularity of TRPO and PPO, we consider both of these algorithms when developing our family of GPI algorithms.
| m | 1c4daeeaac2057b91f828045b8c1800e |
Creating large-scale medical image datasets for training neural networks is a major obstacle due to the complexity of data acquisition, expensive annotations, and privacy concerns {{cite:f0f7d8eb49f7c02f7a0e93042def8256cd79dd87}}, {{cite:2387ce538bc8a2acb72cbfdc6d40107b848fd20b}}.
To alleviate these challenges, a conventional approach is to train deep networks, e.g., ResNet-50 {{cite:3e51db3d1ae8b8b72ef582d4af7c2b889973e790}}, on large-scale natural image datasets such as ImageNet {{cite:d0c5726a5fe74ee4e59e4543a44c143c4a88c2ab}} and subsequently fine-tune them on the target medical domain. However, such schemes are sub-optimal due to the large domain discrepancy between natural images and medical data {{cite:41711d956b9720a3f514027ca4945bf571a58b6f}}, {{cite:f63913aeaad92ea66ac82a9401ceccdb236c5363}}.
This has motivated other techniques for collecting annotated medical datasets across domains and training networks using full {{cite:77c8d0e8601d47319540072e0bdd58fe7e32cb33}}, {{cite:ed1f097dbb61b57e3aea5f5944c8bd2550e215ad}} or semi-supervision {{cite:72aa26f8c10385f796d1db9c139370aabda9c9f3}}.
Nevertheless, the amount of acquired relevant training data in this manner is still limited, which significantly limits the performance of deep neural networks.
{{figure:2a12a87a-384c-4935-bd1a-a4291ee01bbe}} | i | 4628bf6a8d78f71f4b5307b558d3bd32 |
Inspired by Lin and Och {{cite:a9bef9c9f8a9ec911557c086a8860ace8ac7322a}}, we proposed an LCS based post-editing algorithmExecutable source code is available at
https://github.com/SourcecodeSharing/CWSpostediting
(seen in Algorithm REF ) to alleviate the negative impact to CWS.
In the algorithm, we define an extended word segmentation labels set {{formula:e6d22dc5-2462-4920-8f6b-597d6134d81d}} .
{{formula:f68b0075-1f9d-460b-961d-1d697eb60529}} represent begin, middle, end of a multi-character segmentation respectively,
and {{formula:83e07311-be28-48f4-b5f7-2542cb4c65d8}} represents a single character segmentation.
The additional label {{formula:d0bdbec2-e2a2-49e9-9d8c-da4ba3b71c0c}} in {{formula:8e294304-030c-47a0-966a-d1907e86846c}} can be seen as any other labels according to its context.
For example, given a CWS label sequence {{formula:8e4360d4-cd15-4dc9-9c0c-7c1227c161ff}} , the {{formula:9ce6248e-dcde-483e-a865-85303de2cd5f}} should be transformed into label {{formula:b4de6d78-439e-4b4f-9f41-74ba7e2a47a2}} and
in the other case of {{formula:46be6722-fbfa-49c5-9864-1252ea09feb2}} , the {{formula:be12c5c2-e3e6-4002-9869-5bf45609000e}} should be treated as label {{formula:7d834095-13dc-4c77-ada0-a7c92b36ca7d}} .
The above transformation strategy can be based on handcraft rules or machine learning methods.
In this paper, we use the transformation rules written manually.
Table REF also gives an example of post-editing results.
| m | 333b7be288bc1599077ba2915c11502a |
To evaluate the performance of the proposed method, we examine the rooted mean squared error (RMSE) of the changepoint estimate, the empirical coverage of the credible interval (CI), and the length of CI. For each setting of spatial correlation and SNR, we run 100 simulations. Different locations, changepoints and functional data are generated independently in each simulation. For our Bayesian model, to make sure the MCMC chain has already converged, we try several sets of different initial values for all parameters and evaluate the difference between those chains with the Gelman–Rubin diagnostic {{cite:77aed37130206e1f399745958ad17b23d843f304}}. We also apply Geweke’s diagnostic {{cite:4cc57f2181b402b18a820449bbfff12481014e2c}} to determine the burn‐in period.
Through experimentation, we find that 20,000 MCMC iterations with a 15,000 burn-in period and thinning with step size 10, is sufficient to produce nearly iid samples from the posterior distribution. We compute 95% credible intervals as the interval between the 2.5 and 97.5 percentiles of the posteriors for each parameter.
| r | b0bfad3295920328939558a53cfa451d |
In the future, besides improvements on the sub-PeV Galactic diffuse emission modeling and
measurements, ALPs – and more in general exotic physics – searches will
benefit from observations at higher latitude, where also the Galactic (diffuse and source) emission is suppressed.
Interestingly, Tibet AS{{formula:0463314a-0adf-42cf-aeaa-3dc4cbe4a377}} has sensitivity also to high-latitude ({{formula:a061f71a-d598-4c38-b8aa-c77cc628699b}} ) photons, and,
recently, Ref. {{cite:d18d44bfe5621f466b7a62cadd2d069e66d4451a}} has used the full charged cosmic-ray Tibet AS{{formula:155cf3ef-e8dd-45e0-8011-91ba4d70bb62}} measurements to set an
upper limit on the diffuse gamma-ray emission for {{formula:39af38e5-b7ed-4b41-9a8c-408479d51dfc}} .
By using these data as reported in {{cite:d18d44bfe5621f466b7a62cadd2d069e66d4451a}}, we
get, however, a bound which is a factor of about 3 weaker than our constraint in Eq. REF , i.e. using the scenario of maximal IE in the sub-PeV energy range.
Nonetheless, future observations by LHAASO are expected to set world-leading constraints
from high-latitude observations {{cite:0543a87c1045a7b3a5ca54e83705970b337b33d6}}.
