text stringlengths 54 548k | label stringclasses 4
values | id_ stringlengths 32 32 |
|---|---|---|
Although Warner (1965) ({{cite:b2074f7bc892a645e2ad15df15d779a7a3600e36}}) and Simmons (1965) ({{cite:cb89b85152682f54be2bab47d5d4a612667de546}}) RRTs have been used here as illustrations, any Randomized Response Technique can be implemented as a web-based survey using the code given in the manuscript, with the mere modification of the function "randrt()".
| d | 6d7bc57a72a49d21bcae83fb9c3b95f5 |
whereas the series part of the solution around {{formula:4212c3bc-bd9a-47b6-9568-ef92229a5772}} , denoted {{formula:77c2f01c-5d0b-4c43-910d-a554d2b72d03}} , satisfies a similar equation. By a succinct argument featured in van der Put and Singer {{cite:125dd210e42eb0456ee658868db1be09c72807a1}} we have solutions, {{formula:b94831c6-594a-45dc-9218-2a5d25008b2b}} and {{formula:c535a92b-407a-46da-9e4c-01775d817ea8}} , that are convergent in neighborhoods of {{formula:490454e6-6981-49ac-8b2a-18c9edd18eb4}} and {{formula:170d8aec-3569-4187-8723-3efe4e2e88da}} respectively.
| r | c9d05d43e8a619b88dd086bc78e45c40 |
The success of hydrodynamical models in high energy nuclear collisions has attracted the interest of the theoretical physics community to understand the foundations of relativistic fluid dynamics at extreme conditions. One of the most relevant advances towards determining the validity of hydrodynamics in far-from-equilibrium regimes was made by of Heller and Spalinski {{cite:38d697c9764291fd07c0f92b91feb52634025213}}. In their seminal work the authors investigated the convergence properties of the hydrodynamic gradient expansion for a conformal fluid which undergoes longitudinal Bjorken boost invariant expansion {{cite:550368d5fb8e206a0013df5fafcd908ca05feebc}}. Within their approach, the authors showed that this perturbative expansion has zero radius of convergence and thus, it opened to the possibility of a new systematic resummation scheme. The authors also provided strong evidence of the existence of an attractor since most of the initial numerical data merged rapidly into a unique single universal line. The late-time behavior of this universal line is described by a few terms of the perturbative gradient expansion. Surprisingly, it was noticed that the Navier Stokes regime is reached out rapidly when the pressure anisotropies are extremely large. Altogether, these results strongly support the idea of a new theory for far-from-equilibrium fluids whose physical properties are still unknown {{cite:c1d32f097886a24e05e714424cabe8ab0b29ef46}}. Subsequent works have verified and extended these results to more generic conformal setups in different coupling regimes (see recent reviews in this subject {{cite:d2d7a6207907362432e4a57591d5b3fb70175c67}}, {{cite:8602bc8b9ca6ec88583140291217568b9b11589f}}, {{cite:14a8adea8294480b1c8dcf0199dca5f4ec76745a}}, {{cite:77b17bfd318e4a2c1a9d38d0e2fc6dc83b0b3b82}} and references therein).
| i | 917a9ca7f7dab48788e70d5032cd26e4 |
We focused our study on two specific toy models of evaporating black holes, but it would be interesting to generalize our analysis to other models, including the examples studied in {{cite:f729f748722e3922a15e7d398d2c408fa95a62fe}}.
| d | 013bf3f388a9ef04ee33f4f4c3b02270 |
The SDAi method is based on the pre-training of each layer of the encoder. While pre-training is known to prevent overfitting {{cite:fc8d6cefd02ceeb98a749762833d311cdd4c729d}} and combat the vanishing gradient problem, it can mostly be replaced by efficient regularization techniques and the use of the rectifier activation function. Nevertheless, we found pre-training always to be advantageous (smaller imputation error). This finding was based on the comparison (results not shown) where pre-training was replaced by standard random initialization of all encoder weights, followed by the fine-tuning session. We found for the small datasets that pre-training can reduce optimization time and more easily prevent overfitting. For big datasets, the difference between with and without pre-training estimations was less visible.
| d | fdf9a96bab0eb69e7a198c7bff6233dd |
Deep convolutional neural network (CNN) has achieved great success on image classification with abundant labeled data {{cite:086ba89deb2d772274b1010237ed9f0a5bbea396}}. However, for many rare or new-found objects, acquiring so much labeled data is unrealistic, which limits their applications in practical scenarios such as drug discovery {{cite:bee2370c96348a2ff4213e5e5a43e9c0d8c6e74a}} and cold-start recommendations {{cite:7dd252a9703f6d3fc1696bfd347eb2110fe4569a}}. In contrast, humans can quickly learn and recognize novel classes from very few observations. To bridge the gap, Few-Shot Learning (FSL) problem has been proposed and has attracted wide attention recently. It targets at learning transferable knowledge from some base classes with sufficient labeled samples, and then transferring the knowledge to quickly learn a classifier for novel classes with few examples {{cite:95515af4a329f2a7f8fc61d23520f287641a606f}}.
{{figure:b5ff4120-8967-474b-b2d2-1285feeeb6f2}} | i | 9624c3dbf9ae13c6a256facfe858781d |
To exploit the prior linguistic knowledge in the online dictionary for TTS systems, we propose Dict-TTS, which explicitly captures the semantic relevance between the input sentences and the dictionary entries for polyphone disambiguation. In this section, we firstly introduce the overall architecture design of Dict-TTS based on PortaSpeech {{cite:c638e69b3108ac0900bd48c94d1caf1d6e5a072c}}. Then after the comparison between the phoneme-based and character-based TTS systems in the acoustic and semantic space, we design a novel semantics-to-pronunciation attention (S2PA) module to learn grapheme-to-phoneme mappings based on the semantic context; In general, Dict-TTS exploits the prior pronunciation knowledge in the online dictionary with the following steps: Firstly, the character sequence is fed into the self-attention based semantic encoder to obtain the semantic representations of the input character sequence, and we utilize a pre-trained cross-lingual language model {{cite:5c54f687fc82fb22dddebd7b5dd14a2097febc43}} to extract the semantic context information in the dictionary entries; Secondly, we calculate the most relevant dictionary entries of the input graphemes and obtain the corresponding pronunciation sequence contained in the dictionary entries; Finally, the extracted semantic context information and the pronunciations are fed into the linguistic encoder for feature fusion. We describe these designs in detail in the following subsections.
{{figure:67147a2b-5414-4e93-b127-e4e3bf98c562}} | m | 0e375ea3222beb304904085c8f43cdbb |
where {{formula:e8e41c68-99d6-4e8c-a74c-0b51b44509e9}} is the {{formula:48d6dab8-d80d-460a-bf44-f8e2a658ec2d}} -row of {{formula:863cf500-d535-47b0-bc04-c8321557b351}} randomly selected at iteration {{formula:1efb7ac6-32b2-4b54-88b3-ea1d8d155928}} and {{formula:f52e202e-73a1-45c7-99cc-4104094b31c6}} is the {{formula:1160d931-5fd4-4300-8aae-3cec8a1cfdfa}} -entry of {{formula:1538fbaa-034c-4ce9-870f-1db5387b861f}} . Recently, remarkable progress of the Kaczmarz method has been made; see for example {{cite:2d5e37b3da5b604caf2f7bb1fa663915995a623c}}, {{cite:867f524c52cd33860b2e76ef3cc983ed430a8a8b}}, {{cite:858980f479dacaa7cfae82db44b6739e080f8b06}}, {{cite:441692f2027dc750c77713cb34c7c3fa62dd2f51}}, {{cite:428b58d64e25f499cdb3524bd38e3be664067894}}, {{cite:1d319146f7e263fb4d8d92a34da12ab12f39c087}}, {{cite:3eef870f52cc91f5c0ba8d4d33aa57d235b8893b}}, {{cite:a1e9b4487243f3f1fd5cb21314bd54f3b1fca0cb}}, {{cite:8702d20723314ee41450275d797f0ae4a3435dbf}}, {{cite:d98821bf7b645a8004cabcbb1dfb1e5c0db4d8f3}}, {{cite:621f6a628e31dd319a16d789fb3dc6af2626825d}}, {{cite:d97db5e02b633450bdc92e1907a05a2a5920ad93}}.
| i | b9339cd83c8d9decb1f9cf8f68154081 |
The integration of the two presented systems requires some specific temporal integration methods. Indeed, these systems can be seen as convective-diffusive systems with stiff source terms {{cite:251eec8f5923c95fa46cec6f43da1072ac072628}}, {{cite:e4c427665288da69b4e29f8fc30af4d3ffb38dc6}}, {{cite:fe03d9b10fe32d725bf1de584c07c048badfcd0c}}, as presented in Appendix . Some of these source terms (in particular those associated with collisions and ionization/recombination) are numerically stiff since they are associated with physical processes where timescales often are much smaller than the convective timescales. Therefore, these terms are subject to strong CFL constraints leading to small timesteps and drastically increased computational costs. To avoid this, our strategy is to consider a first order Lie splitting approach in order to split the temporal integration on different subsystems. First, the convective terms have been integrated following a third-order predictor-corrector Hyman {{cite:7755007da5819163fedcbeb3f4c37550d239da12}}. Alternately, it is possible to consider a third-order explicit Runge-Kutta method. Then, the diffusive term (i.e., the Spitzer term) has been integrated by solving a non-Fickian (hyperbolic) diffusion equation as described by {{cite:2f9ce20d921006fe45616a94e6e05041a5a5d5e5}}. Note that, in Ebysus, it is possible to integrate this term following an implicit multi-grid method, as described in {{cite:90175cdc88c3c78c4101677e3d7aaf5a87445ff3}}. Finally, the stiff source terms (see {{formula:ad156784-bebe-4b68-9d6e-5d757a460206}} in Appendix ) are integrated following an Ordinary Differential Equation (ODE) solver based on a fifth-order implicit Runge Kutta method also known as Radau IIA method {{cite:e50c8091e284a78e107b6cb5d4cfbdda0e0ed495}}, {{cite:f6e67832e3ae1f2b965dee9365f87165ab489d0b}}. In order to maintain the solenoidal constraint on the magnetic field, a Hodge projection method has been used {{cite:5c185ae4226728878e59b3a103d74c4ef884b3dc}}.
| m | 24450a134518897655ff4d370d6b24fc |
However, while the results of these recent efforts have been encouraging, we recognize that there are significant differences between natural language and code. Along with syntactic structure, code statements have long range correlations, much larger vocabulary than natural language and reduced robustness to minor changes {{cite:86decc1df658ee32e330b50ee66eab364f3beeb7}}. Previous work attempts to capture such long range correlations by using syntactic structure: data and control-flow graphs of code {{cite:0ad310edbb015b0b286aa4025d6d8b2ae20f6503}}, {{cite:0b5d122803a96e171e74bb3a36411d55b7661590}}, {{cite:e4b197a65ef465e5701c749e47d9626c6a522487}}, abstract syntax trees {{cite:d9799d00c27db7e1e3c6a97559e5cce7399f8458}}, {{cite:614165e5db922534f948d403f7fed2d179e170d4}} etc. It has largely focused on applying such techniques on whole code snippets (code snapshots) like individual methods or classes {{cite:95ea5d373a23fa87d2ed9ebdb6720246de5fd4a4}}, {{cite:7af1d6805e7d4193f78bb5befc9b1085df2bc5de}}.
| i | 1916d466762b6e15ec50d186534f405f |
In this section, we firstly define the model and then illustrate the tools that we use to measure the domain growth.
In the last part of this section, we briefly illustrate these methods by discussing the standard Ising case. For more details on the Monte Carlo method and its variations, we refer the reader e.g. to the monographs {{cite:87ed6c3151f1efd46bb83770188f75bb4f1bc8e3}}, {{cite:d1bf2d55523bc486ec713d12b11718f501fbfaab}}.
| m | 866e17a53b06da1cd5892a9aa2f690e8 |
Weighting KL Divergence Loss:
To understand the effect of weighting KL Divergence loss, we assign different values to {{formula:76129318-8473-4164-bcf7-7f5533c9d533}} such as 0.25, 0.5 and 1.0 and analyze the quality of sketches generated both visually and using the Ske-score. Unlike {{cite:f7771232c7344a84d3303070ac563d93df5dc7ed}}, where a higher {{formula:e6d22ceb-3398-435f-9259-853b072e3a76}} produces images closer to the data manifold, in our case, this behaviour is exactly opposite. Figure REF shows the plot of {{formula:d7df5da2-5a27-4ccd-ba9a-200ab7cdb2be}} for different values of {{formula:cffffca3-b410-41ca-bfe3-fb42e8899ed8}} while training the proposed model on cat sketches with GRU and LSTM respectively as the discriminators. The implication of this on the sketch generation is shown in Figure REF . Visually inspecting these sketches, we conclude that a value of 0.5 is ideal for {{formula:01723771-9fdd-472d-aa3c-a0a57390c618}} .
