text stringlengths 54 548k | label stringclasses 4
values | id_ stringlengths 32 32 |
|---|---|---|
{{formula:0129d02f-63cd-4ac4-9774-4893bedd0016}} is the dilogarithm function {{cite:6701adbb873a88466ef748d006c2f9960a93e25d}}, and {{formula:00cc28e3-a653-4de1-af1e-e872277a1f18}} is the principle branch of the Lambert {{formula:edfb6d89-60fb-4b87-b97f-d4113bb9ee78}} function {{cite:64bea5a6a980c31af52be98c4cb109c37c5777f4}}.
| i | 2e727d87fa8885403d6df70cbf00a5b1 |
(8) The Einstein's equations are supposed to be valid in all reference frames {{cite:dc0685bc1e53f49c526e7fd942170f90a4594c42}}, {{cite:c03a041e73bf3ee325952cb5a79cc4bf79cca231}}, {{cite:3d33c1d4a93c2480767bc1ad25f3656fdb167c24}}. However, in our theory the generalized Einstein's equations (REF ) are valid only in some special non-inertial reference frames.
| d | 05224d2745dffe4b98666f82c0ce85eb |
Relationship Retrieval (RR). This can be further divided into three sub-tasks.
(1) Predicate Classification (PredCls): the same as Predicate Det..
(2) Scene Graph Classification (SGCls) (Fig.REF (b)): its input is a ground true bounding box without labels.
(3) Scene Graph Detection (SGDet): detects scene graphs from scratch. It is the same as Relationship Det..
Zero-Shot Relationship Retrieval (ZSRR) {{cite:679ab465132c2c592985c6041b1b86fd259ce56f}}. The visual relationship that ZSRR requires to be tested has never been observed in the training set. Similar to RR, ZSRR has the same three sub-tasks.
Sentence-to-Graph Retrieval (S2GR). Both RR and ZSRR are evaluated at the triplet level. S2RR aims to evaluate the scene graph at the graph level. It uses image caption sentences as queries to retrieve images represented by scene graphs.
{{table:274924f2-d5c5-40ef-94ea-45977371166d}} | m | f99728e0195f71d3160159fb1977a563 |
where now the {{formula:f9aaf9c9-531a-42b2-a0fa-ee487c495235}} are locally determined for {{formula:1edcc726-1987-4367-9eb4-9965b4dd5229}} even and {{formula:15b4d6dc-3236-4c7b-a605-f5aaaa328628}} , {{formula:5c77d225-3198-4460-89ee-7cafbf9190de}} is locally determined and trace-free, the trace of {{formula:0353a297-9484-405c-99d8-1018fe01d850}} is locally
determined, but the trace-free part of {{formula:7a25778d-634e-4408-9364-7242872176ef}} is formally undetermined.
We remark that {{formula:3157c857-5278-4955-a39c-021fb8d58694}} together with {{formula:8dc9f7f8-4a1e-43e2-9751-1d8180c12b80}} determine the asymptotic behavior of {{formula:855898f6-fe42-48d3-9277-15cc4ca72dee}} ({{cite:f6b6e9ea3c5e0040379bb8fa45f9cfa0ef7d13d6}}, {{cite:9b049013ee960f82af71d05e936f6a3493272ae9}}).
| r | 0e3f6ee9362682b57a48b12f6337b83d |
Finally, CGNN {{cite:40c8c2445afe65c16693cb4db745e0d8c2b3e1ec}} represents the rather novel influence of deep learning methods.
We use the implementation provided by the authors, with itself uses Python with the Tensorflow {{cite:1ae5db3f3a0215c8edf7d8dd30736ece8b483f8f}} library.
The most critical hyper-parameter here is, as the authors themselves mention, the number of hidden neurons which we set to a value of {{formula:e9185ba8-66be-471e-a6d2-848ba32084bc}} , as this is the default in the given implementation and delivers generally good results.