{{figure:263b94e2-9ad7-4b2d-8a7e-04a220543a88}} | d | 715c547bf32817d868e32b5de6dfb47a |
If the dimensionalities of input and output vectors do not match, the vectors are padded with zeros during inference, or small scale Gaussian noise during training, to encourage the network to ignore the additional padding dimensions, as done in {{cite:c47ffb681bc1deeb776c1da9754db0a5a7d556a8}}.
| m | 0246b799455845fe19b92db1f6e1cf7d |
Observe that if {{formula:0f819c1c-4bca-4153-8f76-26a8adfe906e}} is a finite Galois extension of {{formula:7e8b3996-6d82-434d-bfda-9be706348c8c}} containing {{formula:3290cbae-fcd6-4cb2-9982-faa5c89db320}} , it then follows from the transitivity of the norm maps that the surjectivity of {{formula:d2502574-abe4-46ab-867a-7d66edf1d030}} is a consequence of the surjectivity of {{formula:187bcb07-2124-42b7-983a-50e6174976ae}} . Hence we might as well assume that {{formula:e53ab13c-47f0-4a8e-9317-0afdb5028eeb}} is a finite Galois extension of {{formula:d1ca1d62-8da7-4332-97e9-743e7bb6fa96}} at the start, and to lighten notation, we shall write {{formula:8b5bc8c2-10ce-4c42-91bb-2c3078729b82}} . By the Universal Coefficient Theorem {{cite:77a313ef05bcb842ad3029c5ba8b9d3d6c95bf41}}, there is a short exact sequence
{{formula:14d958df-25ee-4618-b127-552fcb2839a5}}
| r | e6fbfda598417d46d7d4faa7f15b1040 |
The bare cross section including final-state radiation takes the form {{cite:d2c367e67353bbd72242ac89e25dba37c26c9b36}}, {{cite:3772081d62d6dbf66dfcb9361a0513b38df207d2}}, {{cite:a76a796fa666b2df524320d8e6886ac4af46b28e}}, {{cite:1048c45561df7af56c359d834b691e15041484db}}
{{formula:823d4634-9ffe-4f58-a573-98e86c1434e8}}
| d | 94cd6a6d059dffc93561923bf45292bc |
In order to verify the effectiveness of the proposed MDRN, we compare it with more than 10 classic SID methods, including BM3D {{cite:981e1536c1ae9c28836746bc99607ec6a1505c22}}, RED30 {{cite:597e78e49738831dfa449f330f6a9a8f0ab92400}}, TNRD {{cite:d28a8116d50c2ca1bf8249636160d3aed8b91ec4}}, IRCNN {{cite:79d9eb8f5248e23bb0ccda731a7608bd5c9dd56f}}, DnCNN {{cite:98cc679f94ede7a04e841f485577dd1d7d05af31}}, MemNet {{cite:396c21ada8cad833dd5eab5bfc2476e6f8e67217}}, FFDNet {{cite:5c6e66a76bf386c4aedf49a7215a299edeb7a24c}}, MLEFGN {{cite:1bf10bdffab1e704995ea783a49ad2cae33633b2}}, MFENANN {{cite:a10c7ced50d71711e56da05c96835064fc9bc337}}, and DRNe {{cite:072849f16d38591a501dc6bb4b7c17e91e40f8f4}}.
All aforementioned methods are the most widely used denoiser and all of them achieved the SOTA results at the time.
{{table:9ab952ba-8bec-4a63-ae0e-fe7b34bbd9ab}}{{figure:742d4df0-fc1e-45d0-b819-75d671ba6fb3}} | m | cb8ac68de66b0716044619804b2d683d |
Evolutionary governing equations via VO-RL derivatives of constants: A particularly interesting property of fractional-order Riemann-Liouville operators stems out of their behavior when applied to the fixed-order derivative of a constant. It is found that this fractional order derivative is not equal to zero, unless the order converges to an integer. While this is an unexpected and maybe even unsettling property of such operators, at least in view of classical integer order calculus, we will show that this property has extremely valuable implications for modeling physical systems exhibiting highly nonlinear and discontinues behavior. Mathematically, the RL derivative of a constant {{formula:149b7670-817f-462a-9e3f-7b0bea77a2bf}} {{formula:fee82caf-b1cf-4d3d-b6b1-257c49c85ae6}} to a constant fractional-order {{formula:fbee9234-a0c5-4ec8-b232-a94cd1e5753c}} {{formula:4eab59e5-2a3d-4ed2-a125-7836938a9540}} defined on the interval {{formula:c6e22fe0-fbf0-46a4-aaac-fee6fb9ec92c}} is given as {{cite:2bd08d605511c0b9b02874a8d27603cd138edd69}}:
{{formula:0dd7f636-f98d-4d06-b6dd-52d4c13d155f}}
| m | 16254a78f0b841086c973575db8221de |
Results for Normal to Foggy Weather. Differences in weather conditions can significantly affect visual data. In many applications (i.e., autonomous driving), the object detector needs to perform well in all conditions {{cite:93f27e8df888ad2e60b931e40859be7e89c38880}}. Here we evaluate the effectiveness of our SNR and demonstrate its generalization superiority over the current state-of-the-art for this task. We use Cityscapes dataset as the source domain and Foggy Cityscapes as the target domain (denoted by “Cityscapes {{formula:2dbb7d86-56e4-405e-8186-9edf27e30362}} Foggy Cityscapes”).
| r | 0fb9d7a7eb4ca059df9bf7111a7fd801 |
Let {{formula:ea3a9478-4c1d-4121-a048-26d35d4a2b53}} be the Banach space consisting of continuous functions from {{formula:3230c9af-4895-4ac7-8cbc-ce77e070b86a}} to {{formula:94fa31e5-b701-4480-8401-2f43944f87b9}} with the maximum norm. The following proposition is due to Henry {{cite:d47d704da2619ba4916ec3f3c952dbf0e2af10e8}} (see also Cruz and Ševčovič {{cite:e74adebb5126e620bd1a3b9d07b9748cb842cfca}}).
| r | 4158f7a7c0004c8a6b15fad42126b699 |
Our model of Indirect Active Learning (Figure REF ) assumes that the relationship between {{formula:781af1fe-256d-485c-816e-01cd9e79f673}} and {{formula:6de12eaf-ad49-4f92-be4a-e636d2f3efdb}} is unconfounded given {{formula:ba97445f-7307-4150-8fbf-f69171a9b6a8}} , i.e., that {{formula:aacfb189-fa4d-4eff-82ba-e3701c02243c}} is independent of {{formula:3aee7ca6-793a-4961-bb05-4a5182eef2d2}} and that {{formula:47a9c050-e475-44e6-b2a5-1275ff8d5dba}} is conditionally independent of {{formula:6bc81797-1938-4d1c-9d5c-5e10e3dcedda}} given {{formula:ab4b3b3f-7382-4a5a-8e21-37c8a0d22327}} . In general, there may exist confounders that affect both {{formula:533f0a1a-4c31-4a5c-b0c9-302c91377b45}} and {{formula:9afd9c1e-57ec-44fa-9087-fafa99724a36}} .