{{figure:54d7d9e6-df0f-4631-a99b-c88dd0416e8d}} | d | aa3216c726063696c481c8a7da211f76 |
The SED analysis relies on the following distances and photometry. The distances were estimated based on the parallaxes provided in Gaia EDR3 ({{cite:5f5e11c0df24a9ad0670d43798154d2ded92ef61}}). Following {{cite:576f31dc70e76a3bd0670d2387285ab812423a43}}, the inverse of the parallax was used to estimate the distances when the fractional parallax error (fpe) was smaller than 0.1. A Bayesian method with a geometric prior from {{cite:576f31dc70e76a3bd0670d2387285ab812423a43}} was used to derive the distances in case fpe > 0.1, or when the parallax was negative. The distances for the 209 HAeBes in the sample are listed in Col. 4 of Table REF . In addition, we assumed that a low-quality parallax and thus a potentially spurious distance can be obtained when any of the following criteria based on Gaia EDR3 parallaxes, uncertainties and quality flags are met: First, negative parallax or fpe > 1. Second, renormalised unit weight error (RUWE) > 3. Third, astrometric_excess_noise > 0.5 and astrometric_excess_noise sig > 2. Finally, RUWE > 1.4 and ipd_gof_harmonic_amplitude > 0.1. The last three criteria were defined according to Table 4 of {{cite:5f5e11c0df24a9ad0670d43798154d2ded92ef61}} and Sect. 3.2 of {{cite:6c08a20e3ff7d89ef79f36a51882eb9c3d7033cd}}. The 42 objects indicated with the dagger in Table REF have low-quality parallaxes according to the criteria listed above. Although we also derived the stellar and circumstellar properties of these objects, they were not considered in the analysis (Sect. ). The potential spuriousness of Gaia EDR3 astrometric solutions is subject of ongoing work, for which we adopted conservative criteria. For instance, essentially all stars with low-quality parallaxes based on the very recent work by {{cite:3093e3bb9d031d0c46e2e15236c26de3fe9e0d0b}} (not yet submitted, according to the astro-ph notes) are also identified here, although roughly half of the sources with low-quality parallaxes according to our criteria are not classified as such based on {{cite:3093e3bb9d031d0c46e2e15236c26de3fe9e0d0b}}.
| r | 67a5722d8b1b556920001ef64e658ce5 |
{{cite:7a45521ea17b3734103fdec5be757c0f9b2fcdaa}} first considered the necessary conditions on proxy variables to provably recover the underlying causal effect via direct recovery of the hidden confounder. This work was in turn generalized by {{cite:205815a74027b807c34dd63df1ed65858b178540}}. In their work, it is shown that two types of proxy variables are sufficient to recover the true causal effects without explicitly recovering the confounder. One is an outcome-inducing proxy, which correlates with confounders and causes the outcome, and the other is a treatment-inducing proxy which is caused by confounders and correlates with the treatment. In the flight ticket example, we can use the number of views of the ticket reservation page as the outcome-inducing proxy, and the cost of fuel as the treatment-inducing proxy. Given these proxy variables, {{cite:205815a74027b807c34dd63df1ed65858b178540}} show that the true causal effect can be recovered by solving a Fredholm integral equation, which is referred to as the proxy causal learning (PCL) problem.
| i | 26f02070a2d15738cd4c2535582bb76d |
Instance segmentation is a task to delineate the shape of each distinct object in an image. Most instance segmentation techniques are based on the framework of object detection, such as Mask R-CNN {{cite:70e60e07a4ddbd987ccdd370a46f51a64fbb4925}} and its variants {{cite:d88961e58d9f99817ca33a18b67f96d65fb2815a}}, {{cite:b52a16d9efcd5c13f65a58a84403d668b6650e1c}}, {{cite:28a72a050fab855a4b8e9aa7e5108fa8462d5cd0}}. They detect the bounding box of each instance and then classify whether each pixel within the box belongs to the instance or not. However, despite yielding promising results, they often demand high computational costs due to the pixelwise classification.
| i | 60be33be5bb41f78aefae310e23ac51c |
Similar to {{cite:21dd02997a8dd70c774ac63b46abe35f60c4432a}}, we create a prediction task involving predicting {{formula:80916096-b20f-4d23-9113-1c432a5e313c}} of future blocks. Details are provided in the appendix. For multiple views, we define {{formula:7fc78f6c-f5c6-496b-ac01-f9293c0fbb09}} , where {{formula:1664c782-c391-49f0-9163-7362d1ab9417}} , {{formula:89670509-32ae-4908-ad6e-3a1eb3e54c7d}} and {{formula:93e91ce3-b879-4e12-9710-ddf013e50de5}} represent the input, context feature and composite encoder for view {{formula:0cfe2e3f-b41e-4b14-b57e-397b827a09b5}} respectively.
| m | 6fa8fed13c08736cfafc120a8c303aea |
Networks and Datasets We implement the end-to-end training framework in TensorFlow. We differentiate the metasurface optics layer (from phase profile to the target kernel) and integrate the differentiation into the back-propagation process. We evaluate on three networks: AlexNet {{cite:bdc08fa8bc22ffb62320335bbea7761b5822f66a}}, VGG-16 {{cite:62f6708cc3cc275b8cfc5996a764ba0e055b43da}}, and ResNet50 {{cite:984cce9fef4ff7e5c1105fdf2ab8342a4f74161a}}, all on the CIFAR-100 dataset {{cite:9ac04b3c1d0c2484c27847d349ed5e37cb80f515}}.
{{figure:5730d141-bdad-4d80-9eb2-661e31694482}} | m | 519ca5ba987591f44f1b2e46f79c6ab4 |
While a plethora of BH solutions currently exist in general relativity (GR), astrophysical BHs are expected to be described by the Kerr metric {{cite:882356fabbd509718bc8442c0028911e7d71269d}}, {{cite:4ab0cc3bf36cccefa94aae3c3f2955e998d6f9e8}}. The governing equations of an object of negligible, but conserved, rest mass orbiting around a Kerr BH are integrable, due the stationarity, axisymmetry and the existence of a hidden symmetry; the Carter constant {{cite:32787104453576101f85943b88d1535417b1c132}}. Since integrability implies the absence of chaotic phenomena {{cite:b97475f90d053d2b1be2bfb6c2ff5716625ae515}}, {{cite:ae18d4e73574fd6a13218fd7ea949f13c035b381}}, an EMRI involving a Kerr primary (or any other compact object with integrable geodesics) will produce a waveform with slowly drifting fundamental frequencies and harmonics. If, however, the primary's geometry is significantly distorted due to the presence of accretion disks {{cite:84f68ed1a4fc82d1e7e95058d9f06dc590e9cc00}}, {{cite:0179786f4300874b6c9171cac180a957ede0bc67}} (though their influence is expected to be negligible {{cite:2d5d9717c8c57577f199da45468b359eb8bf9be5}}), or GR does not provide an exact description of geometries surrounding compact objects, then the Kerr metric may not adequately describe supermassive BHs. Including slight, or otherwise, modifications to the Kerr description provides a reasonable framework to deal with the aforementioned subtleties with a drawback; integrability may be broken (though see {{cite:da630a8777b0da5642c0d714ad02e24d1c3d93a2}}, {{cite:df690e0135bc7c85af0966a97451e843adffa7da}}, {{cite:70a994e7cc76af0decb70eeb00f8c2eff6dd3a08}}).
| i | 970cedf1de2537fbc3a89e94eedaae98 |
We find that a loop of half length 1 Mm, with 10{{formula:f053e096-b21c-4f78-a0cb-5c6449ae8316}} ergs deposited within 60 s mimics the observed rise times, peak intensity, and approximate duration. The pre-event ambient background heating, designed to match the base intensity level in AIA 193, maintains plasma at temperature of 0.92 MK and density 7.8{{formula:a84d4484-56b4-4f47-abe4-febaf6f48ca1}} cm{{formula:a1ecf70e-57dc-4497-95d3-cae47194f875}} , which is similar to the cooler loops found in active regions {{cite:c8c14343bc0da21aea23542b94828fdbce961a9b}}. The EBTEL simulation is run for 60 s before a triangular heat pulse is applied for a duration {{formula:6345009f-d2f5-44de-a45f-f152e8d81740}} =60 s. We follow the evolution of the plasma during and after this pulse and use the DEMs obtained from the simulation to predict the intensity light curves for all AIA channels.
{{figure:f573765b-3481-46f2-b5d9-c54fbb5e0aeb}}{{figure:fdf87ba1-092b-4f77-ab6c-a900aad71eb3}}{{figure:3a028b8b-912d-4b9f-9ce2-0f5c73bcc871}} | m | 0f8733fcaa1e35d925b5bdd3af7c251e |
where {{formula:688038c6-ef08-40d7-828a-66bfe1ad31ff}} and {{formula:52689349-cfe3-4df3-8d7e-81a0866e067a}} are features from the teacher and student, {{formula:39fb957f-0217-44e9-b6a0-9c34f3f3faa8}} denotes the imitation (objects) mask generated from the bounding box predictions of the teacher, and {{formula:8f148762-fade-46a7-a227-c0a211cb420b}} is the element-wise multiplication operator. Our experiments were in agreement with {{cite:90b5e7552e4499256a525f27daace14957678780}} in that backbone features prove to be more useful. Also, the imitation (objects) masks {{cite:739e176c55ae92e99548ee59757872383c9e8482}} are being applied to the backbone features (not the detector heads). In order to ensure the shapes and dimensions are compatible for matching, an adaptation block needs to be added after the students feature response. We found out in our experiments that a minimal adaptation size achieves the best performance, so a one layer convolution was used for adaptation. This makes sense since only the main student network without the adaptation is later used for validation, so the knowledge of detection should stay in the actual student model and not the adaptation piece. It is also worth noting that Non-Max Suppression (NMS) is applied to the teacher model predictions to reduce the noisiness of the predicted bounding boxes.
In the case of YOLO-based object detectors {{cite:2a70a6f11c35d1221ed0900f20fa9f928b2e429c}}, the prediction loss component is defined as:
{{formula:5e4451c7-9a57-403d-b367-7925ce381a88}}
| m | 7ae77d2cbf4e5292cac1214e2aaefa4b |
During the experiments we observed that the number of linear regions decrease in many training settings: (a) {{formula:af87b3cc-d272-46c7-ab9b-01bd02c0acb9}} regularization (b) using dropout with {{formula:a866212d-db8e-432e-9060-5a9b43841530}} or {{formula:6cbf0ff4-539d-4ee5-ac96-613a9da1678d}} regularization, (c) adversarial training with {{formula:8bb95127-7375-4939-aea8-c341b6012418}} regularization. Furthermore, we observed the largest decrease in the number of linear regions while using adversarial training with {{formula:b5a4facf-94a5-4291-aa9e-06c43e57de7e}} regularization. The decrease in the number of linear regions due to regularization validates the intuition that linear regions is typically considered to be a proxy for network expressiveness {{cite:0472f3b21b0f1963a008c1d21d66338be1a684c8}}, {{cite:d9bf1afade22c845838085527af1ddbac4b03d3c}}, {{cite:c6a8ead6557482741c6aafc5a4130205a83fbcf6}}. In a sense, we restrict the expressiveness of a network when we perform regularization.
| d | d423ae047cfab355a4fcbe5a72437f46 |
(a)
Unlike the {{formula:70d215e7-c5ed-45c8-8303-2b9015685e36}} schemes, which use multiple stages of evaluation of the BVD algorithm, the current approach requires a single-stage evaluation
(b)
The unlimited linear scheme is evaluated through compact reconstruction which has superior dissipation and dispersive properties compared to explicit schemes (see {{cite:6ec0e9d2cbd493e6ed85132dad2b3a9d1638eb78}}, {{cite:24bde3410f0a319e4ef73c3294ec1f16fe7071f4}}, {{cite:0905140a75058d554576c808fcd4393ecf740dd6}}, {{cite:e220313ed89a2f9ed415b45cf82d5a4f1622f32d}}, {{cite:40261b1d6a3caa330bb3c7abac8e19a0c7cb4229}}
(c)
Finally, since the underlying linear scheme is central, which is non-dissipative, the inherent numerical dissipation in Riemann solver is reduced in the smooth regions of the flow.
| i | beade1a210600ab84263d891b217afde |
Our goal is to learn an embedding space that groups image patches containing the same part of an object.
This is motivated by the hypothesis that object part representations are more general than object representations, as parts can be recombined in multiple ways to different objects.
As an example, a wheel representation can be combined to a car representation but also a bus representation.
Therefore, object part representations should transfer better across datasets.
For that, we aim to design a pretext task that allows for intra-image category learning on an image-patch-level.
Thus, a clustering pretext task is a natural choice.
As shown in Figure REF , we retrieve patch-level optimal cluster assignments from a teacher network and predict them from the student network.
The choice of a clustering pretext task is further supported by empirical evidence showing that clustering pretext outperforms contrastive pretext for dense prediction tasks {{cite:68ee6daa182823465a03cd4823f8b6f5378549f8}}, {{cite:0b7b990c9cc89db7aa421c7398073b5f3a694303}}.
Instead of pretraining models from scratch which requires substantial GPU budgets, we use our loss to fine-tune pretrained neural networks.
Further, this circumvents known cluster stability issues and clusters capturing low-level image features when applied to a patch-level as reported in {{cite:b42f68159ccacf992facb3be2f3db59e85dcaaf3}}.
| m | 26b79343ac61261c2ca421085a8e316a |
As we know,
the Fourier spectral method makes use of
the eigenfunctions of the Laplace operator which are orthogonal to each other with respect to the Sobolev inner product involving derivatives, thus the corresponding algebraic system is diagonal {{cite:217c6bfcb0ae575616aa95b05e704f89ff633a75}}, {{cite:f5a6dc6f94dc05ebfd24477cfa2b6d047a038d66}}, {{cite:9e23aef03c833e5d03be4469c5bfee904faf5ea2}}.
This fact together with the availability of the fast Fourier transform (FFT) makes
the Fourier spectral method be an ideal approximation approach for differential equations
with periodic boundary conditions.
Although the utilization of the genuine orthogonal polynomials/functions in this direct approach usually leads to a highly sparse (e.g., tri-diagonal, penta-diagonal) and well-conditioned algebraic system, however, in many cases, people still want to get
a set of Fourier-like basis functions for a fully diagonalized algebraic system {{cite:fe286049a042793cb2e17341b109e16bda517635}}.
| i | 552d125b81e635b52657c28866df2898 |
Discrete exterior calculus (DEC) constitutes a discrete realization
of the exterior differential forms, and therefore, the right
framework in which to develop a discretization for differential
equations not just on flat space but on manifold {{cite:16433c931cb0d2c8fb0692b325a1f54951efdd18}}, {{cite:fc8ea6083bf2a150f760244b94b60ee116db431e}}, {{cite:c580a4af1f4837b447e8c00e2b159a501d0a4487}}, {{cite:3b3ca92e8b211db7b868bb41e6741115a8441e77}}, {{cite:3dc248a603e719ba855f56f6a0df761060fff14b}}, {{cite:b511d11f01297ae4a2f980e36505eb42638cca3b}}, {{cite:099ca79ce9d8b115560d22c1f72115febb271b4e}}, {{cite:3ca808ea3a678a15558decdc3e897377c757a96b}}, {{cite:e51fa0d2d696961fdd42dcf462c25238e99d3df7}}. The
differential operators such as gradient, divergence, and Laplace
operator
on manifold can also be naturally discretized using DEC.
The numerical solution of
wave equation on space manifold and the time by the methods of DEC
is obtained in this paper. For
this equation, an explicit scheme is derived. The analysis of this
scheme's stability shows that the numerical solution becomes
unstable unless the time step is restricted.
| i | 0e39f0ba8114a3f3703a5007f84bbd57 |
The convexity of {{formula:1377cde2-d8ef-4c78-830e-db4a35d75eb2}} ensures the Taylor expansion of {{formula:0278fe0f-ea4e-468b-a8ab-d1a9eac0bc8d}} at any iterate.
The Lipschitz continuity of {{formula:655daeae-8a40-482f-9d38-c1bf1775d659}} relaxes the thrice continuous differentiability of {{formula:67bdaf0c-8fb3-40fa-a298-5812313047d1}} , as assumed in {{cite:951874e7ef793e9378a5c88dbddcb1cf592f6936}}.
Note that imposing conditions on {{formula:d8d5278a-d869-422b-b3b8-617138f2450f}} is equivalent to imposing the same conditions on its components {{formula:07e9a584-e2b5-4fba-8203-1587807ea72d}} and {{formula:9f644704-9b73-4976-9a51-8f5b6d69d23c}} , although the Lipschitz constant may be different.
The Lipschitz continuity of {{formula:2da4efa9-4caf-462e-909c-4aa3636be24e}} is implied by the compactness of the set {{formula:efa2b541-3872-4db4-bc46-cf27318cf67d}} , while we express it out explicitly with constant {{formula:1f4fe838-c507-453d-83f8-0a0f1b8b0168}} . With the above setup and a simple calculation, we know from (REF ) that {{formula:6748cc4a-0076-450b-b835-2d8a0e9d4c1b}} is also Lipschitz continuous in {{formula:09cd27fa-43ad-4300-945c-52e9ce93cc2f}} .