We use 32 runs each, as recommended by the authors of the algorithm.
| r | 72e823f7c002f37c4f65b58196b1fd74 |
In addition to interesting applications such as the randomness amplification, interactive proofs and quantum games {{cite:a0d5177bbf40dd03b459469ea6e2e8d8ee10feeb}}, {{cite:01e622a41354cd243f349a33d43210de8de07477}}, {{cite:ca18427920ab465dfce8882dbf6c0583f0fe21f9}}, {{cite:afcae6b312b6f35f373fab71a88ba0709fd1e6e2}}, {{cite:f90caa0a73d7ab31191680bca09169b214290630}}, quantum networks allow multipartite tasks. One notable problem is to address the supremacy of quantum networks in the case of multipartite interactive proofs or computational complexities. Its improvement may provide further relevance of quantum networks and classical problems {{cite:f90caa0a73d7ab31191680bca09169b214290630}}.
| d | abd1c4e6c4f0b61176244e576d43a34c |
Building on this analogy, the quantities described in Eq. REF and Eq. REF implicitly model an individual's activity space as a circle, and have been used in a variety of contexts {{cite:e6fb302e0246715552b9fe1cab367ebf9d897876}}, {{cite:939c2b1ecf165136203dc74085fa06dfc101b271}}, {{cite:c41d6dc230618047efef190b4706dd5bc0ab9f11}}, {{cite:251673843a0e34e3faed83fc32ab680a6f78c3a3}}, {{cite:17c9076ac1bc97e3ece3fcdacd0402b6a60f5051}}, {{cite:4e4fb2bb16bbea3676fb308949c8ba5386e027da}}, {{cite:0b3a8811856947c1a4f29a32b6ec7bfb0527945f}}, {{cite:bc0958b42f65ad43667dd8d2dbfd2ae2ccc68988}}, {{cite:70e0df8ec2f21ad6c6a06774eda44023beebb86c}}, {{cite:8268e848a2b6188452a9d0b6dbeadc179a48af4f}}, {{cite:8f948129aa0ab7a0961fd6c959b89b326458dc59}}, {{cite:ca65c5ba74a00829a32e4edade45492e314212a1}}. Yet, recent observational studies have documented polycentricity in human mobility, which an approach identifying only one centre and one spatial scale fails to capture {{cite:d4ce9f545767dd90a186de7266b4ddc08ea4b40c}}, {{cite:2adb9f7d0b07aee2ed4f1db08f83277879a578e5}}, {{cite:327d7b2ceec4401e8cc45b8e87f92731bbf5d5bb}}, {{cite:875aed7c983a53750c72e23061ba154737018620}}.
To address this important limitation, the monocentric model can be extended to account for the polycentric nature of individual activity spaces. Specifically, we propose that an individual's activity space can be described as a set of {{formula:e7c45c84-c109-4ab7-b139-791ef02b0f5f}} ({{formula:5c140c0d-2f8c-4244-96fe-2f4f2bc262d2}} ) circles, with centres {{formula:0fa336e5-0502-433c-bf44-44135255db3b}} . Under this flexicentric model, the moment of inertia (see Eq. REF ) can be generalised as:
{{formula:65416fe9-873b-407a-910d-fc7d920ada36}}
| r | af001e0fc8047b3e989bf9205c4b06e6 |
We created a pipeline to make multi-band photometric measurements of all
objects detected in the optical and infrared imaging in the 13{{formula:d301f7b3-a079-447d-9955-1849eb7b8c16}} field.
First, a master catalogue of optical and infrared detections was created using
SExtractor {{cite:2a86d1ca42645b4810c7c026957d091ffb82117b}} to construct separate
catalogues from the images in each filter.
SExtractor was configured to record
MAG_AUTO, MAGERR_AUTO, FLUX_RADIUS and FLAGS parameters for each
source (FLUX_RADIUS records the radius which contains 50% of the
source flux).
| m | d08292779cfbc362d47686b9499e7456 |
In terms of empirical performance, it has been reported that RR outperforms the full gradient descent method for large-scale instances of problem (REF ); see {{cite:662f1bde268ee7f59fec69f23e05ca949fc1d59d}} for more details. Due to such a superiority, RR is used in a vast variety of engineering fields. Most notably, RR is broadly applied in practice for training deep neural networks; see, e.g., {{cite:662f1bde268ee7f59fec69f23e05ca949fc1d59d}}, {{cite:7d92710c3784893fe7dd3405c915a0b72d85a196}}, {{cite:24fef50e77e113f4e62e838527c8ca2fbc7a8c1a}}, {{cite:5da921f684267a67d75e60024564a826119e4b63}}, {{cite:2986ead42265c8a016692f033907c68b2375a2f9}}, {{cite:f6cfd2e8c8e374c68fae49328659f7b4d0211c6f}}, {{cite:945333aa78831098210c7785470d57b2e341e7bd}}, {{cite:aa9d54c6d1c24a5fa95dbea8aef894c9577b498c}}. This special instance of RR (together with the incremental gradient method) is better known as the backpropagation algorithm for the training of neural networks.