This confounded model coincides with the causal model assumed by instrumental variable methods, where one attempts to use an unconfounded instrumental variable (here, {{formula:32e21009-a1c8-42eb-86ef-e1b8c2a634d4}} ) to deconfound the relationship between {{formula:afedc3ba-8820-4220-87fe-8acfeb6afda3}} and {{formula:7a8f88cb-0f31-4978-9773-51bcae1c2478}} , typically by projecting {{formula:40ae9f64-2fad-4e0e-86b6-74576b01d49b}} onto its variation caused by {{formula:984bb569-a488-4d4a-99e7-45a226dfba7d}} {{cite:b2acf51bf9e58c305bdf2d65192885c275147f19}}.
In the passive case, {{cite:4a389e413a1827d8eec8f1af260d8d2696704874}} recently provided minimax rates for nonparametric instrumental variable regression, but we are not aware of work characterizing potential benefits of actively manipulating the instrumental variable {{formula:437f6ad8-f454-46ae-bad5-4383f6fd8ec0}} in this setting. Indeed, one motivation for the present work is to provide a baseline for such investigations by characterizing minimax-optimal learning in the simpler unconfounded case.
| m | 3039481c0ba5fb1b72d07d9006d6a0c2 |
Thus, Weierstrass Theorem {{cite:e76b2b0c36af55844a6eba10d268dcb839faa466}} assures that (REF ) and () both have a solution. As a consequence, the splitting method above is well-defined and an infinite iterative sequence {{formula:5847539c-b8c4-418e-a8c2-7750863d816d}} is generated. Moreover, {{formula:2e7be240-0990-4ef2-829b-e94a283198c7}} can be seen as an infinite iterative sequence to approximate the minimizer of Optimization (REF ).
| m | 3cb154ab19617af6daaf62e98870a95d |
Physics-informed deep learning methods {{cite:ba577e6603e4c781f4bdd22a83fb639695893fb1}}, {{cite:68a2614be2acc8a85c6868b9c1dda24b7ce0d6cd}}, {{cite:e9ed77292d61f3e36676c57a219ec4e9e5f952bc}}, {{cite:64be13dd4a5119398dbffea0e1aaba068e50efa3}}, {{cite:ab1f0ef5cfb31abf5205ea17ddfe05640d2bcc85}}, {{cite:40be6152d516370190a06248a1bae5e37106ecf8}}, {{cite:e0fdebde7c268e7c0ecee7a4a15c6447951cede6}} have recently achieved great success in solving nonlinear PDE problems that arise in different fields of STEM. This entails using deep neural networks to approximate the solution of PDEs, often in many dimensions, leading to a mesh-free numerical method in place of more established schemes, such as finite elements or finite differences.
| i | 2072dc40969f1cad43fbbae9c61c0c91 |
In this study, we present two deep learning-based methods to align audio to phones and achieve comparable performance against several existing forced alignment tools. These models can be combined as a pipeline to bootstrap phone labels from naturalistic audio data in the wild, which are massive but remain under-exploited. This will have implications for speech corpus creation, phonetic research, and interdisciplinary studies. Given the good performance of {{formula:1daee03f-3301-4cb0-924b-b0d8c521e4d8}} , our method can also be combined with MFA{{cite:490830ca241d61d675f51fce1ff650b3d23f65c7}} to train a large-scale text-independent frame classification model to get rid of text in inference. The W2V2 models
can be accelerated by harnessing the computing power of GPU. For example, in the inference mode, our model aligned the 1600 hours Common Voice dataset in about 10 hours on a 2080Ti GPU without batching. While only tested on phone segmentation, the model also performed well on related tasks such as word or sentence segmentation.
In the future, we will continue to improve the alignment accuracy and extend it to different languages.
| d | ab4fd189e7cefb9d275882a5d145bc81 |
An overview of the results in Tables REF and REF convey that the proposed CIE-Net, employed within the incremental instance segmentation framework, shows neat performance improvement over standard models such as Mask Scoring R-CNN {{cite:90dd5e72cd96b89fb7051814f90bfe0d687f88ec}}, Mask R-CNN {{cite:75e1e6f9bc857234a24d68f377f01eea90f99a57}}, Hybrid Task Cascade {{cite:403737866c48084d92cb7978d6f53ecbe210a8b2}} and YOLACT {{cite:ef42460b17b7ff1b147fe2e6e1d802adb932c2f1}}).
| d | ceced6d4ab27ef100724aa0cf8261864 |
Pre-requisites:
Without loss of generality, we focus on the SOTA multimodal contrastive learning model, VATT {{cite:abd108d0faac79298e5f3131eab3b6767ffe48de}}, as our default subject of study; and we follow its modality-agnostic single-backbone setting due to its compelling performance-efficiency trade-off, and due to the fact that it is the most intuitive setting to modify gradients (shared between paired modalities) w.r.t different end-point objectives.
| m | 0cd9cd541287a6a163c6f6200235e9e3 |
Creutz applied the lattice QCD simulations with the Wilson loop to describe the interquark potential between a
quark and an antiquark {{cite:91a673837c98c74c264f3f58f42bd4e85f4d74f3}} after that
a large amount of effort has been devoted in lattice QCD to
study the multiquark force {{cite:91a673837c98c74c264f3f58f42bd4e85f4d74f3}}, {{cite:17f270399f5fe0ce3f5c70cdaec1c1648161ff6c}}, {{cite:1c8b996be108855c0c4db3c2f1b8fca48cfab85f}}, {{cite:ed2ded07a26ea83f9c932bf95fc1fd35b44fa042}}, {{cite:9ad7a50af3a80eec66bfe43d7351d58340e31a55}}. These potentials are successful to calculate the energy and the mass of tetraquark systems contained purely heavy
quarks {{cite:afeaa04c4b5ddc536e44ed1494f595f3b0c94762}}, {{cite:6feef49ecbe53bdad355637c0dc1ace756167591}}, {{cite:be3d259a6c191280f6b372755625e43cd912116e}}.