That is, for any {{formula:3fa61725-181f-46cd-a0cc-7f061eb6dea1}} ,
{{formula:386c1e11-b1a0-4d38-ba36-a9606d496044}}
| r | a7b7ad19164284088c04096295e669be |
- We propose a novel adversarially robust FSS framework, PNODE, that can handle different kinds of adversarial attacks like FGSM {{cite:a346602f18d6aa45128b6a2de3f20cea73af7c36}}, PGD {{cite:cbcff34c5ab3e7e9f9f40c31f0c44abd6909eeae}} and SMIA {{cite:0697b8ea5513a5751abac81f0eebe22d530d7a47}} differing in intensity and design, even without an expensive adversarial training procedure.
| i | 2bc6f53a0ab7d86d6800dcc0926f0eb3 |
Larger pre-training scale improves intra-domain transfer. We obtain evidence that either natural-natural domain transfer or medical-medical domain transfer can be improved when increasing model and data size during pre-training (Tabs. REF , REF ). This improvement is strongly evident for natural-natural transfer scenario. There, either increasing data size by using ImageNet-21k instead of ImageNet-1k or increasing network model size by using ResNet-152x4 instead of smaller ResNet-50x1 creates strong, consistent boost in transfer performance across all natural target datasets, which is in line with previous observations {{cite:d6a1486d3bd4dd7f68376c4d473a4dc23ea0db55}}, {{cite:e4a760ea2b86b987b6a22e7a211eeb89c1fbe402}}. Transfer improvement is strongly pronounced in few-shot regime (e.g, Figs. REF , REF , Suppl. Figs. REF , REF ), also adding evidence for more data efficient transfer due to larger scale. The situation is less clear for medical-medical transfer scenario (Tab. REF ). There, transfer improvement due to larger pre-training scale is given for full transfer, without clear benefits for few-shot transfer (Figs. REF , Suppl. Figs. REF , REF ) that is so strongly boosted in natural-natural scenario. This hints that nature of dataset domain itself may play an important role in determining how transfer improvement will be affected by pre-training scale.
| d | 0c672f25bd777336ed926c0f8db66035 |
where the first two terms are referred to as one-electron integrals and represent the kinetic energy due to the electrons and the potential energy contributions from electron-nuclei interactions.
The remaining terms are two-electron integrals that describe the potential energy arising from electron-electron interactions and are called Coulomb and exchange integrals.
Using Lagrange multipliers, one can express the Hartree–Fock equation in a compact matrix form, the so-called Roothan–Hall equations,,s,,{{cite:e389cd110324745052867e324bf89b0a13735ae8}}, {{cite:50bed951212eeea5c77345e5e35bd71a17f7fdc9}}, {{cite:f53475cb3a0edbf0b4754005031e5a2db2330b8e}} which allow for an efficient solution:
{{formula:86652ed4-ba7d-4ccd-a644-987504862c3b}}
| m | aae159946e9fe53ef1166c9d1dfc52ef |
An interesting extension of this work is to consider a Bayesian model with various priors on the coefficients of interest. {{cite:d501692642d27e281cfc4780e58342260d8160fc}} consider asymptotic properties of Bayes risk with the horseshoe prior. We believe, investigating through different prior beliefs can result in increased predictive performance of deep partial least squares. Another useful extension is to draw applications to eminent domain with judge characteristics as instruments ({{cite:557ac44ae2059d065d6e97f25afae4b5a0bae1d4}}). In addition to prediction problems, consistency of DPLS allows us to address estimation of the treatment effect. In future, we plan to demonstrate precision of the treatment effect and compare it to other benchmark algorithms.
| d | 27020ef0171fb6cd4986510513e3baef |
The examples we deal with are within the context of two-dimensional Jackiw-Teitelboim (JT) gravity {{cite:8e7fdfaa519243f4d148f0d103419a93a874fd18}}, {{cite:83e198e898eaf8774a81aa9bc83bab454ff1310a}}, {{cite:d20fa0c9288113ebb3437b403ed2b43bbb040245}}, {{cite:ad0e1d2015ed0ceac6f4b3a0e741e3e49ff187ab}}, {{cite:ffcca85ed8a658c26f51c1a6cf0b3a628e461fde}}, {{cite:3ae35ee95a575dc6cc84c4481d6ae2f57ed9c22c}}.
In two of the models, we take two pure-state black holes in two different asymptotically {{formula:2fbc98e6-ac16-4c9d-b3f0-f4e3cf497b01}} spacetimes and then entangle them.
Each pure state is one of the somewhat atypical Kourkoulou-Maldacena {{cite:f1dc68467d092c280cb12d4680e44a7a2b94280f}} states with end-of-the-world (ETW) branes behind the horizon, obtained by a projection acting on one half of a thermofield double.
In the first model, the doubled Penington-Shenker-Stanford-Yang (dPSSY) {{cite:c838fb4c4cb0c963380605f34512a925b03e7e5d}} model, we allow the ETW brane in both black hole interiors to have multiple flavours and entangle this flavour degree of freedom between the two black holes.
In the second model, we consider the JT gravity coupled to a 2d CFT in the bulk, following {{cite:2dafc5f751949e63bef9d040bd80c94b41dd0752}}.
At time {{formula:15918b55-911d-4b26-a60a-f497fe41e3b7}} we couple the two spacetimes (each with a black hole) by imposing transparent boundary conditions between them at the asymptotic boundaries, and then evolve for a long time.
The third model is similar to the second, except that the two black holes before the coupling are taken to be in a thermofield double state.
| i | a793ac0300ace1fdd40f38df33d90bf8 |
+ Shake-Shake {{cite:3d5a583f63b7d72de1bc93354b3a81d8131add00}} {{formula:05ac061d-3dd6-43be-911d-7e3b5b115d74}} {{formula:73b754ab-701b-4ad2-935f-9ac2e762a87c}} - -
| r | 4c79ff12d8b80fce83d1ff291b199067 |
Multi-path choreographic solutions provide such a class. When the single-particle system is non-integrable and has an elliptic periodic orbit, it has, typically, infinitely many elliptic orbits (e.g. around an elliptic periodic orbit there are typically many resonant elliptic orbits, around which there are secondary resonances, and so on {{cite:05297423695bedb587b6a0720527b5c7be3bc9a6}}). Similarly, near a homoclinic loop to a saddle center of a non-integrable system many stable periodic orbits co-exist {{cite:9d3b54ad0aff126c6c98e1c99936a3577bf76add}}, {{cite:1441ecd38e5d17989e47aaf50ffc246244a62b24}}. When the corresponding periodic paths in the configuration space do not intersect, one can obtain KAM-stable motions of repelling particles along several such paths: the particles on the same path must have the same frequency to avoid collisions, but the particles on different paths may have different frequencies. When the paths do intersect, one needs the frequencies of the motion along the different paths to be in resonance. This condition is not actually restrictive: given any finite set of elliptic orbits with arbitrary periods, one can tune the partial energies such that for every two paths the ratio of the periods would become rational. For particles in the billiard this is done just by normalizing the motion speed for each path by the path length.
| d | f9db888db16632fbc2d5eb3255430525 |
We conducted a series of experiments on two well-known datasets, namely MOT17 ({{cite:ae9029d1b33f7b1de7ed8cba35c95e0f08b142af}}) and UA-DETRAC ({{cite:84bf593f2ffa3e41f0f79575350dcee339b56873}}). TADN outperformed the state of the art, delivering superior MOTA performance in all cases, proving to be a viable alternative to typical computationally expensive data association techniques. In this section, first we present the two benchmarks we evaluated our method on followed by our implementation details. Next, we present our results along with a qualitative analysis including particular cases of interest regarding the strengths and weaknesses of our method. Last, an extensive ablation study follows investigating and comparing the performance of our method for a curated selection of various configurations.
| r | 3fea92573e36d42b19d5f22954653407 |
As mentioned in Section , another method – APPNP – can also provably prevents over-smoothing {{cite:54460d3995662d9da118a49be44af5791b7887c2}}. The authors of this study use the fact that the PPR propagation will converge to {{formula:726685b4-40c0-4e1b-95b6-a326f06f9f10}} , where {{formula:4ca0fcad-02fd-4442-a7b7-b5093e8dd019}} is independent on the node label information provided in the training data. Each row of {{formula:a374dffc-aaa6-413a-af4e-eb3446043a97}} still depends on {{formula:e5b1a2f6-211c-413b-a069-da73daf06fa9}} and thus APPNP will not suffer from the over-smoothing effect. However, since {{formula:211f29f9-e6ad-42f9-88ee-a9b03307e09e}} is independent of the label information, it can cause undesired consequences that we discuss in what follows.
{{figure:c53480f7-27c9-4a5c-8efc-4ff858d9016e}} | d | 07f23c05d083b1e9880a7de90d7d1d4e |
As we known, although the quenched quark models have obtained lots of success in the last decades, the effects of the sea quarks and gluons interactions are not taken into account. The unquenched models, considering all kinds of additional effects, have been developed, and widely used to describe the hadron spectra, such as the coupled channel model {{cite:f33079c2d9598fedd5d7d9851474330c90b2b0fb}}, {{cite:9c4b3146ca4b5f428f5b80beeaf7d86bd1df7d63}}, {{cite:3b256672492e02d4c1cbf1b31394be3058a7cd21}}, {{cite:1407178e1d546403bb4405bb2c9eac5091451ec1}}, {{cite:7f58f75c115275cba323d8186c8223bc3c11632e}} and the screened potential model {{cite:6bad528bcd552d6b8a8649cc8b2427411f1af36d}}, {{cite:4ec6f89ae7f6c824b1d40da843791ded7eff4165}}, {{cite:3b2e824151f5b1e81a6893578954a97d74c9275b}}, {{cite:951da95b13045d5cb87b03183ce149ac00f76c75}}, {{cite:66d752747d12a8578c4b839b86f7890ffee0393a}}, {{cite:75ecd232e01cf0e74524fd71802617603b2847b2}}. In the quenched potential model, the potentials mainly contain a coulomb term at short distances and the linear confining interaction at large distances. However this is not appreciate in the large mass range, since the linear potential, which is expected to be dominant in large mass region, will be
screened or softened by the vacuum polarization effects of dynamical fermions {{cite:ad8cd4926c0ad67e0ed5f140ed2b008252d95101}}, {{cite:382dc7010bcab9243fa00c2d6357964b6039e46f}}, i.e., the unquenched effects reflecting the sea quarks or gluons contributions to some extent.
Clearly, the unquenched effects can lead to important influence for higher radial and orbital excited hadrons, which means that the predicted masses of the higher excited states will be smaller than the ones of the general liner potential models. Comparing with the coupled channel model, the screened potential model is simpler, and have been successfully used to describe the spectra of the charmed-strange meson {{cite:4ec6f89ae7f6c824b1d40da843791ded7eff4165}}, charm meson {{cite:6bad528bcd552d6b8a8649cc8b2427411f1af36d}}, excited {{formula:df896e2c-a423-462e-8aba-bae20e356d51}} mesons {{cite:b977f119a237e2c2fa78a9fff9175078b2e6a00f}}, {{cite:a57040ce79e797f5636473a3090d27769076ae1b}}, charmonium {{cite:951da95b13045d5cb87b03183ce149ac00f76c75}}, {{cite:382dc7010bcab9243fa00c2d6357964b6039e46f}}, {{cite:66d752747d12a8578c4b839b86f7890ffee0393a}}, and bottomonium {{cite:3b2e824151f5b1e81a6893578954a97d74c9275b}}, {{cite:38af636f402478b7f8b0284702429af49d973341}}.
{{table:f696efd0-1614-450a-936d-a6af7017ba86}} | i | c37f7c43df83864a14a17b29f1f4ad76 |
Although modern language models are reasonably effective in deciding whether a short text (premise) infers or contradicts a claim (hypothesis), their performances fall when the size of the premise increases. Therefore, effective retrieval of related sentences from a reference corpus is crucial for accurate verification. Recent approaches on claim verification are indeed based on a retrieve-and-verify paradigm where given a claim, text passages related to the claim are first retrieved, then a classifier is employed to decide whether the retrieved text supports, refutes or is neutral about the claim.
Thorne et al. {{cite:8a400b09a4113d48d671ead0514eef7214eaf80c}} describe a general framework for verifying claims over a large corpus of text documents consisting in three steps. The document retrieval step extracts from the corpus a set of documents related to the claim, which are likely to contain the evidence for or against the claim. Then a sentence selection step performs a refinement by identifying in the retrieved documents a restricted set of sentences related to the claim, which we call the evidence. The last step, claim verification, decides whether the claim or its negation can be inferred by the selected set of sentences and classifies the claim into one of the labels supported, refuted or not enough info. Recently proposed claim verification methods are mostly based on this framework {{cite:e4a8d2e2394221e5868aa14a8c8e71b3305d9ead}}, {{cite:161ac0db96f73316c9e4d3b4162415c940aaa321}}, {{cite:46cd4cb8bbd2d096e11841e653f30da22153077c}}, {{cite:a99f0312829b7944dbfef645b66e046bfc7f6ee1}}, {{cite:4b13328d0655c31e4512990f0a6046ce39dcc27c}}, {{cite:bf8652fe092b1b99288bf818b6e38ae862f75f8b}}, {{cite:df701fa09bdbeced83c28599db54fa6785d7cfac}}, {{cite:b470cd35619e4ba77271312a50a7f1deb58b4e13}}.
| i | 45232d1e8fbab628ef2ac65809c229a1 |
In the case of adversarial training we also run a single image analysis to detect the frequencies useful for adversarial robustness and those responsible for adversarial weakness. Here too, masks learned on adversarially trained networks concentrate more towards lower frequencies compared to those learned on vanilla networks. Furthermore, the analysis showed that only a sparse, class-specific set of frequencies is needed to classify an image. Surprisingly, mask-filtered images in this case are not recognizable and resemble texture-like patterns, supporting the idea that ANNs use fundamentally different classification strategies from humans to achieve robust generalization [{{cite:a40dca960bf5c5e003a0be24751bc9b240369fea}}].
| d | 3b9f9c50cd0245cb5acf4b0169bee1d4 |
Here, we circumvent this second step:
instead, we enhance the relative nerve functor to write down the functor {{formula:29583f08-791a-4473-b79c-23318af6e360}} , which takes its values directly in the {{formula:68335727-ea9e-499b-bee2-b8bdaa501289}} -category {{formula:6fd9414b-83bc-462b-b6d6-f7138cdfc630}} , and prove that this already exhibits the {{formula:947e02ec-393b-48b8-a989-8543b4383666}} -categorical localisation (like {{cite:c385b71082c8e1d054428b099e51c95da34d1818}}, we avoid the homotopy coherent nerve {{cite:f3c521ef0b05dd118473aaaf81a5e5d737c3c516}}, {{cite:a9ad80d696f49453a135d5902f65c57108a16836}} in doing so).
At the same time, we obtain a framework which unifies three perspectives on {{formula:15d852ca-f2e0-45c5-9971-e0b7dfa5b9b6}} -valued functors:
they can be presented as strict {{formula:e64dce9c-5e19-4bd8-8e80-a59937e7d394}} -valued diagrams, or as left fibrations over {{formula:1a184ee5-a86d-4c56-b5da-1f80ad6538d5}} , or they can be viewed directly as objects in {{formula:f64d3c24-4156-48a0-b74d-57ad87ecadd0}} .