| m | c7d76177e72bf380705f4917bf93d2fd |
In this paper, we study a deep neural network approximation for the energy and the underlying electronic fields in a unit cell of magnesium subjected to strain. We generate data by repeatedly solving the unit cell problem and use it to train a deep neural network. An important challenge is the representation of the electronic fields: these are elements of infinite dimensional function spaces whereas neural networks typically approximate maps between finite dimensional spaces. Therefore, we use the approach that combines model reduction and neural networks for high-fidelity approximations of maps between function spaces {{cite:0eb6defed34882a5b9a5de3e61542c7f7c79df5e}}. We show excellent, specifically chemical, accuracy of the trained neural network approximation over a range of strain. Further, the approximation is able to learn the onset of an instability.
| i | 3372602acbc23a485b5209b476221ecf |
The predicted radio SZ has a steep scaling towards low frequencies (see Fig. REF ). As such, it could become an important foreground for standard 21cm fluctuation measurements. The signal is expected to spatially correlate with the standard Compton-{{formula:67949e48-a99f-4d92-bb54-effa2d4f0f78}} map and therefore can be modelled by combining SZ and 21cm measurements. This is of particular interest to radio experiments such as LOFAR {{cite:b3f059e8e5de8f582e7f22c5ec4c918818df4f00}} and the Square Kilometer Array {{cite:86d2489f2a46b77179913426f01b004b2c200a2e}}, and CMB experiments like the Simons Observatory {{cite:6b39bb06a4bd3fed4d7f28be901caea4652e6f75}} and CMB-S4 {{cite:4ac6c0f1ee47cc3739e0750789dbaae1696e6144}}, which all promise high resolution wide-area maps of the sky.
| d | 02c09526bc1a785161503523e0e36689 |
Spin-pumping {{cite:66822dc504ee7b11edb6805bb4000d4cd3f09411}}, {{cite:354033cdff0a487851de8044b4026fb82c019de8}} is an effect where a spin current is injected from a ferromagnet with a precessing magnetization into a material in contact with the ferromagnet. Its reciprocal effect is spin-transfer torque {{cite:1e41d03950779b5a39cad075da6b4c4a29d29cc2}}, and the two constitute central phenomena in the field of spintronics {{cite:65fd8c1383d568fffaf4665f9df517d2ec933c5d}}.
| i | b8ce5a9260129a5d5ba3afd5e8b1b14d |
After an ablation study to select parameters, we choose a few top performers to bench mark our algorithm. On the ECCV22-MIA {{cite:095a2bf1dc28f23e1c47f11e37f353a3bfc33467}} validation dataset, the benchmark results are listed in Table 2. In this table, the F1 scores are presented.
| r | 405057544350635a16d8a39931340661 |
Since {{formula:c208d490-b60e-4f4e-ac3f-8e0dae11a36d}} , {{formula:1273a2f2-7639-45e5-b1a7-a554b0c7f535}} is bounded from above and, hence, {{formula:95e54464-124b-4875-8b1b-bdd36a67a882}} . Applying {{cite:0d6f2f9fed396f9ad298f5ee45d19f93b379bab4}}, it now follows that
{{formula:c2f9ef57-4e41-40c2-9150-154e3d2d8056}}
| r | 0933e0ad11fb7cb30818e745187f13ea |
Statistical methods also can be roughly divided into two estimation paradigms, i.e., estimation by definition and correlation between visibility and feature respectively, and both of them are data-driven methods. For definition estimation, it is to estimate the farthest visible object in the scene. For example, Pomerleau et al. detected lane lines in the image captured by the on-board camera and then estimated visibility leveraging the attenuation of contrast {{cite:ab870954bca068bc760bd1b06a2826e191c423d1}}. For correlation estimation, it is to discover the underlying relationship between visibility and features extracted from source data (e.g., images). For instance, Hallowell et al. attempted to find the relationship between visibility and edge attenuation using the images captured by the fixed-position camera in sunny weather {{cite:18ce4de2b05113d514a12d1676c50d7b43d60b38}}. Following that, so many works have been done in improving either feature extraction or correlation function {{cite:21bd7e030ad8f8e7056672e87edd34ecab3e0747}}, {{cite:2b4b17ce254aa85b887e6eace9c4be7242d8e792}}, {{cite:88a3d3ba200720e1d5c378f9ffa3b1caa71fd9a0}}, {{cite:f6a0ccb6eb56a6590b021af94672dc6726e177e3}}, {{cite:fcfc268fc18a08e2a6ec405547afa8a34dc2767b}}, {{cite:04cae474431847d02fabc2aa4923503a5467913c}}, {{cite:b97b7f819b1b7ee3f5373e3361f2585fd121da64}}, {{cite:8c00ed86e5511804d1f257c35e7302d763f6eb40}}, {{cite:cc67de185798d4e126907c7019d0c761cb16f821}}.