The experimental discoveries of multiquark hadrons
reveal new aspects of the interquark force such as the quark confinement force, the
color-magnetic interaction and the diquark correlation {{cite:7ae1c4a24fec4d15187bd85da4f9d67e6ae2fbb2}}.
According to these, the proper Hamiltonian for
the quark-model calculation of multiquarks has been suggested to
investigate the interquark force in the multiquark system directly based on QCD {{cite:6102b75a9e4f6679e82e5edecc4451503920c02b}}, {{cite:ae4fd3bcf8ff396a8dc2b7388e14b793af46c540}}.
| i | 0fdabb4f544218ea48254faf41d96f46 |
Instead of using only {{formula:e7594918-5424-4c7a-a0f3-54b1942336bb}} and {{formula:afa56060-16f9-4f29-8104-24c5dfb80125}} orbitals for the even- and odd parity layers, one can also think of using {{formula:12425c30-0b54-4c9b-93ac-d8804968ec56}} and {{formula:2e7d6e44-5696-407b-8257-70f9162771a5}} orbitals, respectively, which will accompany electronic correlations. An electron correlation can be a curse for the stability of Weyl points, however, it can also be a blessing for incorporating several many-body non-trivial phenomena within the Weyl matrix. For example, it has been shown earlier that in the 5{{formula:12137013-d764-442d-9198-54cb1de5d716}} electron URu{{formula:0e157fe5-b821-4d42-8958-c3cc907c1066}} Si{{formula:856e3b80-c471-4a37-a447-91705152b1a1}} , the Fermi surface nesting renders the main band and the shadow bands to touch at the Fermi level (but not at the {{formula:beb51adc-398b-4f54-98f3-f516150f7286}} point) with Dirac-like dispersions,{{cite:9179ef13e961726e512f1419d2c2f0d354e281b2}} which resembles the so-called `accidental degenerate points',{{cite:4b9f3c3d01b37a4d03b03f322017612c80b5fc26}} which we call now Weyl points.{{cite:772ebf533b22e0968d07678f669fa2c159c9ce46}}
| d | 9f68ca3fcf6bebdf1245a2d33b0725e2 |
Firstly, we consider mmWave point-to-point MIMO systems.
The transmitter has {{formula:f8f7278f-637e-44d6-b3de-ce5de2e9de1e}} TAs and the receiver has {{formula:391e65ea-436f-4b73-b6bf-945b25a9df57}} RAs.
The number of RF chains at the BS is set to {{formula:dcd0533d-67d8-4400-b3f4-aa4c7e69b61b}} .
The number of data streams {{formula:5b205bf5-bb09-4b4a-acfa-1151482386e1}} is set to 4.
The number of propagation paths is 8.
The azimuth AoD {{formula:b84caa9b-7544-4c62-95b8-95b870752b3d}} is uniformly distributed in the interval {{formula:f94ce5ff-d983-4adb-abe2-6d691de0b37f}} and elevation AoD {{formula:ad64c577-42be-4dc8-adf6-eee1fda1ce12}} is uniformly distributed in the interval {{formula:67d30236-93b2-4d5e-8233-02a84d2a1074}} {{cite:76a04a57b06fe40d0dde95f8b37482ba00f1a64f}}.
| r | a2f31068c3048cefc3c2e52113ebd9f2 |
Non-parametric algorithms feature a key quantity known as scan statistic, for example CUSUM statistic of {{cite:37ad07c57e0649a4371d073f65eed06d61739439}}, which is required to `scan' the dataset to identify the change points. We propose a scan statistic based on influence functions proposed by {{cite:bb4c7258f011089eba3b4cf5e4702d4be5b74cee}} that can handle outliers and heavy-tails, to deal additionally with contamination. We consider a contamination model, where the
outliers (corrupted data) are correlated to each other and to inliers (uncorrupted data). The inliers can also be correlated to one another. The resulting robust non-parametric algorithm RC-Cat announces a change if the scan statistic exceeds a pre-specified threshold, provided the scan statistic is a local maximum. This additional sophistication of local search methods was the introduced in {{cite:b0bb2c47df0c6239a6649d165f46ec25fe99d39f}} and developed for the robust version in {{cite:25a4298b6c7cbf859d24658d72d6d852298d4941}}, to mainly avoid overestimation of change points.
A natural way to theoretically evaluate change detection algorithms is to establish consistency of the estimated change point indexes as the number of samples increases. In particular, we show that RC-Cat is consistent in the presence of contamination, i.e, as the number of data points {{formula:f2ce43dc-b048-45b3-9b19-6081a8b48e43}} ,
{{formula:25f408b3-0272-471c-bb6f-79ac688af769}}
where {{formula:0075ccca-fabf-422d-bb14-5291d3d45800}} is the number of true change points located at {{formula:875da86e-72e4-4b70-be48-9f497fb6b9da}} and {{formula:9bdcda2d-cfc3-4ff7-9cbd-5bfefc288d71}} is the number of detected change points announced at {{formula:328ea763-5713-4235-9825-c4815cb8061f}} , and {{formula:a20497cb-dc96-4ffc-80b9-d0401a0ffee7}} is parameter that related to the window length considered.
| r | 15d6a599b8d5eff34f57d48ff6d4ef50 |
Let {{formula:9e50d849-2403-44e2-aaff-4692a0498871}} be the sequence generated by the algorithm (REF ) with the primal step-size {{formula:ab245868-735a-4688-97e6-bc2638171335}} .