Furthermore, we make manifest the full homotopical meaning of the rectification of simplicial diagrams {{formula:1cd7140a-1db1-4de9-ba76-cd3367a014fc}} from {{cite:c385b71082c8e1d054428b099e51c95da34d1818}}:
it is really the left fibration presented by the coherent diagram of spaces {{formula:dd2490ea-c9b8-4c19-a683-714f8d796d8b}} described by {{formula:a60b2a2c-33a2-411f-9bef-d8934eb35421}} .
| i | d733bd1c629fbc4fed710b0375439dce |
Our proposed SecretGen works under a wide range of scenarios regarding different types of prior knowledge.
See Table REF for the scenarios under which SecretGen and existing methods can be applied.
Although EMI theoretically works in {{formula:b2612e3a-9669-44ff-9810-c4ab8f53b44d}} cases with ground truth labels, its performance and efficiency dramatically suffer against deep models.
Although SecretGen still requires non-sensitive private data as prior knowledge, such assumption is realistic as image corruption is often leveraged for privacy protection by individuals {{cite:04168fe2d3ef28d622684edad3dace99230a6a3b}}.
Furthermore, if the knowledge of ground truth labels is available, it can be incorporated into SecretGen conveniently.
More details are deferred to the appendix.
| d | 6ef2018fb39d407bfe0efc25a249252d |
where {{formula:a8fd42fa-1f2b-465a-a9c5-8bef1ae10c8c}} is a vector of {{formula:2fb76d77-be49-4794-b3fd-c3e4c95d2aea}} latent factors, {{formula:6d8a7da6-d86a-444b-bdb8-7532303d0fa4}} is a matrix of factor loadings and {{formula:3da97d9c-45a9-46dc-ba81-e982295226da}} is a vector of idiosyncratic errors. The matrix {{formula:91a9249d-e191-47c5-84c4-14eba0883098}} can be estimated using the first {{formula:8a731ee2-e7ec-4ffd-83b1-c1ae578854c4}} eigenvectors in the eigendecomposition of the (rescaled) genetic covariance matrix {{formula:6ad16657-a2cc-41e7-9695-8b2d5884e154}} . This covariance matrix and its eigendecomposition can be approximated if one has access to a reference dataset. The number {{formula:77adcabc-0d0b-4949-badf-7b1b57619348}} of latent factors can be selected either empirically, as is common in factor analysis, or in a more disciplined way, for example using the factor penalization method of {{cite:73ae7ea746f37cb210f7da240b329088f378af75}}.
| m | f3ea62c18bf2ad613c13bf6c76662b99 |
With the recent advances in the low Earth orbit (LEO) satellites industry, a wide range of applications are anticipated to benefit from more than 4,700 LEO satellites launched in space {{cite:fa2bdfa4e76b6315c2e9fdb0a6613f24fcc9ca8e}}. One main application is providing seamless global coverage and low-latency ultra-distance communication {{cite:c0d80c0e0554154920fadebade7f62a638be4e69}}. As a powerful mathematical tool, stochastic geometry is one of the few modeling methods that can provide analytical results for crucial performance metrics of satellite networks such as interference power, coverage probability, and latency {{cite:653c92dda772f44c09e56d1353a1d56463877f4d}}, {{cite:f834f425c2646a1a5dd664073f501e648e62c78f}}. Some literature has studied the coverage probability of satellite networks by modeling LEO satellite locations as Poisson point process (PPP) {{cite:7c00816cabe03e2a11f1ffbe5025a7f3b5c57170}}, {{cite:fdba8a242bce63d4ac1590123a7d83511cc37869}}, binomial point process (BPP) {{cite:5a11fc5fbc4762214fabdf956401f5bf8b8816de}}, {{cite:65944f1757b1045436d39256d444739ad749df9e}}, {{cite:2d30b682f54e017856c88a60a03d90bb1df57d42}} and non-homogeneous Poisson point process (NPPP) {{cite:a1105edcfcf381235142fca75bb68695846e8775}}. Although stochastic geometry facilitate the analysis of the performance of satellite networks, the satellites rotate along fixed orbits in reality rather than follow the point process distribution on the entire spherical surface. Because of the difference between the stochastic geometry-based model and the actual orbit model, whether the stochastic geometry method is suitable for satellite network performance analysis has been the focus of discussion.
| i | c5eecfd994ef05065b1a0a29e4f210a1 |
In addition, the overlapping area could also include some stars originated from within the Milky Way disk.
{{cite:d7b78cbe7bce12d0a033d94a0eec43143cee8bbb}} proposed two disk heating mechanisms to form the metal-rich halo stars: runaway stars
and radial migration. Runaway stars are young stars that are formed
in the disk and then ejected from their birthplace.
Based on the discussion of metallicity and spectral type in {{cite:d7b78cbe7bce12d0a033d94a0eec43143cee8bbb}},
combined with the characteristics of our sample, we can conclude that runaway stars are a minor component of
the observed metal-rich stellar halo.
| d | 43f40880065c65b09060c83ee7143953 |
Theorem 1 (Talagrand Inequality) {{cite:64e28a55e268e3eebb497ced7998df25d876b62f}}
Let {{formula:c048a021-0241-41f2-a455-376a9b4a8c56}} be a reference probability measure with density {{formula:a1c422af-099d-4504-b6a6-7c3fb54ee4e8}} . We say {{formula:7335da9d-dfc7-4f0c-87a2-867e830f49db}} satisfies {{formula:2c464273-de72-4133-9882-73086b52b8fc}} , i.e., Talagrand inequality with parameter {{formula:00e71d33-4804-4a56-985b-78ebd0f6f63f}} , if for any {{formula:fd0555dd-0f35-4bb8-8171-5aa3baa056e1}} ,
{{formula:44f73bd0-2f5e-4139-b2a6-f3c32b881931}}
| r | cbf7867ebfa67fa47273aa2c6c6385ff |
In this work, we only calculate the accretion mode transition condition for an ADAF with magnetic outflows at the transition radius. In principle, the critical mass accretion rate for an ADAF with magnetic outflows given by Equation (REF ) is only valid at {{formula:7e2853d9-e4a4-4826-903c-6a4c337c5f2b}} . However, the field advection is very efficient in the ADAF, i.e., the field threading the ADAF at {{formula:2e7a8f83-16f9-4d91-94d1-c0a1adada20a}} will be dragged inwards, and field strength will be substantially amplified in the central part of the flow {{cite:4d05c2175a33c65ed487032eea9ac18c0f25b19d}}. It is found that the value of {{formula:f67a55c4-cfa3-4de2-a0b9-da4f098e843b}} increases radially in the ADAF {{cite:4d05c2175a33c65ed487032eea9ac18c0f25b19d}}, which implies that the critical mass accretion rate of an ADAF with magnetic outflows will increase with decreasing radius (see Equations REF and REF ). The ADAF is shrinking with the rising of the mass feeding rate of the outer disk. We note that the field advection was calculated only for an ADAF without outflows in {{cite:4d05c2175a33c65ed487032eea9ac18c0f25b19d}}. If the angular momentum carried away by the outflows from the ADAF is properly considered, the field will be more enhanced than an ADAF without outflows, because the radial velocity of the ADAF with outflows is higher than its counterpart without flows. It implies that the whole ADAF can survive if the mass accretion rate is lower than the critical value derived at the transition radius. Our model calculation of the critical mass accretion rate is valid in general for an ADAF with magnetic outflows. The calculation of the global structure of such an ADAF with magnetically driven outflows is desired, and it may help understanding the detailed physics of the hard to soft state transition in XRBs, which is beyond the scope of this work.
{{figure:5d5e5ffc-f2ef-485a-aa23-8f2cd2c1813b}}{{figure:cefb996b-11f5-4b57-aa9b-50a74e78c3de}}{{figure:6c1795a5-c645-4641-911a-8613f5c95a5f}} | d | 627e810fa9c321e9bbec15a23d56d5ca |
Lemma A.14 {{cite:fa58404ed7249615294fcf6921f19ad39c8425bc}}[Theorem 6.8]
Weak convergence in {{formula:aa6e2827-878b-441a-8467-ea1d203cef52}} is equivalent to convergence with respect to {{formula:93262836-ffa2-4497-9a14-98d9c3885861}} , {{formula:3cd07b56-eae3-4dd6-b1e3-08682d65b1aa}} is a complete and separable metric space.
| r | b9e5854a5536a801480d5fae8aafaf9b |
Proposition 2.1 ({{cite:70127fbef6c85b00c1d0de36648081ede04272a5}})
Let {{formula:f52636b8-b083-472a-a6f0-d51231c89118}} be three {{formula:c6800056-6782-4b1d-a226-76d5e4e14685}} matrices and {{formula:36e01e35-db5c-4dcc-80f5-cc60ca785231}} be two {{formula:5136aa91-50cf-4057-a1c8-80f15d8b7821}} diagonal matrices. Then
{{formula:534aa90b-0758-4b55-9284-925ce6e397b6}}
| r | f6afa49c2c20e5ef2c5840e6964bf16a |
Among all the machine learning frameworks, reinforcement learning (RL) has an unique advantage that the interactive training process is aligned with human decision making. During the past years, two basic optimization ways, Q-learning and Actor-Critic have been developed to solve many RL problems, as well as the combination of the two. Take financial applications for example, Pastore et al. {{cite:657563b5be0c02cc6fef550974ca09886f872271}} analyze data on 46 players from online games in financial markets and test whether the Q-Learning could capture these players' behaviours based on a riskiness measure. Their results indicate that not all players are short-sighted, which contradicts with the naive-investor hypothesis. Besides, Li et al. {{cite:e3c2cc08610bfef8d48f54f0b823adac90e120b8}} study the benefits of the three different classical deep RL models DQN {{cite:9d608c89f5222d5bd57f9eaf0807125cbfd20a4c}}, Double DQN {{cite:1c7187d61bffb262ba0b3c525266932d35c80ffa}} and Dueling DQN {{cite:de56052637c7c16a5b39e56e9d00172e00ff392c}} by predicting the stock price and conclude that DQN has the best performance. In addition, some researchers simulate and compare the improved deep RL method with the Adaboost algorithm and propose a hybrid solution. Interestingly, Lee et al. (2019) {{cite:9bf8ff4bfa58c390c3f134b2860788f9d5a673ed}} apply CNN to a deep Q-network which takes stock prices and volumes chart images as input to predict global stock markets. On the policy gradient side, Kang et al. (2018) {{cite:818882d9474038a47d19b75cbd9e0cfb9999760d}} utilize the state-of-art Asynchronous Advantage Actor-Critic algorithm (A3C {{cite:5428af0f9c507e392edc599a7d2e4fb803c86091}}) to solve the portfolio management problem and design a standalone deep RL model. Moreover, Li et al. {{cite:ce43d77a889d01d4b94d9399b72ce8b37f64178c}} propose an adaptive deep deterministic RL method (Adaptive DDPG) for some portfolio allocation task, which incorporates optimistic or pessimistic deep RL algorithm based on the influence of prediction error. Through analyzing data from Dow Jones 30 component stocks, the trading strategy outperforms traditional DDPG method {{cite:8fab83a0743feee6735c3bd6017fbb7aca72f762}}. Sutta {{cite:fe0677725d36d43f3c412842fc472857bd9dfd73}} compares the performance of an AI agent to the results of the buy-and-hold strategy and the expert trader by testing 15 years' forex data market with a paired t-Test. The findings show that AI can beat the buy {{formula:abe61f16-41b6-48e0-8e92-06526ca331d6}} hold strategy and commodity trading advisor in FOREX for both EURUSD and USDJPY currency pairs.
| i | 2a560c4b643432d76fd5f79ed879f61e |
One of the themes that has been developed in different researches,
is the construction of wormholes with matter sources that does not
violate the null energy condition, but as we know the effective
stress-energy tensor in fact violate the null energy condition
{{cite:1d73e44c925ac9b860bb48733abab09f067790f8}}. In the {{formula:012ddd6b-bac6-47e8-8b56-bb479ad2d3b9}} theory, have been built wormholes in
which the fluid considered, anisotropic, does not violate the energy
conditions WEC and NEC, this is possible due the election of {{formula:56f57d94-5053-4e94-887d-f09b476a9ba0}}
and the state equation {{cite:bb7bfe5c9c98a81bf995e186fd8644b7b0f84945}} or the Einstein–Cartan
theory {{cite:f9856d7fa9bb710c04b79c8d1726114dbd3600f5}} between others.
| i | e8440e6b3a5b139989e744fa66854c63 |
The experiment was implemented in Python leveraging the symbolic framework CasADi {{cite:496c6e4a8b2dc1d19050a402473add2fa6cba23c}} and its interface to the IPOPT solver {{cite:b524a171e1c9f1aecd1df666ea7bb407ce2afb48}}. The source code and simulation results can be found in the following repository: https://github.com/ FilippoAiraldi/learning-safety-in-mpc-based-rl.
| r | 282a72a2f5ec5e8e866eacef69898c54 |
The WDL function satisfies the following condition, which can be proved using the same method as shown in Refs. {{cite:9a26dea764566611cb487814c579b5bb6584af05}}, {{cite:63218ccf4c8742db5f64eaabf63804e97b5b330c}}:
{{formula:6e0c64b0-1e5f-4f8a-97d5-0349909e77dd}}
| m | 1b47231643e9782731c1608227689a3a |
Furthermore, the standard disk is thought to be more suitable for the formation of powerful jets in the magnetized accretion ejection structure (e.g., {{cite:d1ac9e08f5ceae9bccce916935d13fc79f0b900f}}), because the radial magnetic tension overcomes the toroidal one at the disk surface when the disk becomes too thick. {{cite:ff330caa3c59ee56e28551d7936fdbec7d4dd048}} have advanced these works to propose a jet-emitting disk model, explaining the canonical spectral states of black hole X-ray binaries. In this framework, spectral energy distributions of the jet-emitting disk model have been investigated ({{cite:bb2fcad96796656f12a6fe43af4fc951b4dc6b9b}}).
| d | 420d6ce45162dd71ce16ee4e1750452c |
This study also sheds light on a new type of sensitivity of MRI reconstruction algorithms. At present there is growing interest in identifying sensitivities of such algorithms {{cite:b7b84f5c86a7e83c1c4ac4d781b95692efba061d}}, {{cite:edb34ee08cd81df21961c76614322d1c3fd9bc27}}, {{cite:4d52ffc00d5b19383098cb3506ac0b2eb2e56569}}, {{cite:03272b40c6b15d04e0c025397287063254d70913}}, {{cite:a989cb7668cf168b693d52f09501f7ecf86be7c2}}, {{cite:19cbe6e59a0cc91ec006249b314fa1c812ab51f4}}. However,
recent studies focused mainly on investigating sensitivities with respect to adversarial attacks. While these attacks are an important research tool, they are not observed in practice since MRI scanners are closed systems. Here, on the other hand, we focused on sensitivity related to a more common cause: off-label usage of public databases. While reviewing papers, we noticed that such usage is becoming increasingly more common due to the growing availability of public databases that offer various types of MRI data. Subtle inverse crime I may be common since MR images found in public databases are often based on images produced by commercial scanners, where the preprocessing pipeline described in Figure REF a is often applied by default. Additionally, Subtle inverse crime II may be common since JPEG images are highly prevalent; 73.3% of the Internet websites contain JPEG-format data {{cite:75d85fc4fd517631ae139d927887a53bc02e5870}}. These factors suggest that the subtle inverse crimes might be more common that intuitively expected.
| d | 06276ff0ca7a6a5547bf2ec838dfc169 |
Grasp Representation.