| m | 87218201f8188d0685753f6b5ddf720b |
This section presents the experimental results concerning the DBNs and the proposed approaches, i.e., the A-DBN and the G-DBN, applied to the task of event recognition. All models follow the work of Ng et al. {{cite:fe093175a2ebdb304dc7a5f78873b9399aef6170}}, using 6 frames per video clip uniformly distributed over time.
| r | c1adcca858464fc00fcc55a4b62da804 |
Most of the existing methods learn representations in high dimensional vector spaces. Consequently, the curse of dimensionality {{cite:b441cfeb43e700c83eea091a2b2176750c71ad7b}} leads to the formation of hub nodes {{cite:4ebb18fb146eaca4cfccec240d3e1d2ac7632a7c}} that become nearest neighbours of many smaller degree nodes.
In EA, we observe this phenomenon occurring where an entity gets aligned to many smaller degree entities. For instance, in the alignment results of BootEA on the EN-FR{{formula:90210ab1-954a-407c-9839-270b160ded78}} dataset, the French entity “Beillé" gets aligned to 31 English entities.
To analyze the effect of hubness in the alignment results, we formulate a {{formula:3117a656-97a3-4386-abc7-085e1b8fbdeb}} -score (hubness score) as:
{{formula:519ffabf-fb59-44d7-928e-8f031f43b829}}
| m | 7b317722f7733bc43c91be0e913b33cf |
Tables REF compares the validation errors of ResNet-50 using standalone and hybrid training strategies, respectively. The results are very coherent with the ones of AlexNet. For the standalone training protocol, Our proposed SVD-Padé method improves the ordinary SVD algorithm by 0.3 % in validation error and outperforms iSQRT-COV {{cite:bbc431f3be5d78aab843499fe53030d29cf8a2a5}} by 0.2 %. When it comes to the hybrid training strategy, these SVD variants have achieved competitive performances against iSQRT-COV {{cite:bbc431f3be5d78aab843499fe53030d29cf8a2a5}} and even outperform by a narrow margin. Among all the methods, our SVD-Padé remedy has the best performance and lead iSQRT-COV {{cite:bbc431f3be5d78aab843499fe53030d29cf8a2a5}} by 0.2 % in validation error. We also believe that warming up more epochs when switching to SVD methods can a bring larger performance gain.
| r | 245fd976e581acf74d9ca45f58398b4d |
Overlapped speech detection. Finding good and reproducible baselines for the overlapped speech detection task proved to be a difficult task. We thank Kunesova et al. {{cite:e716e6041e388e7de71a165efa56e1ce4c351f5f}} and Landini et al. {{cite:70b9a908088446f1d696e61349e49633efaffd41}} for sharing the output of their detection pipelines. Results are reported in Table REF that shows that, like for voice activity detection, our segmentation model can be used successfully for overlapped speech detection, even though it was not initially trained for this particular task. It outperforms pyannote 1.1 overlapped speech detection which we believe was the previous state of the art {{cite:ea2c6b46d9e9b123e0268cf4b1561273ad55edc4}}.
| r | aacf847cd78e9b808c00da28151fb1ad |
In Section 2,
we have seen that the propensity score estimation problem reduces to the density ratio estimation problem. Density ratio estimation (DRE), the problem of estimating the ratio of two density functions for two different populations, is a fundamental problem in machine learning {{cite:86f2bab8e92c557e6fc6422726f48211ebd7926d}}.
By partitioning the sample into two groups based on the response status, we can apply the DRE method and thus obtain the inverse propensity scores.
One important method of DRE is so called the maximum entropy method, which minimizes the KL divergence (or negative entropy)
subject to a normalization constraint {{cite:a378d1df51555311bc49195bb8d909138b9bbc85}}.
| m | 9bf92aab7c040184438fbc4557c51d00 |
COCO object detection and segmentation.