Following {{cite:c9f1c07600d91c542065afcbe541c5446587aef9}}, the update (REF ) satisfies that
{{formula:04b407b1-d7da-4684-88b9-760c7a300771}}
| d | c4caa3c54b57fe94b1623da314432b6f |
Attention Model {{cite:e678332a349bd7c30c0c5f75d81a8bf738dcd1e6}}, {{cite:fa579b4c01fb7e665a4556dad469f6c241d08c92}} was first introduced in machine translation task and the attention weights were later widely used in natural language processing tasks as explanations in neural networks {{cite:09ca82bfee670d39c733535789831d56fbe1f4e3}}, {{cite:76b727cc6395c66b767af2cc963adb21030a8f64}}. Other than Natural Language Processing (NLP) tasks, the attention mechanism is also a near-ubiquitous method in recommendation tasks used as explanations in some works. {{cite:b3dbe52c1612d22d04f68219671bc5742fc252a7}}, {{cite:4786fb438b651d62f413072944505de16fcbec9b}}, {{cite:9515431cdccb77a3db9ce21d90d7670225e21623}}, {{cite:5eaad9cc88c730da0b1dd3a2cb66e702c86c50a3}}
However, there are different opinions on whether attention mechanism could be used as a way to explain data {{cite:9c612692a15f95a5bf79e532a38d6cecd46d1dd6}}, {{cite:ce248b83674bcfc326beed8d405d980db2842112}}, {{cite:5d501c6243907cf9674055b390b0d8f00854c0c8}}.
| d | 61e3c26a5b0863c158f03a6b32c174b9 |
As shown in {{cite:7ec696b931add09f37ad6ce7b32238769e5758ad}}, {{cite:d0063e9675000b6a365264eda3c26f145e31dc20}}, the norm of the Moreau envelope — {{formula:2f1c11c1-d311-46d6-90d6-4e984c6713ac}} — defines an alternative stationarity measure for problem (REF ) that is equivalent to the natural residual if {{formula:9f85e009-3bb8-4d6b-8848-6a975cb4c92e}} is chosen sufficiently small. A more explicit derivation of this connection is provided in lem:env-to-nat.
| r | e3bf2f2159ff2a0be12262a4a32d156c |
A number of authors have used this technique {{cite:fc7a1e70a1e6df34b7dd03f28bd69faab8073a27}} and tentatively found results consistent with those of {{cite:a3170010ad80cec831c836f373abef355814fd0a}}, i.e., that in giant elliptical galaxies the mass to light ratio is such that one must have an IMF that produces less light per unit mass than the Milky Way IMF.
| m | 515b3f3110f18c0de3910eb578c44a0f |
ReLICv2 demonstrates for the first time that representations learned without access to labels can consistently outperform a strong, supervised baseline on ImageNet. In terms of a like-for-like comparison using ResNet50 encoders, ReLICv2 represents a substantial improvement over current state-of-art.
This is a direct consequence of incorporating better strategies for selecting positive and negative points in the ReLIC framework as suggested by the theoretical results of {{cite:c8930904044fdd2517f095e12c28dcb6284650e0}}.
| d | 88a2c77a33e915fe0e21d55345e2b243 |
However, these methods have limitations in extending to other applications such as game graphics or metaverse. For such tasks, a canonical view of a target should be known, whether it is given or predicted by the model. In reality, a facial image given by a user is not always in frontal view, but rather from an arbitrary angle. However, existing GAN inversion methods are initially aimed to reconstruct the original image thus they cannot synthesize the frontal view of an image taken from a random perspective. Although several methods {{cite:b1200a11f97c99ff9cffd543d86e2fc2fce37256}}, {{cite:27562ea010ebcd57e514d2df8a79cbd19a94e2e3}}, {{cite:7cdd602fb16ea3bb2f0470041b4fff2740c77bb8}} have shown the successful pose control by discovering a direction related to pose in the latent space of StyleGAN, they haven't found accurate mapping for frontalizing an arbitrary image in an unsupervised manner. InterFaceGAN {{cite:b1200a11f97c99ff9cffd543d86e2fc2fce37256}} is capable of obtaining canonical pose by using a semantic hyperplane, but it requires supervision (landmark) for binary classification of yaw in order to calculate the hyperplane.
Several off-the-shelf frontalization methods {{cite:10cb0bbe395e13f8b71b4fdce647340e58dc5b87}}, {{cite:27b621351c0406a5a2398d2bc7a1c9ba6b38d2c2}} capable of handling high-resolution images can be considered here to obtain a canonical view of stylized images. However, most face frontalization methods are trained with real facial images, thus their performance would decrease significantly when applied to stylized images because of the domain gap. Application of a similar approach frontalize {{formula:8ee9ecb7-ff7c-4e4e-8fac-b658dff8376a}} stylize also presents degenerated results that identities of original images are not well maintained in the output.
{{figure:65b768b8-3b78-4b10-a23c-8dc1038fac79}} | i | 48141eaf45f91fc776933d36f9f83a00 |
Vision Transformer (ViT) {{cite:e1b1a7310f1dcf82e9828ac21b116cd17fa55428}} architectures have recently gained traction as an effective alternative to CNNs for computer vision tasks {{cite:1753e644f4369512c4dcc797287e90cd9131bcac}}, achieving impressive performance despite fewer inductive biases.