The autonomous visual grasping tasks generally start from collecting visual images of the object by sensory input, which will then be processed to generate an effective grasp configuration to maximise the probability of grasp success. Considering a parallel-plate gripper, the grasp representation {{formula:613c0503-b5c0-4b79-b381-1bf4b89278f9}} {{cite:5f6c029d4443f1f5a998701d33159a8dd86ebfb7}} is formulated as a 5-tuple:
{{formula:320d91fb-fe32-4486-88ad-9f8da0265dba}}
| m | fe712bfcb60921ef556c58949fe215ea |
A real-life cybersecurity application is the actual platform to use the CyberLearning model that typically examines the behavior of the network, finding the security patterns for profiling the normal behavior, and thus detects the anomalies or associated attacks. Although an ANN model has its hidden layers for computing, it also affects on the significance of the features. For instance, an ANN model with the selected security features gives significant accuracy for detecting anomalies and multi-attacks, as discussed briefly in Section . Although we use the security datasets UNSW-NB15 {{cite:b4d4e05aa20f3aaf36b083eeb351175a55308db2}} and NSL-KDD {{cite:3520ed8df7d2b81ba39004e291f8ddaddfa1c3e2}} while building the security model, our analysis is also applicable to other application domains in the area of cybersecurity, including IoT security. Several deep learning networks such as Convolutional neural network (CNN), recurrent neural network (RNN), Long Short-Term Memory (LSTM), deep belief network (DBN), or an autoencoder, etc. could be effective while working on a huge number of datasets. Typically deep learning algorithms perform well when the data volumes are large {{cite:76f878dc2dc92a53d49793112c079da6498df752}} {{cite:10a7b5fa2e282195941be09f69375e371241bf4c}}. In addition, noisy instance analysis {{cite:adca3729c209e8e72a0c8644e1e7a14da68ec8e8}}, incorporating contextual information {{cite:0aaeaeb8a8bbe69460de012b281b4f8e1bd4e6cf}} {{cite:509bc7ba6e28acf405f8368ba6b2df28534879b0}}, or recency analysis considering recent patterns in data {{cite:2940fde84186a13de72bdd7e00cad490362c4d97}}, could be another potential research dimensions in the area. Overall, we believe that our CyberLearning model including a comprehensive experimental analysis opens a promising path for future research in the domain of cybersecurity, while working on machine learning-based security modeling, to make the security model lightweight and more applicable in the area.
| d | 7d08ab7fa2043bad1b6e36a096b88743 |
In this paper, we studied the non linear memory in the bulk of the space time around null infinity. Considering the large-{{formula:17424ba6-821e-4154-a8d8-81cba6bd2f57}} expansion of the BMS conserved charges at each order, we showed that despite the existence of a set of integrable charges at order {{formula:a445e3f3-087f-4369-8f46-4ac5913b34ad}} , the memory is not zero contrary to the case in the leading order. That may be related to the fact that the physical meaning of the NP charges is not very clear especially in the non-linearised theory of gravity while the interpretation of the charge in the leading order is the Bondi mass aspect. However, we only considered the supertranslation charges at subleading orders. Large-{{formula:c3d6f816-bcc6-4c64-a052-d687032308d1}} expansion of the superrotation charges which may be related to some conserved physical charges and the connection to the subleading memory remains to be explored in future works. However, the observable for the spin memory is a time delay See {{cite:239a7b56ef2fb56c057972212a4ff99b65f8d05c}}, {{cite:d3e2a9b2563c849cd13f451bf2f9a9d42c1276c3}} for the connection of the spin memory and superrotations. while in this paper the displacement memory is considered. Moreover, the gravitational waves and therefore the infinite towers of memories are characterized by the asymptotic shear {{formula:abc11f7a-64f6-4de5-ba05-adfac50c7486}} and according to {{cite:5ac21898105c4a30252d7bf6a502172a592a383d}}, the subleading memories are defined by different choice of the asymptotic shear. However as it is obvious from Eqs. (REF ) and (REF ), the displacement memories are not characterized only by the shearThe tower of memories should be only dependent on the functions {{formula:012f384a-a4b3-461d-8040-160f1cedb2d5}} and {{formula:21bdf16c-b96b-453f-a96a-2461e3ade727}} (See Appendix ). However, the subleading memory at order {{formula:f46568a7-6d56-4090-87e2-dffc0e0b307c}} and {{formula:139382ff-65fa-49ba-b244-a814d4dcacf4}} also depends on the functions {{formula:af5f00f2-bdd7-4e0a-9b93-73a8208782b9}} , {{formula:c3f56b9b-3253-4d2f-87e0-471d933e7ff4}} , {{formula:e2b42d56-b32e-48a2-a095-30529f439621}} and {{formula:e5464bd7-dfb1-4d8b-9381-eef501efe13a}} . Suppose in a physical system we let the Bondi mass be conserved. Therefore, there will be no gravitational displacement memory at null infinity and also not around the null boundary. which is related to the flux at infinity. We may solve this in a future work by considering the tower of dual BMS charges {{cite:818b51d36394a9db479c611c2475edfe37fbd3c0}}. Also despite the leading order in which the memory is produced by the non-integrable part of the leading charge, the non-integrable parts of the subleading supertranslation charges do not play any role in the subleading memory effect. Finally, in addition to the standard memory, a displacement in the radial direction has been calculated. It is shown that the shift in the {{formula:03bd0cc4-05b0-4408-b19f-c07f9742c541}} direction shows itself only away from null infinity and it is exactly the change in the relative radius between two detectors when the supertranslations act on the initial separation vector between two observers before passage of the gravitational waves. This confirms the method we used in the body of the paper led to consistent results with previous works.
| d | fd5770c0f0cab2f941c8abe9249aa074 |
We apply the real-space DMRG
algorithm {{cite:7fac7c03054fbf572481c9814680395e65c1f0c5}}, {{cite:55fdffd4b9b141211dd76c295a3945ef7964107d}}, {{cite:88eed2915a19f39e4aefb139e1981a581bb1370d}}
to the Hamiltonian (REF ).
Complex-valued and long-range electron transfer amplitudes
and anti-periodic boundary conditions require
an elaborate DMRG code that was
originally designed for calculations
in quantum chemistry utilizing various optimization protocols
based on quantum information theory.{{cite:fd515fd41f7a74a0e4efe5504ead86067a924ad0}}
| m | 37cdb70811350c907498cf220c9d4678 |
where {{formula:12df38a6-0552-4964-ba02-5b18935809e3}} . Following properties of the function {{formula:133cf973-cb19-4ca1-a7cd-6ba7b4bc6807}} are of our interest (see {{formula:92ec635f-e5dd-45fb-95b1-04ce49798b63}} and {{formula:6e3154d9-f39b-4205-8540-2fd97f5aabf6}} of {{cite:25fa08545571e51879c66ff4599370ac89919b1d}}, and {{cite:fd6ca44662fc9d0a9a2c57a4dddbe2e139227713}})
{{formula:1f8d35d2-f3f7-4163-9d5a-7fec4e4d8166}}
| m | 8b2a97413163de9b9d694b8f92eb4ad0 |
{{formula:9425b8ca-a16b-4591-894d-c188c576d203}} Comparison of the Action Classification Accuracy.
Since our method and hotspot {{cite:2f944d00070cda43b4e2ee84d5ee6d80ce813cd9}} only use action class as supervision during affordance learning, the accuracy of action recognition affects the results of affordance grounding. We compare the performance of action classification between ours and hotspot {{cite:2f944d00070cda43b4e2ee84d5ee6d80ce813cd9}} on the EPIC {{cite:b126be848c038f1faabf88bc213f5231e8e44dda}} dataset. We only choose the eight most frequent classes and calculate the confusion matrix, as shown in Fig. REF . Our method performs better on “open”, “close”, “take”, “put”, and “wash”, demonstrating that it can better distinguish different actions by combining hand position and action-related affordance cues to improve the result of action classification.
{{table:5ee77738-2b8a-4819-8d58-f84f5fdfcfd9}}{{figure:df0f052a-a4e5-469d-86d6-1486b3703918}} | r | 90fd99dbf936566a63ee710c9490180c |
Theorem 1.8 ( {{cite:2706aa7ba9d2177c3627db56f9ca0822e81c5e36}}) Let {{formula:6382212d-0506-40aa-aadc-c9fc29306d3d}} be a complete metric space
with metric {{formula:f13d3436-8f34-4acd-bb6e-bf849a030e74}} and {{formula:e464c416-374e-4bb1-8fd2-a3e61e8b59b8}} a lower semi-continuous
functional not identically {{formula:efb4dbe8-c5bf-4676-ae2b-1d07f71a4a34}} which is bounded from below.
Let {{formula:1ad20a2c-c4be-4cc7-83c6-5c98c67368f0}} and {{formula:412a149e-85bd-4535-907c-ad9f73d56276}} such that
{{formula:56c14d02-bec9-4124-9965-94921f045447}}
Then for any given {{formula:1a44e99e-417e-45a5-9d99-656988b6e15d}} , there exists {{formula:a4cb2d29-13ee-4a42-9bf7-bc5b57124f45}}
such that {{formula:b44c8650-8df9-4ba5-b3d0-17045c9f5caf}} , {{formula:5eac29a5-49c3-4d26-9b99-0495bd9bde7f}} , and
{{formula:bf6e95fc-c919-4bdb-9e7b-3a36269563c6}}
| i | 19269a664465da461c465f799a3cc7c5 |
The empirical insights are as expected if we consider the mechanisms behind those XAI methods. For instance, considering {{formula:a96c31cd-1d1b-4b78-9715-c6628c91e3ac}} , DeepLift and Integrated Gradients are more robust, since they use the reference point to avoid the discontinuous gradients (large curvature) that mislead the attribution maps {{cite:920d839e23f1e045d5e6a41eba2bfc65b1662d25}}. On the other hand, DeepLift and Integrated Gradients become vulnerable to {{formula:78a03da1-5c8b-4706-a7f5-8b0d4fe7ac71}} . Because misclassification and misinterpretation are rare events, it is very likely that perturbed inputs inside the norm ball have consistent interpretation with the seed. Consequently, the integration from the reference point which averages the attribution map over several points are prone to produce the consistent interpretations. We discuss more in Appendix to back up the empirical observations.
| m | e14706bfe2904a42303657138315a11f |
Domain walls (DWs), i.e. the cross border of regions with homogeneous but differently
oriented magnetization play a central role in magnetism
{{cite:4af660c397ff3214b6f2c6b4963be8f26d7040de}}. A particularly interesting suggestion is to exploit
DWs for high density storage in a "racetrack" shift memory
{{cite:0589d1dd08d3fe462e206faa6356089db5c2fc9d}}, {{cite:ebf0688c99302ed251665af7dcad0f6697618474}}. The shifting is brought about by passing a
spin-polarized current that exerts a torque on the DWs
{{cite:91531e8c06d58786d1767c0be90be4eef8f76c78}}, {{cite:6359ff80247dac8df069262d77ff3ba21ec560f5}}. While the proposal is technologically attractive
it is hampered by large energy dissipation due to the high current
densities needed. Recent advances resulted in an increase of the DW's
velocity at lower current densities {{cite:72bb78294b72f4f1be8a9e2733e6bdb64f42ee46}}, {{cite:5bbad1fc1ffba6d946cd51c34b1edca6cd44a329}}, {{cite:e278ba63356aa7cc15eecc80da23720f03669890}}.
It is however of interest to explore qualitatively new ways for
controlling DWs. Here we show that magnonic current may serve as an
efficient tool for the driving of DWs.
| i | a556f56082d35fd44b8d61889a263612 |
Next, we use two cooperative domains from the Stanford Intelligent Systems Laboratory (SISL) {{cite:2ebb389c18fc4d40b38871b072d943b438a3d24a}}. The experiments with these domains have two phases — training and execution. The algorithms train for 1000 games in the training phase and then enter an execution phase, where they execute the trained policy for 100 games (no further training). We choose to set the value of advisor influence {{formula:7cde0f37-e6aa-4957-97f0-251b2638a13b}} to 0.8 at the start of training and linearly decay it the same way as in the above experiments with Pommerman (since we are using good advisors). In the execution phase, there is no more influence of the advisor and no further exploratory actions, for all algorithms.
{{figure:9f6b4668-0c86-4470-99f2-c5f5bfc403d6}} | r | 9427986b6d29aba644e5a2168e724066 |
On MJD 55266, the maximum observed {{formula:cc846f60-e694-444c-8732-4b771770a078}} -ray energy was {{formula:794a5bee-28ca-4f25-b84f-1436dd7c08d2}} TeV which is produced from the interaction of Fermi accelerated proton of energy {{formula:63f5fbcf-c33b-46b4-ae36-5541581d4dfd}} TeV in the jet with the photons in the lower part of the SSC spectrum. In the two-zone SSC model of {{cite:1bbf50d7be67b3914a391feb1e45e38375a0b86f}} for this day the SSC spectrum starts around {{formula:6ef0735d-4403-4475-a7af-c0c8acd71eb8}} Hz (Figure 8(b) of {{cite:1bbf50d7be67b3914a391feb1e45e38375a0b86f}}). For this value of {{formula:b724bca6-cba6-4a60-9498-f0814a600a02}} , the minimum value of the bulk Lorentz factor estimated in our model is {{formula:ae3fc683-5e0f-47a2-b060-d4e5dc67df54}} . Similar or smaller values for the bulk Lorentz factor are obtained for other days. On the other hand, in the one-zone SSC model (Figure 8(a) of {{cite:1bbf50d7be67b3914a391feb1e45e38375a0b86f}}), the SSC spectrum starts above {{formula:84e69ad9-9dc9-494e-9e68-a2166f6f4651}} Hz and for this value of {{formula:d1197794-ae07-4c2a-9a21-2b110f5dafe2}} we get minimum
{{formula:86ecca62-8a67-4610-b8dc-009160382227}} which is too high.
| r | f3bdac3d819fae4a5473c8afc4285213 |
Besides the excellent experimental prospects from cLFV-dedicated experiments (COMET/Mu2e and Mu3e {{cite:7868c7edd0375e024611865c61887d0aecca962e}}), Belle II recently started its first data-taking run.
Being a {{formula:de682909-ec27-4bb0-8a74-8828fc7c0e4a}} -factory, it aims not only to scrutinise the {{formula:665f103c-5613-43e5-8f61-f17f007c9d06}} -anomalies, but also to produce copious amounts of charmed mesons and {{formula:3742e90f-49ed-4182-aa5d-50e03aca17e7}} leptons, allowing to reach an unprecedented precision and sensitivity on related flavour processes in the near future.