In Table REF , we compare the performance of our approach to existing pre-training schemes by fine-tuning a Mask R-CNN FPN model. We followed the training settings in {{cite:ed7db9d7c131b864b3287f3c82cf0536bc279b13}}: the model is trained on the COCO train2017 subset and tested on the mini-val subset. Synchronised batch normalisation (SyncBN) is used in the backbone, FPN, and detection heads. CCOP achieves a 1.4 and 1.3 absolute improvement on AP{{formula:71f8bcf5-2b25-41b5-8efe-4ff745d718ff}} and AP{{formula:09cf98d8-b12b-4b00-a356-bab3cd7f19c6}} compared to the MoCo v2 baseline, and surpasses other recent pre-training approaches including those designed for object-level and pixel-level tasks such as DenseCL {{cite:ed7db9d7c131b864b3287f3c82cf0536bc279b13}} and Self-EMD {{cite:f0757033d8f30637f0c30c93e09420438d700490}}.
| r | e39a4ff8368df65ea5d5c9fa74a76a66 |
This result also has implications for optimization. There has been considerable interest recently in sharpness-aware optimizers (for example {{cite:e2831dfadaacc6f7b94335e918eb5f303f5e7fde}}). This result indicates that minimizing sharpness in the top direction is not possible and probably not desirable, while leaving open the possibility that minimizing sharpness in the bulk directions may still be valuable. We might also speculate on a connection to the tendency of adaptive optimizers such as Adam {{cite:6e3ff0cd78fdf3f11fc1b44911fa2c7fa990f308}} to have lower generalization performance than stochastic gradient descent {{cite:4fe4a24f23334a957829b8dca57179edd38a1267}} and to the instabilities that famously plague the training of generative adversarial networks.
| d | 189da752c1b2f20a908e3fa47b77626f |
The most direct way of observing the impact of the nuclear structure via this method is through comparing observables measured in collisions of species that are close in mass. Isobars, i.e., nuclides having the same mass number, are ideal candidates for such studies {{cite:b19f1ac62534e8e4a7a1e5c4f4828c3fbc5714d0}}, as explicitly demonstrated by experimental data from {{formula:5fcf3591-0798-4c4f-9baf-c44e01870174}} Zr+{{formula:51e6f905-13c6-407f-a663-364609bea434}} Zr and {{formula:f8faa857-df93-45f9-be6c-96e33f26c519}} Ru+{{formula:e8eb5fe4-a5b2-490e-9e1b-9f4ebe56c23a}} Ru collisions, collected in 2018 at the BNL RHIC and released three years later {{cite:53fd6737905da337c05094298e42b0128a7e3a08}}. Given two isobars, {{formula:d02dbef7-7254-42f0-b16a-1991990a3269}} and {{formula:cb593b63-bed4-48c0-86a6-7d3961976f69}} , and a given observable, {{formula:eabd875d-276f-435d-b3e4-08023e04269f}} , we ask the following question:
{{formula:b738775a-afa5-4097-80f1-0a6e0565ed89}}
| m | 7ae07a4f3b469bf46b5694aad418ccb7 |
In the present contribution, we aim to assess in detail the full potential of the nudging approach for unsteady flow estimation in an aerodynamic context, based on URANS modelling and synthetic velocity data. The unsteady turbulent flow around a square cylinder at {{formula:f092967d-5eef-4b8a-8cc8-6dc8f6eeed46}} is chosen as flow configuration {{cite:548a0896bb520e0f0ae3ec88a2c2937c370faf45}}, {{cite:296ecbbcc47b6c713da09375620d616dbc0863dd}} because it exhibits several flow phenomena occurring at various time and spatial scales that are of interest for aeronautical application. They include low-frequency quasi-periodic vortex shedding (VS) past the square cylinder, along with broad-band high-frequency Kelvin-Helmholtz (KH) fluctuations in the shear layers at the top and bottom sides of the cylinder. URANS simulations aim at capturing the fluctuation of the statistically-averaged flow, and are most accurate when the fluctuation frequency, such that of vortex shedding, is much lower than that of the turbulence itself, which is modelled. For the present flow configuration, {{cite:582bf2cf1f65909d9b2e3632bb18e02b7aec39e4}} showed that, while being inevitably more expensive than RANS simulations, URANS simulations significantly improve the prediction of the time-independent mean flow. On the other hand, they do not necessarily capture the small-scale high-frequency fluctuations associated to Kelvin-Helmholtz vortices and, more generally, their accuracy remains determined by the adequacy of the closure model for the Reynolds stress tensor {{cite:d3aef0873b49c5020ce735b0bc92378ae5655b4d}}, {{cite:7b5814a28e7c188bfae66170ad81fd8e4ddc9243}}, here the Spalart-Allmaras model {{cite:bbe315bef4bf4b631dee58b7892243e0714cdcb4}}. The objective of the paper is thus to assess the potential of the nudging method in unsteady flow estimation, not only for the large-scale low-frequency fluctuations that are solved by standard URANS simulations, but also for the small-scale high-frequency fluctuations that are not captured by these same simulations. Synthetic observations for the considered two-dimensional URANS simulations will be generated through a spanwise average of a three-dimensional DNS of the present flow configuration. Nudging will then be employed to enhance the URANS predictions and reconstructing the full reference flow from sparse velocity observations, whose spatial density will be systematically varied. A crucial aspect of the present study will be to correctly reproduce flow structures evolving coherently in space and time. This will be assessed through the spectral proper orthogonal decomposition {{cite:ce791d8eaa82fffa1e961b4fca6a065f2aee4268}}.