With their in-built self-attention mechanisms, and improved calibration under distribution shift over their CNN counterparts {{cite:b1fed7be94ea5e1587c0b27d2024a81ea24c5642}}, ViTs may be particularly well-suited to domain adaptation {{cite:812aa1f663d089066c81f60f0660455ec3dc3111}}, {{cite:cf800eb4f8d3cc0cd6f9a22d0d715bec7daa7ece}}.
| i | c7541af1cbef6a81c645b3f7abf40c4d |
Our experiments for texture recognition were performed using the Materials in Context (MINC) Database {{cite:f87693424b7e3fa6e3348df03b039a09b696ecaa}}. Specifically, its subset MINC-2500 is adopted. MINC-2500 contains 57,500 images of 23 classes. Each class contains 2,500 images. We use the train-validation-test split 1 provided in the dataset, with 2125 training images, 125 validation images and 250 testing images for each class. We employ the HyperMS-SSIM tensorflow-compression implementation available in {{cite:b1089a3b337d8eec6be0efcedefa74f4033944a0}}, which provides 8 trained models for HyperMS-SSIM corresponding to different quality levels, and thus rates, named as “bmshj2018-hyperprior-msssin-[1-8]” (1 for lowest quality/rate and 8 for highest quality/rate). The compression is done by applying the pretrained “bmshj2018-hyperprior-msssim-1” (HyperMS-SSIM-1), “bmshj2018-hyperprior-msssim-4” (HyperMS-SSIM-4) and “bmshj2018-hyperprior-msssim-8” (HyperMS-SSIM-8) models which correspond to the lowest, middle, and highest quality/rate, respectively, on all images in the MINC-2500 dataset.
| r | 63d31a7c93a44d6dd2d5f7fb6fd9dc3b |
Despite the fact that Fields' results have achieved the so-called uniform reduction in
the sense of Olver {{cite:4b742516f9a1ab85449394069316f64b5ba2079c}}, they are found to be too complicated for any practical
application; see, e.g., Erdélyi {{cite:a7d584e7ac9608f8319d0b75b582cd280a797ba9}}, Olver {{cite:4b742516f9a1ab85449394069316f64b5ba2079c}}, and Wong {{cite:ca47722ff6d9c0818ec55473995350a2cd208e89}}.
For example, Olver commented that
It may be desirable to investigate
whether any simplifications are feasible since the results in {{cite:e546157238dc2c100b222a2f8d6c29346e4dc3b5}} are rather complicated
to apply in their present form.
| m | 25974fcc3337fb80ebc0681b2112b86e |
We compare the pre-trained models of GraphSage/GAT
in {{cite:43fdbaec0eae8dfb89d4fd50dde81dfc6f7c5934}} with GraphSage/GAT of the same architecture (5
layers, 300 dimensional hidden units and global mean pooling) trained in
2STG+.
The pre-trained models are fine-tuned on the datasets.
The mean and standard deviation of test accuracy of 5 splits are reported in Tables REF and REF .
Despite being pre-trained on a much smaller dataset, 2STG+ achieves better accuracy in {{formula:50aae657-7b0b-4713-ab49-24cc3302a1e4}} of the considered cases: up to {{formula:a67e41b6-f51f-4d14-8f7c-875b6760db2d}}
points in the benchmark
datasets and up to {{formula:6265ef34-7177-446b-b95d-27e9e5fb1acb}} points in the NYC Taxi
datasets. This validates our claims in Table REF .
| m | 4240ba0633b2a17311f32340783fdb85 |
We recall some basic notions, introduced in {{cite:c6c718178a88df1f4e65e1b9473d68afa01ed23b}}, {{cite:2f9c23017a2dbae12db784bd11f45cfe929f0836}}, {{cite:5961aed91a7abc81baa51985abc9274a6ad24274}} related to Hom-algebras and while dealing of any binary operation we will use juxtaposition in order to reduce the
number of braces i.e., e.g., for {{formula:b5f214f8-d4f3-479f-9a04-e28a823c73af}} {{formula:893c74ec-bec6-4a2a-b005-e7bcd3be63d3}} means {{formula:4ee393a8-c03b-4074-b9ce-95cc2458ddb0}}
Also for the map {{formula:83deb92e-a150-4051-8ba0-67440e36dbc4}} we will sometimes {{formula:8c4b132b-ec13-49ef-92c0-64736bba6b8f}} as {{formula:a6b37d79-0aa9-482d-9591-c265ba0b291e}} or {{formula:182b350b-5e2a-45fa-89e5-05a88fc5b55f}} for {{formula:36827704-d310-443d-ad28-8dcb51c4be32}} and if {{formula:9569bd9d-abbe-4bcf-b359-90a9d62d63f2}} is another vector
space, {{formula:a4faee10-f7ed-4a98-82f0-41e3dd131152}} (resp. {{formula:b41494ce-24b9-4555-9067-c1a729ec991c}} ) denote the twist isomorphism {{formula:65511524-7284-4728-b97b-d838de6a3927}} (resp.
{{formula:238da7d4-37a8-4306-970a-d725322e80b0}} ).
| r | dc081de1ea2196df1b969a82b9f95899 |
The seed binary is treated as two sink particles with a sink radius of {{formula:d188c16e-e87f-456a-a51c-8384799a6365}} ,
following {{cite:b9283f6fb9ddcf66eadb35c765141b52d6d5e63e}}.
The SPH particles are removed from the computational domain once they
fall into the sink radius of each seed or reach the outer boundary, without any feedback to the seeds.
| m | d28f447c4db433b8493ea75c187bd133 |
There are several interesting avenues for generalizing our theory. The proposed inference as control method can be extended to the more general case of observer taking external actions in addition to the internal predictions, in order to maximize external rewards while minimizing both external action costs (e.g. movements) and internal computational costs. This could be modeled as a typical LQG control problem {{cite:01f6122c703fc98987089a2c87b891f2b09a1f9d}} for external variables by including appropriate entries in matrices {{formula:89c34ebf-c4f2-48aa-820a-e7e784148471}} , {{formula:ebf66058-fe89-4b30-9d40-62b10aac4a07}} , and {{formula:63c4c92c-08b2-42e4-8832-47dc58234667}} in addition to the internal controls we use for predictive coding. Though our results are limited to just a simple one dimensional linear dynamics with uncorrelated noise, our future work will explore these computational constraints in more complex multivariate and nonlinear tasks. Accounting for graph-structured inferences {{cite:23e4f8ab908ff42354e8586cacd3b989efa2f533}} and controllability {{cite:2552e469e790055d0a7ead936d0a63ad794745d8}} may provide additional constraints on the brain's inference processes, and additional predictions for experiments.
| d | 1b37dc39ce43b69edfb5203f34f827ce |
The set {{formula:18db4856-6400-46f3-964f-35c17a11a9da}} is the (disjoint) union of two sets {{formula:62ed8839-d1c8-4761-9907-333723405a20}} and {{formula:86dccd22-3fce-4d02-b75e-3da54d6791a3}} , consisting of rank 0 elliptic curves with ordinary and supersingular reduction at {{formula:a1af83ca-b505-41b5-ad03-00a938a86702}} , respectively.