Going beyond a random scan, and to a comprehensive statistical analysis (a global fit of the LQ couplings) allows to explore the experimental prospects on how vector LQ models can be further probed in the near future {{cite:f8f50bb6e3920f2aa0dd792a248f9dbf62c7570e}}.
By means of Monte-Carlo techniques we thus estimate ranges for numerous observables to be investigated by Belle II {{cite:f2c632918a03929bf15794bb115673d1964f49eb}}, for three LQ mass benchmark points {{formula:38406df7-e065-47bf-9e24-3028f7f2cf86}} .
{{figure:c7d7c724-0e91-4244-9d77-7d0dd1365299}} | r | 69b7d4866e06b739c32985f2820d85fe |
Processing-in-Memory (PIM) architectures {{cite:c611f61cff7993dcfc5526c706e7f33206efc180}}, {{cite:c692a39fb7e9423edda4a591d5866c987290f9af}} promise to alleviate the memory bandwidth bottleneck.
The UPMEM PIM architecture {{cite:87b6d06f66dc4e7dfdba0cc5e2816c11687e85e4}} is the first PIM system to be commercialized in real hardware.
The hardware has been shown to be effective at accelerating memory-bound workloads {{cite:c8c74c86254f9a0af0af5c158742b2be36eb87af}}, {{cite:b8fc460312c0fe6000cba3602fe4266e579a5efa}}, including variant calling {{cite:70ac6b5f35125b4ad2486eebb7b264fc5dc29a11}}.
In this work, we introduce the first efficient PIM implementation of WFA for the UPMEM architecture.
| i | 1ea392bab4fb253558eb37e54b798113 |
During the past two decades, the standard {{formula:72bfcbeb-8ef3-49a3-8cd2-dc66c578cc0d}} -cold dark matter ({{formula:b058fe45-bc20-40b4-81ee-fc64b3c15e38}} CDM) model has been successfully confirmed by a large number of observations such as Type Ia supernovae (SNe Ia) {{cite:c1ec3ea21c62c267daeb65296589d3c8edc7392c}}, {{cite:8bcce7ed34fed175df662cce36a6cb29bd444ff0}}, baryon acoustic oscillations (BAO) {{cite:caaffd837207c6cae8b445c9509825c1778b2fbe}}, {{cite:ee2c75d84e9dfde2fc2f65b87f663b699ab0fe9e}} and cosmic microwave background (CMB) {{cite:59dfe23ca29079f2af89dd04804ac5616ee9e8fd}}, {{cite:5801f3debfd7c261227973405ac00ab0441ddfdc}}, {{cite:031f4d7fdcd89989e456b3e636c8d8741c2023af}}. However, it is imperfect and faces at least four intractable problems: (i)
the locally direct measurement of the Hubble Constant {{formula:e0396eec-85e5-43b2-8371-c43225c1d5b1}} from the Hubble Space Telescope (HST) {{cite:dfc180a124bf6ebae5415f4a3daa8dca6091424f}} is {{formula:2b4fdd53-f87c-43c2-9f84-c1af6adaabe4}} higher than the globally derived {{formula:ee7e7e4b-95f9-4a53-94c6-f01c5c3e3072}} value from the Planck-2018 CMB data under the assumption of {{formula:157e2984-11b9-4184-a9b4-d5b0df08ee25}} CDM {{cite:031f4d7fdcd89989e456b3e636c8d8741c2023af}}; (ii) present-day matter density fluctuation amplitude {{formula:0ad67a7a-c910-4251-9df4-044a95e15606}} measured by several low-redshift probes including weak gravitational lensing {{cite:b8e82609a34199da50c4987a8ed5147d8f93d869}}, cluster counts {{cite:99a488b4d4706ab8bfa8c957d2fc42de96bd3116}} and redshift space distortions {{cite:fa0d38085166db8b7a8e09495e34e3c3c141a007}} is lower than that indirectly measured by the Planck-2018 CMB observations under {{formula:8ec77011-6516-460c-bc8d-0a0ee0ba0cc2}} CDM {{cite:031f4d7fdcd89989e456b3e636c8d8741c2023af}}; (iii) the value of cosmological constant from current observations is much smaller than that from the theoretical calculation; (iv) why the present-day densities of dark matter and dark energy (DE) are of the same order {{cite:01a18ff776c20e3c8d97186cc9b378e2a04f9464}}? These pendent problems have initiated intense discussions of possible new physics and underlying systematic uncertainties of the measurements in the community (see Refs.{{cite:7402fbd9d6eed62a5c87695a74ff2e52ef671e70}}, {{cite:6e32cac40f6c30d11812c8366c502e03eb4c8d13}}, {{cite:55487383e18d663a5b3f1ade9a00d90189ffbade}} for recent reviews).
| i | d26e7840789cc707420d1b2cdea0efdd |
In the last chapter the analytic functional calculus for quaternionic functions is obtained (see Theorem REF ), as a particular case of the general functional calculus given by Theorem REF . We also recapture one of the main results from {{cite:5ee8b38dc385a251fc22512f1e87c15804331658}}, showing that the slice regularity of a quaterninic function is equivalent with its property of being the Cauchy transform of a stem function (Theorem REF ). The main ingredients of this result are the
representation formula (REF ) and a similar result from
{{cite:833680bc89bfeacc5420b624b4d4fbeb57cbbfd8}} (Lemma 4.3.8).
| i | 912cc8f33e2c6ceb7b2b7ab147154c93 |
Since the level at which EDE models differ from {{formula:bbf0c408-e188-4bcd-bb4d-62cc08f3c88c}} CDM are at the {{formula:f3c5c139-ac98-449c-ac24-e87d3c2984f3}} -ish level, small changes in analyses or model assumptions that one might guess were “below the radar” can actually affect the conclusions. For example, the galaxy-clustering constraint to {{formula:97998fa5-8201-4dee-bb82-6b906a28ee40}} changes by about 0.7 km sec{{formula:0dffd6de-417e-467c-bce6-526413b8e607}} Mpc{{formula:77ce95cd-066c-45b8-8798-38be6b23c7f7}} if the dark energy is assumed to be a cosmological constant or described by a more general equation-of-state parameter {{cite:81aa20809a7b91146f77a91f41694d5cd57d520c}}. Similar shifts arise from different assumptions about still-undetermined neutrino masses. Although these shifts are small, at the {{formula:003b79d1-ab68-491a-81f3-6853e1e1085e}} level, they can change a result from the {{formula:924b6738-224f-4bbe-bcd2-5753bf031670}} threshold to a {{formula:32e11844-5b50-4e68-8357-0d49036cf18b}} result. At this level, conclusions can also depend on the interpretation of the statistics.
For example, Refs. {{cite:9486203fe117f52c56260de5f53cba42815fc3f9}}, {{cite:bfc0a30d9a2203fe1ea17cd2abf10f165551134f}}, {{cite:91046a31afc5f6739c2320c7c405f77e811cfb99}} argued that current data favored {{formula:5964fd73-47c4-4616-b9ed-21fac297a780}} CDM over EDE at the
{{formula:d4f8f07f-c1b8-4724-8740-d845fdacdf7b}} level, while others {{cite:3a43570dd0ca21c441bf8740c4aa809738433fd2}}, {{cite:7c01aecabdc8d9acf45516416c076fb4598734b8}}
warn that the conclusions may reflect the choice of priors.
| r | d78f777ab208e9b3cfa7559cb0d39548 |
{{formula:41022fa9-3555-43e2-ab4b-c762fde6281f}} is an instance of random variable of assitive information that is drawn from multivariate Gaussian distribution {{formula:24bb8bcc-d08e-43e1-b2d9-1e789f7a622c}} , {{formula:057b9e11-302a-46e8-ae28-45d653e1bb4c}} is the parameters of encoder {{formula:301b4f92-8a6a-4b98-9f5f-4d2f9521e362}} and {{formula:83a7cc4c-f11b-4924-be90-c4bce31e9c7e}} is an identity matrix, a Lagrange multiplier parameter {{formula:30957beb-778a-45f9-a595-f5e30d0a7718}} 0 is used to control the trade-off between the encoding the necessary information and reaching maximal compression. Yet this does not lead to a learnable model, since this mutual information is intractable. With the help of variational approximation, specifically, deep variational information bottleneck {{cite:85d705baf503df151f08318525fc3d2a445b598a}}, we are able to parameterize this model using a neural network. We then derive and optimize a variational lower bound of such objective as
{{formula:22e68f4b-93fe-4f2a-87a3-44eb220b2840}}
| m | 3df171c0790bd9ccc6ececafa4e7062d |
The baseline methods RANSAC {{cite:b5c382a923c55f1b68fa2638858f137ea0ad7ccc}} and FGR {{cite:b8f47883f321e23067f759e5e28a0fffdcca0119}} have been implemented in Open3D {{cite:4eb8e214599f3c556bd04d62e155e06e76f8ef89}}. For GC-RANSAC {{cite:6128ad019f3d65b07ed8ecebf73cfc5ef783f31b}} and TEASER {{cite:42db4a21c04cc95fac1bdb6fc1b7560ced299fdd}}, we use the official implementations. Note that we use TEASER with reciprocal check; otherwise, it takes an extremely long time for testing when the number of input correspondences becomes large. For DGR {{cite:61d382d33b3e9ad29adb1088f60cdb9a62b91190}}, we use its official implementation and the released pre-trained model.
Due to the unsatisfactory results of publicly released code,
we re-implement SM {{cite:9c2f656ae45eebb8b6d345994338cd0529478ea4}} and 3DRegNet {{cite:eb140f1dddac5280130f5d4ef400174dcffb1bbb}}, with the implementation details as follows.
| m | ec0b4602938d44e1e6b393e142d0930a |
Our analysis relies on the contraction of hidden representations in total variation stated in Assumption REF . As discussed in Section , the standard Doeblin's condition is sufficient for the contraction {{cite:a00b71ea7e84504d8f5e19432d8b66c790d1eaf3}}. We conjecture that it is possible to prove Doeblin's condition for the chain of hidden representations: Gaussian weights and full-rank inputs allow the exploration required in Doeblin's condition, thereby ensuring the contraction of the distribution of the hidden representation.
| d | 5bd6603d143eed0c035f43880a83a715 |
Here, we have not addressed the inference of interior parameters from specific exoplanet data but have outlined the physical reasoning and implications of magma ocean interiors on transit radii and physical state of the planet. This interior model introduces a higher degree of non-linearity compared to other interior models {{cite:ba301a035119a6e6d872e75ce2d76dfc358231c7}}, {{cite:1910499293d862501ccab759fd8c94e45e8e89d7}}, {{cite:90722959c1a75aea141a2a30c3379154d98a8562}}, as there are stronger inter-dependencies between interior parameters. It does not, however, introduce a higher degree of degeneracy. Thus using our model in an inference scheme does not lead to larger uncertainties on predicted interior properties. The new model comes with a higher computational cost: it is two times slower than the previous models from {{cite:1910499293d862501ccab759fd8c94e45e8e89d7}}. However, inference schemes can be done more efficientlySome model parameters can be computed extremely fast (e.g., planet mass), while others require a detailed structure computation (e.g., planet radius). For a more efficient inference, a detailed structure computation is only done, if a newly proposed set of model parameters is not rejected, assuming a perfect fit for model parameter that would require an expensive computation. such that the overall computational cost for an inference analysis that includes our new model is two times faster than previously {{cite:1910499293d862501ccab759fd8c94e45e8e89d7}}.
| d | 631a2ff28bce9fa25881bfc81fba08f6 |
Purification Comparison.
Table REF compares the purification performance with the state-of-the-art noise detection methods based on the training dynamics, including AUM {{cite:9c49235c2214af7a326b40205991549e685a558f}} and INCV {{cite:9b0662dd42c5d61501c075e68dc483ec031a131a}}.
We notice that the performance of AUM and INCV dreadfully declines when detecting label noise among only a small set of data, which is inevitable in online task-free setting,
whereas SPR can filter superbly even with a small set of data.
Even a non-stochastic version of our Self-Centered filter performs better than the baselines. Encouragingly, our method is further improved by introducing stochastic ensembles.
| r | 45113554a622b0aa6f95e5183ea3b479 |
Using the policy gradient theorem {{cite:d6703a19a8ad9183bd5d0468742cd35ddf194b90}}, we can write the gradient as follows:
{{formula:c196ad6d-8613-455e-aec7-75be30a7c1fb}}
| m | cff4236d3fd8941a5771e839204a556e |
Model rankings. First, M4C {{cite:62629027736550445f1015884ed477c06eddfe68}} performs the best among the evaluated models. Interestingly, it is smaller than many of the more sophisticated model architectures that score higher on VQA. This is probably due to the importance of the ability to read and reason about text in the image for answering some AdVQA questions. Second, among models that can't read text, VisualBERT {{cite:3257f848b8aa8abd31b461ab20695f10661a4e41}} is the best model, despite (or perhaps because of?) it being simple and having fewer parameters compared to other models. Third, the adversarially-trained VILLA {{cite:6d22a199a1da2cf03413e0b01d26dde19c5380ec}} model performs surprisingly poorly. While it may be more robust to statistically-generated adversarial examples, it appears to be less so against human-adversarial examples. Fourth, we find that all of these models perform poorly compared to humans, while model performance on VQA is much closer to that of humans.
{{table:a54b4fbc-1524-4e3a-b4d3-fadfca446263}} | d | b771d289a133bdd7e6b6adc9c4957622 |
An Android app was designed to capture 34 images of a scene at 8M size and fixed ISO, fixed lens position, with the auto-white-balance enabled, across different exposure times. 34 exposure times varying from 0.5 sec to 1/4000 sec in steps of 1/3 standard exposure value were used to generate a ground-truth irradiance map image ({{formula:524eec56-07d0-4493-82ba-e1770343ac4e}} ) of the scene using the method introduced in {{cite:23c1efc4b0870755678dd72f8976782e23c8d67f}}. The above process was repeated for 10 different indoor and outdoor scenes. The complete dataset is made available for public use at http://www.utdallas.edu/{{formula:f0c56072-65e7-4c65-b98d-80e5da482719}} kehtar/ImageEnhancement/
| r | d0b0ae64b48dc29a9421f1af1549036d |
Summarizing, within the framework of a theoretical model described in the
Introduction, we have studied time evolution of the spin part of the singlet
wave function of two electrons in the presence of external static
homogeneous magnetic and exchange fields. In order to obtain the exact
solution to this quantum-mechanical problem, we have had to revise in
Section 2 the traditional approach {{cite:6b4c668c2cdd0dfb495c9d22efc68217393d92e4}} to the spin singlet, because it
does not take adequately into account the property of invariance under
rotations of the coordinate system. Basing our own approach in Section 2
solely on this invariance property and using the theory of spinor invariants
{{cite:469283e091ae8a72238ceb966b5c687dec6bd9f1}}, {{cite:fd62bd3aacddac86ee53350fd3f01c03c62d64f1}}, we have derived the generalized representation of the spin
singlet [Eq.(REF )] whose fundamental feature is that the spins are in
mutually time-reversed states.
| d | 0f1318a369cd467a7f33515733217b3b |
The proof follows from {{cite:cecbfe8dfa97cf15bda5d26becf510a408ab749a}} and {{cite:fcf89472f7a2b5c024f6eab4bc6cb6ba6e0a7432}}.
| d | e5389aec03bfd08c90c4581eb813d1fa |
Fig.REF shows also that the overall capacitance exhibits a local minimum at the origin for different values of {{formula:f49c2e35-745d-4a0b-9ae0-f79f1dc09a3e}} , and an asymptotic increase towards the value of {{formula:9b528f31-db27-4158-bf51-f6a41a5e98be}} at higher polarization. Otherwise, between these two limiting cases, the shapes of the profiles for {{formula:e6fbfc92-fdec-47f4-8300-78e18cc03a5f}} are in concordance with the shapes obtained with the traditional GCS model for which {{formula:af17b11e-4330-4a15-bf30-3953801b9202}} (compare with Fig.REF ).