| i | ee36cd96ceea78b9b3bbf25e3b9bceac |
Although equilibrium shapes of membrane systems are relatively well understood {{cite:16abc01bdf9ca55f362fd263a4eca3c4b17ab9eb}}, {{cite:9a29b67dcc6ef2ab297f9f8cc9fd2d098f711b7b}}, {{cite:cf98372c3bfeeef6a0aac06ab706707975d227e4}}, {{cite:1e0ca69c851d2d7932a9ab0011921834747d40e6}}, {{cite:5eb709e5ebaad3d5fa2e6b7528d446cefb6b2fba}}, {{cite:27dfd3df23976364a4c65cff275b032a41460746}}, {{cite:c571e48616267ca2c45ae291350c319765cb8865}}, the description of their out-of-equilibrium behavior remains a major challenge.
In this article, we propose a quantitative physical model for the out-of-equilibrium dynamics of lipid vesicle induced by surfactant. We show how the dynamics of the first microscopic pore can be either short or long-lived, depending on the surfactant and membrane properties. The driving mechanism for this behavior is the solubilization of the lipid bilayer, which induces an area reduction of the vesicle at almost constant volume. The progressive reduction of the area to volume ratio produces an increase in membrane tension, eventually leading to membrane rupture, and the opening of a large micrometer-sized pore. Interestingly, two possible scenarios occur at this point (Fig. REF (a)): either the pore closes in about a second after opening (short-lived pore), or the pore stays open for a long time, typically between ten seconds and a minute, before closing (long-lived pore). Then, as area reduction of the vesicle continues, subsequent series of short-lived pores occur independently of the first pore dynamics, until total solubilization of the lipid vesicle is completed.
| d | c888efa7b2b376a969141e8d0eba2034 |
However, as is illustrated in Fig. REF , PSNR is not a good metric to assess the effect of hiding visual information. Fig. REF (b) has a lower PSNR than Fig. REF (c), while the visual information (e.g., the birds and the leaf) are more easily identified in Fig. REF (b) than in Fig. REF (c). To solve this problem, a feature extraction network is employed to measure the effect of hiding visual information. The feature extraction method was initially used to measure the similarity between two images {{cite:f0d045b2d4a2572da25bcc06b147c854e91b13e9}} and then successfully used to measure the difference between two images {{cite:2a306e22297258d4349ddc5452c3fc82e6a6a9cb}}.
{{figure:19aec468-e1f8-4fbe-95e6-59e95fb647f9}} | m | 97c2cf87aea575fd73dbfc9bc2a6e6bd |
A recurrent or large transformer {{cite:8fd7cd2f3b20e47e35effd07d9412bd1560b9d11}}, {{cite:71661f39b686a43095de5788c10058717c48f34f}} network architecture may further increase the performance as the network would be able to infer the rules over a longer sequence of observations and eventually correct for mistakes done by a purely feed-forward convolutional network, which may also enable a deep learning model with stronger generalization as compared to the level observed in the current study. Recurrent or transformer architectures may also expand tractable time horizon for predicting multiple CA state steps.