For {{formula:1cbc5fc2-8fec-41ff-bba8-f8c8ec78fe7f}} , write {{formula:aae93ad7-3335-4d31-b94e-3ca9308703d9}} for the set of isomorphism classes of elliptic curves over {{formula:05f34602-6925-4b44-9ba6-4b0c922f35e3}} of height {{formula:0c79d290-eb05-4e5c-94d8-d94cf89b9763}} .
If {{formula:62db30bb-013c-4d66-88e5-b0175a0a9d8e}} is a subset of {{formula:75b7e4d5-0392-4124-bd4c-1112e6eb6793}} , set {{formula:1d1c786c-d5e9-4ad4-8f90-bf2626f7f75f}} .
It is conjectured that when ordered by height, discriminant or conductor, half of the elliptic curves over {{formula:61c8cb2a-30b1-4c5d-8126-25b99c3f654e}} have rank 0 (see for example {{cite:6f49bbfba0c1981e9979227855af73e88eb83400}} or {{cite:297ce03e2960d00b4fc1d34a85b8b8398a7eced3}}).
If {{formula:3534b94c-56fe-4aa0-a33e-9006ec0ead14}} has good ordinary reduction at {{formula:dfaca6d7-5865-4bb6-a8fb-8ea57fee8499}} , then the Euler characteristic formula {{formula:2ceb2080-48f7-43ad-98ac-790bab07ee61}} states that
{{formula:d62311f9-86da-493e-ab00-49575cd16d6f}}
| r | 8f6e3cd574aa3a779347b7f1b801a296 |
Under genetic drift, the eventual fate of any allele is fixation or loss. Although drift may seem an additional complication, it also has an interesting effect, namely to allow access to parts of the fitness landscape that were inaccessible from a given state of a deterministic population {{cite:41d78f2a332c539ca21315762a8126c4a89680eb}}. In this sense, genetic drift smooths the landscape, and although stochastic effects are introduced, the expected trajectories are somewhat regularized, because the populations can easily escape suboptimal peaks {{cite:367b312f201ac772973090562b232c33629a9698}}, {{cite:bf5dab3a67ddba1ddfdbd63ff3483edd9e2c17b9}}, converging to fitter states. In this sense, drift aids adaptation, allowing alleles to jump across peaks by mutation and genetic drift {{cite:ba51cca6dcd766b52e15fde5e77ae9f0fc52fcb4}}, {{cite:40b5b39e5a5a8d5db75b5252850fe606df4b1aae}}, {{cite:a37163c9210e2204ce66a6f4add0ad3b9f995878}}.
| d | 92267c62230e4995e22f161fed73b679 |
Bayesian optimization (BO) offers an efficient alternative when the tuning objective can be effectively modeled by a surrogate regression {{cite:c3aa28040bd5e6598603df5dc2bb940578b204a3}}, {{cite:10cfbdbe3434ef936d7eb51e8ce5893f13138f7f}}, or when one can take advantage of related tasks {{cite:645a9396dbf7b5bdc3a13751fe5a259c57592d14}} or strong priors over problem structure {{cite:bcd006947f524f65e10b9a855afe05c3559e0277}}, {{cite:39b2bc675fa640f8dfb43ae952d49d875243b3e8}}. BO optimizes an expensive function by iteratively building a relatively cheap probabilistic surrogate and evaluating a carefully balanced combination of uncertain and promising regions (exploration vs. exploitation).
| i | b9248e602077b0d653680f953a0f0338 |
We have proposed, analyzed, and implemented methods that accelerate planning performance and optimize solutions. The key is to quickly compute bang-bang time-optimal controls using analytic solutions, to produce both metrics and steering methods. Although the study has been limited to RRTs, we expect it could enhance other sampling-based planning methods that rely on distance metrics or steering, such as probabilistic roadmaps {{cite:e3bb6ffbee3e69107bf5db298ccd349108d2cb71}}, {{cite:12318e2d6ea27a96458e79beefcb0dd78d8e088b}} or expansive space trees {{cite:5657d0ca76eb1f5790b06f1b4b03e04466ae36d5}}. One of the key observations of our experiments is that plan-and-optimize is superior when applicable: It is more reliable to explore the C-space first, lift the solution into the state space, and then use bang-bang optimization. However, this applies only for rest-to-rest problems; for more general problems, a bang-bang enhanced RRT could be applied to bring each of {{formula:8128bee7-cd26-4cca-a4bf-dd0b60a5fd26}} and {{formula:8681f8a6-0dcb-4174-9ffa-c506165c5db9}} to zero velocity by biasing samples to the {{formula:d7d2c735-c383-47bd-8e49-a7942ffbf76a}} plane.
| d | 86fd9ca132161df36f024d0a662f7de6 |
where {{formula:8cdc1344-495a-410a-992d-4e0777f0d876}} is a polynomial in {{formula:9e28b9d4-9877-4e80-87af-707ae21f75e7}} . For fixed values of the constants of motion, the spectral curve, {{formula:1a52c7b0-6fd2-4481-b024-c0f845004f66}} , defines a Riemann surface. A link between the initial phase space and the Riemann surface can be established from the set {{formula:f2e6f546-d4ad-4c33-88ad-dd769b218c50}} where {{formula:b1532494-2bbb-4648-9845-0dffaae9e4b1}} is a solution of the implicit equation {{formula:0d8761db-2ad4-480d-b3f6-9b1de4839c65}} and {{formula:d2875160-fe22-49a8-b647-e13e7d22f5a9}} . The set of coordinates {{formula:4aa30cac-4755-4480-9854-538655312477}} is related with a vector bundle that can be defined through the eigenvectors associated to the eigenvalues given by {{formula:2e727b37-a4d6-4069-a86e-e5e983159643}} {{cite:04ca0532668d32a5e3361f779f9623631d9b6c7a}}. Note also that we can replace {{formula:beec54b5-90b4-4728-89f7-f0d48e9ee104}} by {{formula:2586ec15-4b2e-4d03-b133-fc4e44dec06e}} everywhere.