Now because of the interdependence of internal subsystems for {{formula:2a9e57ae-12c3-47d0-a1a6-e9275ed0a542}} , this results is an extra contribution to the amount of entropy of the system, i.e. the entropy {{formula:0f625b45-1798-4655-81db-1f4c4ea98bf9}} of two subsystems {{formula:3b11dd82-50c3-4caa-8a1d-a2c43a89a28f}} and {{formula:f6b487d8-593a-4641-9139-6176154806da}} {{cite:559d329d338c1da917d29b4bfb6f3e3c19721a76}}. This non-additivity of entropy is the basis of a non-extensive Tsallis statistics.
Tsallis proposed the generalized entropy
{{formula:be79f049-994d-474e-bdf8-ad51bdb7acb7}} {{cite:31a45183738cd89587431e3e46bd3928bd4e6b8e}}, {{cite:f6ddb5389248f97fef657f034c6ed3027c299a13}}, {{cite:559d329d338c1da917d29b4bfb6f3e3c19721a76}}, {{cite:ae662b22ca4fde8e516250a0c780d72e317a88c1}}, where {{formula:79fe6ffe-09e5-4333-8124-d655f799af58}} is a positive constant, {{formula:903f876c-9f91-4b21-b58d-a10f5fa5703f}} ,
and the quantities {{formula:bfc97202-390e-483a-85de-4a76c538bffd}} represent the probabilities for the occurrence of the {{formula:16459105-a0e8-4787-9898-21864dc45849}} microstate and satisfy {{formula:b64a8e91-afa3-42c2-8438-a68bb27721c8}} . The function {{formula:69bbf629-b05b-4552-9e85-eef0f41dfd0e}} denotes the {{formula:048428cf-ec05-4ee1-8478-4e776f86ba24}} -logarithm, inverse of the {{formula:533c4f91-3520-4a2b-b756-44aa53583d73}} -exponential (i.e. {{formula:6a801c79-cf5a-4a28-8ab9-f49062fa2066}} ) {{cite:559d329d338c1da917d29b4bfb6f3e3c19721a76}}. In the limit of {{formula:1495eb60-650c-4d08-b7be-dc0efb77c576}} , one recovers the classical form of Boltzmann-Gibbs entropy
{{formula:a2da6858-b5ef-4224-ac2a-66716afeb9d4}}
where {{formula:6d9cd629-e45e-4cb1-bfb3-e87135f9d3c3}} is the Boltzmann constant {{cite:31a45183738cd89587431e3e46bd3928bd4e6b8e}}.
Such additional amount of entropy, which is related to the value of {{formula:544b728a-2147-48dc-8817-a1a4da9ae274}} , affects negatively the net density of charge in the diffuse layer of the EDL structure compared to the traditional case when {{formula:07feaed9-7591-4ba3-a2da-9087f09daf7f}} . This in turn lowers the overall EDL capacitance as depicted in Fig.REF .
| d | 68bc0aab18d8051560c54b9138f50328 |
When comparing simulation results with observations, it is important to pay attention to the way quantities are evaluated in simulations. In particular,
estimates of metallicity (as well as other thermal properties) can be derived in different ways depending on the weight {{formula:8fd9a3c9-746e-4166-9409-30c3c245b2f3}} used to compute the average {{formula:d689e85e-b35a-4951-a60b-68e5e5683fbe}} value, that is {{formula:b1f88792-2670-400c-bcdd-2ed3a4871237}} . Typical weights are the mass of the gas or its emissivity in the X-ray band.
In general, flatter metallicity profiles in simulations are better reproduced when a projected emission-weighted estimate is pursued, whereas the mass-weighted three-dimensional metallicity typically shows a somewhat steeper decrease with radius {{cite:f342bb8f45ddbcb4cec4067cc67dad9447dd3ace}}.
In this perspective, the issue of a fair comparison between gas properties in simulations and observations has also been tackled via the generation of detailed X-ray synthetic observations out of numerical simulations, properly taking into account the specific characteristics of X-ray telescopes.
With such techniques, {{cite:f5cffb4d51706efb22f826fa462310e8ab77d0ae}} showed that an observational-like reconstruction of the iron and oxygen abundances with mock XMM-Newton observations of simulated clusters and groups is in good agreement with the intrinsic simulation value.
More recently, {{cite:612629551fcc1b83fa80d71c62bbd791d68a0fcc}} employed synthetic observations of the X-ray Integral Field Unit (X-IFU) on board the next-generation European X-ray observatory Athena to reconstruct chemical properties of the ICM from simulated galaxy clusters. The authors showed that the metallicity values obtained from X-IFU spectra match well the emission-measure-estimate computed directly from the simulations.
| r | fe9a6d29e92388363f4141a87d86c030 |
In order to evaluate the capability of our model to adapt to sophisticated scenarios, we illustrate in Fig. REF the side-by-side comparisons of our DCPose network against state-of-the-art approaches. Each column depicts different scenario complications including rapid motion, nearby-person, occlusions and video defocus, whereas each row displays the joint detection results from different methods. We compare our a) DCPose against 3 state-of-the-art methods, namely b) SimpleBaseline {{cite:300fe85edcbda753429982db88393fc4669f9b1b}}, c) HRNet-W48 {{cite:3ae2840f5e1884b5224a6560a6162a5cbb029a5e}} and d) PoseWarper {{cite:e1c8f4e91da9e5efbfc4de05f61e49850c15fcdc}}. It is observed that our method yields more robust detection for such challenging cases. SimpleBaseline and HRNet-W48 are trained on static images and fail to capture temporal dependency among video frames, resulting in suboptimal joint detection. On the other hand, PoseWarper leverages the temporal information between video frames to warp the initial pose heatmaps but only employing one adjacent video frame as the auxiliary frame. Our DCPose network makes full use of the temporal context by processing consecutive frames in dual directions. Through our principled design of PTM and PRF modules for better encoding of these information as well as PCN for refining pose detection, our method achieves new state-of-the-art both quantitatively and qualitatively.
| r | 6e596ad6045b0c0eb0ab917267a33e4f |
From Table REF , in general one can observe the benefit of incorporating the attention mechanism in improving the generalization performance of the models across multiple subjects. For some models this improvement is more significant than other. In particular, the positive influence of the attention mechanism on the EEGNet can clearly be seen for setup 2. From Table REF , it can also be seen that as the number of training subjects increases, i.e. going from setup 1 to setup 2, the accuracy is improved and in general the influence of the added attention mechanisms is more pronounced.
Overall, the AA-MultiviewNet with self-attention yields the best result, with the maximum accuracy of {{formula:f18ba0a5-4213-41ff-9faa-98342e6007c5}} . In addition, when considering the performances related to setup 1, using the attention mechanism does not consistently give better results than using the vanilla models.
One can explain these results under two assumptions. The first one is the lack of samples in setup 1, preventing the attention mechanism to learn important features, and thus to be effective as generalizing to unseen subjects. The second one is the increased complexity of the attention models. Since the attention-augmented convolutional layer requires additional intermediate convolutions as well as non-trivial operations like tiling, transposing, and reshaping {{cite:b5c9c4bba889b5651e9c762c6833371cf84edd64}}, we think it negatively impacts the performance when given a smaller number of samples.
| r | b9d6e1df71e695f3e732f26573018c60 |
In the white-box setting, the owner has full access to the pirated DNN.
The watermark can be encoded into the model's parameters {{cite:c744d288d07e95fe395672ce4bad560f4f39d504}} or intermediate outputs {{cite:e4e2935e1e8313a287e49903da1d90155e04a1e1}}.
The owner can also insert extra modules into the DNN's intermediate layers for OV {{cite:755c94042b6b304be981fd82cb69777322856c3b}}.
As for the black-box setting, the owner can only interact with the pirated model through an API.
Watermarking schemes for this case usually resort to backdoors {{cite:f9adda4eaea41c24e5e2bbe91c6d31ba33e9872f}}, {{cite:3a925dd9025b6b9651f03516af7a10909269b376}}.
| i | e46d9842db0d971d4cc4dce28e6c545c |
We compare our method with several baseline methods to demonstrate the performance. (1) PointRNN (LSTM) and PointRNN (GRU): PointRNN is proposed in {{cite:eb91231e820f5a6b4063e26abe40d9447da54abf}} and we call the two variations using LSTM and GRU as PointRNN (LSTM) and PointRNN (GRU). The networks are re-trained on the two datasets due to the different preprocessing of point clouds in {{cite:eb91231e820f5a6b4063e26abe40d9447da54abf}}. (2) PointNet++ (LSTM) and PointNet++ (GRU): We define two additional baselines based on the encoder-decoder architecture of PointNet++ {{cite:04f0012c9974ca7fc3baa5f6d7c46c166ff976f9}}. Firstly the input point clouds are encoded into one-dimensional global features. After that, LSTM or GRU is utilized to model the temporal information and the generated global feature is decoded into the predicted per-point motions to generate future point clouds. (3) Scene flow: We utilize the 3D scene flow estimation network FlowNet3D {{cite:4a6cb9c4e1b9635cffd2acff1cd59b751187a9bb}} to estimate the 3D scene flow between the last two input point clouds and use the estimated scene flow to predict the future point clouds. FlowNet3D is finetuned on the two datasets to achieve better performance.
{{figure:b4b98af5-51b9-41b5-b8e2-82e3a75505a3}} | m | 6d74dc4ea25106fd4ebcef585dc97b84 |
Actually he had, as far as I know, four famous encounters with gravity that went into print.
The first one was in a paper with Henk van Dam {{cite:a9e0839bb41fb81df4782b0fc4a874016a6670c9}} on the lack of a continuous limit from massive vector gauge theory and Einstein gravity theories to their massless and concluded that the latter should be strictly massless.
The second one on the one loop gravity
action {{cite:e96ed8939dd0c966386ffed41e984f8295057b15}}, with Gerard 't Hooft. The third one was a Letter on the effect of the Higgs condensate
as source of the gravitational field brought up by Linde {{cite:af9f895d77f359ff46b8d4e22f0e5aee48fb335f}} a few months before. The fourth one in his lecture course on gravity at the Les Houches school{{cite:f0b24ddf51fde5d30382100a0f888ca952601b4e}}.
| i | 78636a2e39eecd64306f6982f36112e8 |
Although self-learning-based RSISR approaches such as ZSSR {{cite:a2880934b87f7dd5a39ed28df09093702b0a5b11}}, KernelGAN {{cite:7f0084741ad87bd291b271d355be8d46f6f66484}}, and DBPI {{cite:cd6614112f2d8396b613d99acce3ac4f39b66b67}} can be easily adapted to LR input images, they generally have two main shortcomings due to the self-supervised training strategy. First, the optimization of SR models only utilizes the internal information of the LR input, while a great deal of external information is neglected. Second, these methods are usually time-consuming in the testing phase because of online training. To overcome these disadvantages, meta-learning is introduced into recent self-learning-based SR methods {{cite:595e68db948b4a02f076728cc254e17ed25e1bf8}}, {{cite:ddbd91ff1e4d43b8251d4460881de92cfd18ca1f}}. Based on ZSSR {{cite:a2880934b87f7dd5a39ed28df09093702b0a5b11}}, Soh et al. {{cite:595e68db948b4a02f076728cc254e17ed25e1bf8}} present the meta-transfer learning for zero-shot SR (MZSR), which consists of three steps, i.e., large-scale training, meta-transfer learning, and meta-test. The source code of MZSR {{cite:595e68db948b4a02f076728cc254e17ed25e1bf8}} is available at https://www.github.com/JWSoh/MZSR In order to ease the training of the SR network and the meta-learning, the large-scale training step first trains an eight-layer SR network with the pixel-wise {{formula:32070373-c84c-4d67-8d40-282ba94db473}} loss on the large-scale dataset DIV2K {{cite:401e786701a2e9b238fbe53d23f87d0378830a7a}}. The meta-transfer learning process aims to find a generic initial point for internal learning following the Model-Agnostic Meta-Learning {{cite:fe0acbfcaa403ebeb81fc77b90d315c38d6e384b}}, making the model can be quickly adapted to new image conditions within a few gradient updates. In the meta-test phase, the input test image is first degraded to produce example pairs for model parameter update, and then it is fed into the updated model to generate an SR result. Thanks to the meta-transfer learning strategy, MZSR {{cite:595e68db948b4a02f076728cc254e17ed25e1bf8}} achieves competitive performance in terms of both the quality of the super-resolved image and the running time. In {{cite:ddbd91ff1e4d43b8251d4460881de92cfd18ca1f}}, Park et al. also propose to improve the performance of SOTA SR networks such as RCAN {{cite:b3c3c411be2719d80f8cf7917ad123bfd4a509b0}} using the meta-learning strategy, without changing the original architectures. The source code of MLSR {{cite:ddbd91ff1e4d43b8251d4460881de92cfd18ca1f}} is available at https://github.com/parkseobin/MLSR On the whole, meta-learning-based SR approaches have strengths in reconstruction quality, generalization capability, and processing efficiency.
{{figure:23cc32d7-aaef-4b5d-8891-f1d7f1586058}}{{figure:6d6d0ba5-f1b5-441a-bd53-73f0b0a1c56d}}{{figure:3ef0b585-2683-4ccd-8886-010d5fe9ccd7}}{{figure:0af6cb92-8759-416a-8fbc-6e9dd9e88bf3}}{{table:04118ba2-2491-4d3c-9428-d647227a6976}} | m | 9d9a5d984b0d76b135f2578fc0e325db |
Among them, the tensor tubal rank induced by t-SVD can characterize the low-rank structure of a tensor very well {{cite:dfbf47a4c1e0046fc8fb6beb5480aaa5fe217482}}.
Its convex relaxation is the tensor nuclear norm (TNN) {{cite:952eaab9e367a60635744ff3d89e60930edc2d70}}.
TNN is effective to keep the intrinsic structure of tensors;
t-SVD can be calculated easily in the Fourier domain and TNN minimization problem can be efficiently solved by convex optimization algorithms.
Hence, TNN has attracted extensive attention for HSI restoration problems in recent years {{cite:2d891bc6f27a8c81fe44b6cfb39b2df983d621a1}}, {{cite:3c7f98963631dc3b0d4000d8bd3861119ed3f2b9}}, {{cite:c2141d8482458979fbc5727e8d0108bd73ee7ebf}}.