| d | 44640b0ec888ee1ddea6bd62192ed7ce |
Training with the adversarial loss, alone, usually introduces additional artifacts, and previous works often use MSE loss to induce the learning {{cite:8736b4706318b3f7b069e734ee52733aa091d831}}. However, as shown by {{cite:384267b86bf62597744c2c789f365b5695f93bda}}, MSE loss does not handle streak artifacts very well. Therefore, we adopt the choice of {{cite:384267b86bf62597744c2c789f365b5695f93bda}} by using a perceptual loss to induce the learning and give more realistic outputs. Let {{formula:727b80e6-1865-4f30-b587-ac3821090763}} denote the feature maps extracted by the {{formula:fa49eb3d-dff6-40ae-ba8a-250a3f58092f}} -th layer of the perceptual network {{formula:c63e42eb-724c-47ec-b315-04389d6186cd}} and {{formula:29ca6ea1-147c-4c4b-95eb-97cb3e43a9af}} denote the number of elements in {{formula:4d72e147-82ae-49dc-92d5-43cf0923290a}} , the perceptual loss can be computed by
{{formula:66568825-8ccb-4cba-9b9a-2de23b154d4e}}
| m | 516fc0430c03ea68c91eec6007375251 |
The ConvLSTM cell-based neural networks used here are efficient in encoding {{formula:b7579a3e-895c-4638-8b00-d96d654f9b1f}} spatio-temporal information and representation of features. Equally importantly, FUVAI takes advantage of the attention mechanism followed by multi-scale features in each decoder block, and achieves better accuracy in both segmentation and classification. The multi-scale feature information decoder is vital in the fetal body part segmentation task, given their considerable variations in terms of size, shape and location. The approach of relating segmentation performance to the amount of multi-scale feature representations is also true for 2D convolutional neural network-based methods, which rely solely on spatial features {{cite:aa3d0564d95e00ab081f3623adbbb0413e1a682c}}. For more details please refer to a separate ablation study which we published in a conference abstract {{cite:053103b514bca37ee79fd6a96ae15ad3cfcb11a1}}.
| d | 2109b58a8043d5599b26f1f131ead7a8 |
In this section, we show the comparison of our method RAP and EOT {{cite:2f0bcbfea9698700a1e6ec3ecc8357bb9e17b709}}, VT {{cite:70b431edea6d26354f22bf8048026fc8c7dfff3d}}, EMI {{cite:adb2db914797b7123f3b385dbe616b4de11e9a8a}}, and Ghost Net {{cite:f1e81cbb51d6de043b774f3b0f00cf418c2cc98a}} attack methods.
| r | ec1c65eb48acb48ee7d5efb74c9a907c |
The propagating nonlinear pulses were found in an implicit form, and turned out to be close to Korteweg-de Vries (KdV) solitons {{cite:0832678666e6d796bbd640c16c16dc8b39904193}} in some asymptotic limit; both
numerical studies on the model proposed and experiments verified the theoretical predictions {{cite:1f546c3a8d30363062c191daa6d3356082b0df11}}, {{cite:26ace60c0fb10dea6ebb21f3d0f39947ea4833a5}}.
The same structure was extensively revisited more recently,
with the use of
an analytical model based on the transmission line approach,
and numerical simulations. Both the analytical and numerical results
were found to be in good agreement with pertinent experimental observations of
solitons {{cite:7d932f44b79f136f0765790e8b8604d9d1ae81a2}}, {{cite:a4359a34da7ff29e3e0b943da66a295914b12e3c}}, {{cite:1540182e8101a99f65819f1a390061a73d0c2ec1}}.
Other waveguide structures,
loaded with arrays of elastic membranes
or side holes, were also predicted
to support solitons
in airborne acoustics
{{cite:575d122c52f98515117b2ecffc7bd7cdbaaa6344}}, {{cite:03fab3cfe1790913db3b122e503070e041d36670}}. In these latter works, where the waveguide dispersion
was introduced by periodicity, it was shown that the
predicted solitons were in fact envelope solitons
obeying effective nonlinear Schrödinger (NLS) equations.
| i | 453aec25469805e4f5182c64b54d24d1 |
Rather than relying on the difficult sparse outliers identification as done in {{cite:884b88b1b1e38c2d545860d8588a9c420a94af49}}, the main idea of this work comes from {{formula:a4ca7ca1-2d1c-437c-b2b7-0c2cc44d0af4}} -space random subsampling, which can eliminate sparse outliers in {{formula:89b51f48-5b73-42d8-b026-a52fa43c0ea0}} -space from motions in a probabilistic sense.
In fact, this is ideally fit to the deep learning framework thanks to its close relationship with bootstrap aggregation {{cite:4ce597927131bfe784288f6f374ca76088ae93a7}}, which has been recently explored in deep learning MR reconstruction {{cite:74248c8caa208c202b611c14ce24c91e73dcdf29}}.
In particular, if a neural network has been trained using motion-free training data as a reconstruction network from undersampled {{formula:6986dac2-4cef-48ef-af19-2c14f2049bf1}} -space data, the bootstrap subsampling in the {{formula:3e42ff3d-7ba3-461f-8376-f25636f61e4c}} -space domain can eliminate some of the sparse {{formula:86172a15-4ff3-4fea-a9f1-e114b4a1ce3c}} -space outliers from the motion, so that corresponding neural network creates images with fewer motion artifacts.