| r | 2192d895452b3e4a08f35575cd130ed1 |
We use faster r-cnn {{cite:8f58806319c6825eb82ab6f00988db6abf65fad4}} and mask r-cnn {{cite:1b5f1a753bdf5878defe296a733fccd422540e5b}} as baselines for detecting graphical objects in annual reports. We use publicly available implementations of faster r-cnn {{cite:ee53abcfaec02d4c1927b612113316c0417e167c}} and mask r-cnn {{cite:960dd54ce5c0a6636bf8929c1264417a6499ccba}} for this experiment. We train both the models with the images of training set of iiit-ar-13k dataset for the empirical studies.
| m | 674925dbc3c34cf3e86049fbd62be179 |
Carrying double different flavors, the {{formula:a281f088-10fa-4268-8976-3f6be97d1ef7}} tetraquark states can not decay into a heavy quarkonium plus a light meson via annihilating a pair of heavy quark-antiquark. There is only one kind of two-meson strong decay threshold {{formula:88167817-5145-40d2-93dc-6592b0d93864}} for the {{formula:9a43d9c1-7e1d-44a5-9979-384b3dad3bde}} tetraquark systems. To date, only one ground state of {{formula:93c0df39-f73e-4fe7-a606-aa6606fb7f67}} meson has been discovered and confirmed experimentally with {{formula:b0ee0d00-80a8-488a-8dc2-e38487d1fc48}} GeV {{cite:91368e258d325ef508e330f24b8680b8e93ee0a2}}, {{cite:ad79fbdae521fed98414fdef26695e5a184baa7a}}.
The spectroscopy of the other {{formula:1e594e06-d8cc-412f-ac2e-37946ed28ba3}} mesons has been calculated in the relativistic quark model {{cite:cc1f4f28f4138a3941b01e425af22f8dfe6e3da8}}, in which the
mass of {{formula:0fc779a6-60af-450b-914b-fcc5010e60cc}} with {{formula:19f4e8c8-dedf-45ef-91f6-b147301ae26b}} was predicted to be {{formula:54327d89-b9f0-4dfc-8148-49efeb5ab855}} GeV.
| d | ea00453c19c24a119c4254c97274c71b |
Neural SDF {{formula:020599d8-bd74-4f66-8596-49eef4046cd1}}
maps a 3D location {{formula:a0ab8ecb-3081-48b9-9106-973c3143ffb0}} to an SDF value {{formula:c42195e6-2c08-47ff-a1df-c8096a1f1af9}} and a 256D local geometric feature descriptor {{formula:3db90ebe-baf9-48d2-897d-65a751a51689}} , as in recent works {{cite:8930fcfe77e411eeba54b09f7b02e61f63ebedd7}}, {{cite:8727dfd22fcdeb796a173a8330e5f231583d74a5}}, {{cite:976e678b43f9cc1921aea9348a795a58a3d771ac}}; this feature descriptor can then be fed into our neural material networks.
| m | badc1c522243ed13d9ba11bdbf296846 |
Results: The quantitative comparisons between our proposed method and the others are presented in Tab. REF and Tab. REF . The first interesting observation is that the widely-used CRF Loss {{cite:00b802948937a5420a81e98e16db7d973b99fe8e}} achieves the worst performance than all other methods. The reason may be the CRF Loss {{cite:00b802948937a5420a81e98e16db7d973b99fe8e}} is specifically designed for natural image segmentation tasks and is not suitable to handle the CT images with low contrast and non-enhancement. Then, we found that the network is capable of leveraging the scribble annotation more efficiently by encouraging to produce more confident predictions. Moreover, compared with the entropy minimization term, our proposed intra-class intensity variance minimization achieves better results, the mean {{formula:45080966-6025-41cb-92ea-a0d2481f9617}} of 73.90% {{formula:803e2f82-8a1a-4586-a6eb-dc264088af0c}} 74.99%. In addition, combining the entropy minimization and intra-class intensity variance minimization, the model achieves the best performance than the others and improves the mean {{formula:bd40a3e9-e6e5-4458-8e4b-834e7439db40}} from 72.06% to 76.81%. These results demonstrate that also most weakly supervised methods achieve better results than using partially cross-entropy loss, except the CRF Loss. It is noteworthy to mention that the scribble annotations save more than 96% of label costs than dense annotations. Finally, we find large size organs weakly supervised segmentation results are very close to fully supervised, especially in the liver, spleen, kidney, stomach, head of the femur. However, the small size organs still can not be segmented well, it also points out the research direction going. It also shows that with further research, weakly-supervised learning may further reduce the label costs in the future.
| r | 5c5ddde4d62dbe1d5a2e3989f4c94e5c |
Previous results have modeled the spread of information in graphs and diseases in populations for either single-type graphs with general degree distributions {{cite:74841e7c294856157f3c0ad64678c8f00af9d04c}}, {{cite:d99813248ce272500c89ad7725315a53ceb1b5dd}}, {{cite:86f9d4333643654b6bfe8f3cdaa17136c39b2883}}, {{cite:a63a7652f77767b6a566a6fbef8e7f2d48d44cf3}}, {{cite:73b05c3aaf3018f950a26e9fe7f81f7c8ab33399}}, or for multi-type graphs limited to Poisson {{cite:7d66d158496292349b74f8698a015ddce89d6564}}, {{cite:b15368da937993c70be03d250d17b97301237330}}, {{cite:04bf64d2952ae744f4cc0e04da1a61adef686937}} or Poisson-like {{cite:32d2e26efaa3fad98b4aa251c84e4c9b26dcc6cd}} degree distributions.
In this work we derived a method for computing the size of the large out-component of multi-type graphs with a more complex degree distribution and presented an example of its application to the case of epidemiological research.
| d | 0b1a500b96679a8755babbfabb852c4a |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.