However, during the definition of TNN, there are three kinds of prior knowledge that are underutilized for further exploiting the low-rankness in HSIs.
Firstly, in the Fourier transform domain of HSIs,
the low-frequency slices carry the profile information of HSIs,
while the high-frequency slices mainly carry the detail and noise information of HSIs.
Secondly, in each frequency slices,
bigger singular values mainly contain information on clean data and smaller singular values mainly contain information on noise.
Thirdly, low-rankness not only exists in the spectral dimension but also lies in the spatial dimensions {{cite:85184b90a6efe728e4ef421cd0d0d67c546b5687}}.
The classical TNN only takes the Fourier transform to connect the spatial dimensions with the spectral dimension
and lacks flexibility for handling different correlations along with different modes of HSIs {{cite:2d891bc6f27a8c81fe44b6cfb39b2df983d621a1}}.
| i | a702bcb004b7ad83b32253d8410e8a2b |
Given a monocular video captured from a freely moving hand-held camera, our method reconstructs a neural scene representation that decouples moving objects from the static environment, assuming a constant illumination and known camera poses (e.g. calibrated with COLMAP {{cite:ee3f8f104a33eecc3ff8fefdfb0e3be8a56a8535}}).
As illustrated in Figure REF , our method achieves this by learning separate radiance fields for static and dynamic portions of the scene, and doing so in a fully self-supervised fashion.
We describe our architecture (Section REF ), detail of our self-supervised losses (Section REF ), and describe how, while shadows are not explicitly modeled by NeRFs, a simple technique for their effective removal is attainable (Section REF ).
| m | bbc7da455b94b845d000c43818bbc8e0 |
However, when restricting to linear perturbations around a static background the right-hand side vanishes while the left-hand side does not. To not mask the transport coefficient that appears in the linearized equations of motion it comes out to be more convenient to eliminate {{formula:a085a882-46d5-4ddf-aaf9-3ae439f87b40}} instead of {{formula:53a4208c-ca67-4be9-8039-89cce48cdda9}} . This is consistent with the well-known result of the fluid/gravity correspondence {{cite:043c9c7740a719d597dd987c4a8294a580f9defa}} where the coefficient {{formula:41794535-c7b7-4ab9-8778-7d7fe8fd7cb1}} associated with the tensor {{formula:65edc4fb-2510-491c-91ee-39dfe97c517d}} is shown to vanishes for the {{formula:90bf6ef6-3dae-443e-a5ff-6acba8a0198c}} SYM conformal fluid. Besides, this second-order transport coefficient vanishes for a conformal fluid in arbitrary dimensions {{cite:29b9a3a4dc1393420e5f9d6b544f773e26d3d88f}}, {{cite:a18fcd9b7905cd2cd4df678802de7ae6313b565b}}.
| d | a3e75516b633c97304004e4a3bc10127 |
Lemma 2.4 {{cite:4908545c468742f670962d92bfca685153c52aa8}}
Let {{formula:1cce143b-6daa-4789-8b72-7bd864e167d7}} be an Hermitian relation and be bounded from below with lower bound {{formula:24048ed0-d154-40b0-ae30-51d409500ec0}} . Then
{{formula:15771544-fc2c-4464-b911-750d9389b88c}} , and the deficiency index {{formula:981271f9-e2ed-4baa-9c36-6f2a89731ae1}} is constant for all
{{formula:c75952d9-27b7-4c82-87a8-dea345e40d52}} .
| r | 2b514e6c3987a686fabbc6af9267cd81 |
The numerical study was performed using a commercial electromagnetism solver tool COMSOL Multiphysics (Version 5.5) {{cite:c04bb0d1f7da67c3e9f1dfc3ca0beea851c62918}}. We used the wave optics module to solve Maxwell equations in our simulation domain for the wavelength range 500–1100 nm. The dielectric function for gold was taken from Johnson and Christy {{cite:07817ba59cd7d3f8db1325454e3dd2f9cd8dd141}}. The nanoparticle dimer was placed in the middle of a hollow multi-layered core-shell structure (Fig. REF ), which consists of the modelling region. Air were used as the background index and perfectly matched layer absorbing boundary conditions were applied to prevent reflections at the simulation boundaries. Optical properties of the gold dimer were calculated by performing electro-magnetism (EM) calculations in the scattered field formulation, which has yielded reliable results for optical properties of plasmonic nanoparticles {{cite:68ea1983486bca2bd99902ec86a79159616d24e0}}. The background EM field was taken into account in absence of the scatterer (gold dimer), and the scattered field was then solved for the presence of the scatterer in the modelling domain. The total field consisted of the background and the scattered field together. The scattering cross-sections of the nanoparticle dimer were calculated by integrating a sphere over it and applying scattering boundary conditions. The reliability of our model was verified through by comparison with Mie theory for the visible wavelength range (see Supplementary Note 1). The corners and edges of nanodisc and nanocube dimers were rounded by a radius of 5 nm to counter the EM field effect related artefacts, which can result due to sharp edges or corners of the nanoparticle. A blue-shift in the scattering spectra was observed as a consequence of rounding the edges of the nanocube dimer (see Supplementary Note 6), which has been found in agreement with existing studies detailing the effect of rounding on the scattering properties of nanocubes {{cite:c02d199f01b715d8f31c36774f1d3ecf9ebdef03}}.
{{figure:ab4b040a-4fc5-4783-a13c-9b3f0a0d384f}} | m | b7a8946bfc47a138a60e35e5b5704723 |
For the client-side optimizer, our default configuration {{cite:b412fcbf5b4602e515d9f3d3a656f25fd09c6eb2}} uses Adam to train the model. Our experiments indicate that using Adam as a local optimizer performs worse than using SGD because a cross-device FL system expects stateless clients, whereas Adam has stateful parameters. This was also found to be the case in {{cite:6a3283c296456b40327ea31da59dda73d948b751}}. Thus, we use the SGD optimizer for clients. We set {{formula:65bbaa27-32d5-4c7c-8b6a-ece32db8a325}} defined in Algorithm to {{formula:06c3eaea-3b96-4b9c-be0a-701487444673}} for all experiments as we found this to be the best value.
| m | 9a447c957a93c47c248613ceccb55b0a |
Our results have implications to the analysis of the distribution of elements in supernova remnants against explosion mechanisms. There is an ongoing dispute on whether the morphological structures of clumps, rings, and filaments in some supernova remnants result only from instabilities in the explosion as the delayed neutrino mechanism predicts (e.g., {{cite:57b9d79bb0ea1fe8ceb78fbfb413c890819d4598}}, {{cite:4394a76ccd126776f19bac52b98201ee44ae1603}}, {{cite:bc378e400c0f82864d80c5600b6fd5a1b35dcb67}}, {{cite:b7d41e0cecc089a95ab7d6de80ab662532abd2c7}}, {{cite:6d3817d943914fdcbaf8d843b11e4f1920829521}}), or whether in addition to the instabilities jets also play a role as the jittering jets explosion mechanism predicts (e.g., {{cite:4a68262a145740bc2a2848bdee21d11aa9188bf3}}, {{cite:7bac65a416de8067e05f5bd60c6a7aaa193ede6d}}, {{cite:b783436e2fac60d3235248bfc2860c73314fcdd5}}, {{cite:06aced9fbd3b3c61d1a0e3d9babb160f21cea82a}}, {{cite:0bd8662709431e8e2520cb6564573a6e64bb4b21}}).
The most recent dispute is over the supernova remnant SNR 0540-69.3, for which {{cite:69580010d67cfe4775940c48cfe7b6be86f0e3b5}} adopt the delayed neutrino mechanism and argue that instabilities alone explain the rings and clumps in this supernova remnant, while {{cite:9be8a32759a751bbcd88ff83b71f950d1033f767}} claims for imprints of jittering jets as well as instabilities in this supernova remnant.
The present study does not touch directly the dispute over the explosion mechanism of CCSNe. Our study adds to the rich variety of processes that jets might play in CCSNe, and by that indirectly might be used in analysing the morphologies of CCSN remnants. For example, instabilities in the explosion process itself might form clumps and filaments in the ejecta (something which we did not include in the simulations). The jets that we simulate here can shape these clumps and filaments to a bubble-like structure. Namely, the dense shell around the large bubble that we present as the red zone in the upper left panel of Fig. REF , might itself be composed of filaments and clumps. A long filament near the polar zone of the dense shell might acquire a ring-like shape.
| d | 66e20a45f94652dd04b58779d2c3b784 |
Unlike local approaches, global approaches are designed to be invariant to the initial transformation error. Some methods such as GO-ICP {{cite:5331027d20f1ea01aa0604d3ca22a0e245f498c5}}, GOGMA {{cite:c1a939963e7a8a72574b4b60299aa4d8dea08c08}}, and GOSMA {{cite:63e99f1d0e85340a9e07b504a7a913b66d1d8bd1}} search the SE(3) space using branch-and-bound techniques. Other methods {{cite:fcb281a005088e57ec1993696e61e34635319a36}}, {{cite:d5d755210aac4d165c1e1aac578980ba36e719c8}}, {{cite:948e394b567a6cc5ea7cd8e4a2a3f59835103bbd}} match the feature with robust optimization. However, these methods are unsuitable for real-time applications due to their large computation time. Fast Global Registration (FGR) {{cite:27c70f0c91f1100144d1902b0e637fa4a61fa336}} is presented to address this issue, achieving a similar speed to that of many local methods. To further improve the accuracy of the registration result, recent work handles the registration problem via learned global representations. DeepGMR {{cite:94feee266dc067b09c43cd581680b9993405a547}} represents the global feature through GMM distributions and EquivReg {{cite:5b322455df3b7684b1491c195582d927a2c7456a}} takes rotation-equivariant implicit feature embeddings as its global representations. Nevertheless, these learning-based methods often struggle with distribution differences.
| m | 9fac160f597137337e6bef65384e93c9 |
Today, many websites have QA forums, where users can post their questions and answer other users’ questions. However, they usually take time to wait for responses. Moreover, data for question answering has become enormous, which means new questions inevitably have duplicate meanings from the questions in the database. In order to reduce latency and effort, QA systems based on information retrieval (IR) retrieving a good answer from the answer collection is essential. QA relies on open domain datasets such as texts on the web or closed domain datasets such as collections of medical papers like PubMed {{cite:f508a8e636418ab5fcba00ab533369713b775fe4}} to find relevant passages. Moreover, in the COVID-19 pandemic, people care more about their health, and the number of questions posted on health forums has increased rapidly. Therefore, QA in the medical domain plays an important role.
Lexical gaps between queries and relevant documents that occur when both use different words to describe similar contents have been a significant issue. Table REF shows a typical example of this issue in our dataset. Previous studies applied word embeddings to estimate semantic similarity between texts to solve {{cite:755a207a307afbba2c85179918567eeea4a7d7aa}}. Various research studies approached deep neural networks and BERT to extract semantically meaningful texts {{cite:42d908db8a5eaf5d323e99ec6bdaa75d64cbb77e}}. Primarily, SBERT has recently achieved state-of-the-art performance on several tasks, including retrieval tasks {{cite:cf0766ada8301b7c64f5192738ddcdda2d3b88d6}}. This paper focuses on exploring fine-tuned SBERT models with MNR.
| i | 2f7343ab0f802aa871743da2f0c4649d |
This section describes the ADMM algorithm to solve the constrained optimization of the SNAPE method as stated in Eq. 13. The ADMM algorithm has originally been proposed by {{cite:c82c16890dd127db875d3d8a0033ec83634c7a75}} to find the infimum of variational problems that appear in continuum mechanics. The equivalent representation {{cite:8f17c44dad0ab4b5b164d70cd2bdfb63173465ff}} of the optimization problem in Eq. 13 is given as
{{formula:fe1b3ee2-0f7c-44cc-b229-9074433d7073}}
| m | a5d9d7894356fe77bdd7859e5f180a7a |
Related work.
Gupta and Kumar {{cite:137f5d5a87517ea307eade7a99e0a23910b46211}} proposed the geometric version of the SINR model, where signal decays as a fixed
polynomial of distance; it has since been the default in analytic and simulations studies.
They also initiated the average-case analysis of network capacity, giving rise to a large body on “scaling laws”.
Moscibroda and Wattenhofer {{cite:5e411eca46a369f7c281233f6e6d9701dc99763e}} initiated worst-case analysis in the SINR model.
| i | ff3fc393747741c09ccd4b4a89dfcff2 |
It may be also worth to consider the explosion model for EMC (e.g., Sofue 2017) more seriously based on the 3D parameters.
The properties such as an order of magnitude less mass than CMZ, several times larger thickness than Arms I and II, and the round LVD (not parallelogram) at high latitudes, can be explained by the explosion model.
The explosion hypothesis is also consistent with a variety of energetic shocks and outflows observed in radio, X and gamma rays {{cite:87aecf399ac96e277747ec9b1d5bd73515f974d3}}, {{cite:97b184798bcfa4337f8f7dff20737aa014f6ea66}}, {{cite:56de3af8cc216a2fe6cf4d49deb1897234dbd400}}, which are a natural consequence of the feedback {{cite:cca1c564f16f04c6b9ee3e028cdfc6bfe0c0b50e}} of accretion of gas to the nucleus by the bar models.
| d | 5666ee4fc77dd432da4d7a88bb250a68 |
where {{formula:461c08fa-65b3-4228-aac0-cd84b2b698bb}} and {{formula:a7db7fe7-0ad5-43e4-9be9-7eda06017ad9}} denote annihilation and creation operators of the fermionic d-d transitions, and {{formula:231d7ed5-8bb3-48fd-bc28-cd8aa009fc09}} , {{formula:e833e21a-60b2-4c53-ad61-8011472dd8ac}} are operators of the 7.5 THz {{formula:eb951afb-3d82-43c1-9b28-4505d753b133}} phonon field. The first term describes the d-d electronic transition with energy {{formula:27369785-44a9-4fc0-b944-4903bcd5c4eb}} as a two-level system coupled to a phonon field with a coupling constant of {{formula:9818556a-fc53-42ac-a773-77a7c9ad20c8}} . The second term is the total energy of non-interacting phonons of energy {{formula:0ae3490d-321a-4115-b282-6878e1f3a315}} . The coupling constant can be written as {{formula:932dc6d0-f9f3-40d7-aae5-7e8768ac45df}} {{cite:962fe0a88f6bd007bc5bcb71e10ffed17b584a56}}, where {{formula:bb6e4a31-8c41-4c5a-8d8b-96d7a11783b5}} is called the dimensionless Huang-Rhys factor, which quantifies electron-phonon coupling strength. This Hamiltonian is exactly diagonalizable with a unitary transformation {{cite:962fe0a88f6bd007bc5bcb71e10ffed17b584a56}} and the spectral function {{formula:34fd5d12-f402-44da-9d4f-6f09ac0e1414}} has a form of Poisson distribution:
{{formula:0a0d2dcb-65bc-4b1a-b60b-af855264a22c}}
| d | 24230b2d5c4b8b616139e3db701edf3e |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.