Then, during the aggregation step, the loss of image quality from the subsample data can be restored.
To deal with the potential bias from the use of the clean data for training, we use the unpaired training strategy using cycleGAN
for the reconstruction of high quality images from undersampled {{formula:9775f8fe-badb-4cfe-9e72-3884234ef60b}} -space data.
Since the neural network is trained using only clean data without motion artifact simulation and acquisition, our method is so flexible that can be applied to correct various motion artifact from both of simulated data and real data.
In particular, we successfully demonstrate that the intricate motions artifacts from TSM in Gd-EOB-DTPA-enhanced MR can be effectively and robustly removed using the proposed method as will be shown in experiments.
| i | 01a5f6322cf7e70cd0d6b7e5590834c8 |
Let us consider now MWI. The main difference between this interpretation and the typicalistic formulation is that the latter assumes the position of the particle as the primitive ontology, i.e., it adds the sample space {{formula:e95f2d11-9f83-40cd-a097-0ca1f9c237d4}} to the mathematical formalism of the theory. On the contrary, any ontology in addition to the wave function is rejected by the supporters of the MWI, even if in some formulations the existence of some privileged degrees of freedoms like position or momentum is recognized (see for example {{cite:335497c0fd0c51c02dcc718e4da055651b7ae834}}).
| d | 621fecd26e2ecc18a624a172655446cc |
BLE is Laplacian Eigenmaps (LE) {{cite:14c7850d817e4ab94e640b3858e33352a71f5252}} with the most informative view, i.e., one that achieves the best performance with LE.
BLLE is Locality Linear Embedding (LLE) {{cite:3dad87d0e79fec3f425dc1e94c5b83daba6b9a24}} with the most informative view, similar to BLE.
MSE {{cite:38fa267f4c3b530c777ce944fff9cc10f4dd2849}} is a multi-view spectral method based on global coordinate alignment.
CCA {{cite:962a6ced503df33d04ec5097179e9c28d43531be}} is used to deal with multi-view problems by maximizing the cross correlation between two views.
Co-reg {{cite:9460aae8407476e2831f733059604cc807dcad3c}} is a multi-view spectral embedding by regularizing different views to be close to each other.
AMGL {{cite:4bbbc4c9a43227c9ba55653c6a392129b63ff47d}} is an auto-weighted multiple graph learning method, which could allocate ideal weight for each view automatically.
| m | ce903fa66aece5880cbd7b3bf80713a2 |
Another interesting direction of future research is assessing the power of quantum processors {{cite:1f76bd756a31b98333a82a15f06ead6442e5547a}}, {{cite:547cf701e926b57185e8ca16c1ead7389bb8e8d8}}. It has been often pointed out that the number of qubits in a quantum processor is not an appropriate measure, because the presence of noise generally limits the set of achievable computations. Hence, it is important to have more inclusive measures, that take into account not only the raw number of qubits, but also the size of the set of operations implementable on them. The approach of quantum gate compression sheds light into this problem by providing a rigorous notion of “effective size” of a quantum processor, defined as the minimum number of qubits on which the operations of the processor can be compressed with sufficiently high fidelity.
| d | e94d08189d09227c48d337e2d152f41c |
The resulting dataset contained a total of 4259 sequences of length up to 910 characters, with vocabulary of 21 amino acids (A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, X, Y). We use 60/20/20 ratio to create train/val/test split, resulting in 2548, 876 and 835 sequences respectively, maintaining the positive to negative ratio in each split. A LSTM model enhanced with a single head attention mechanism was trained on this RdRp sequence dataset to predict their zoonotic potential (as a binary classification task). The model architecture was comprised of a single encoded layer with single head attention {{cite:d2a0027a16e0dc42d0e0e31d51c9d8cd54c52631}}. The model accepts in the sequence of amino acids {{formula:06c411ac-b67c-4e26-a0ce-c4283986666c}} and learns an embedding {{formula:1f6dd2f6-3e44-4aa2-942a-959699e9cf48}} , which is passed to a dropout layer followed by a RELU and a softmax layer. Best test accuracy of {{formula:b9527230-661a-455a-98d7-a92d9cb4f2f1}} was achieved after hyper-parameter search. More details are in SI section REF
| m | 524438524b84b0b640fa4efd8816bc2e |
The function {{formula:b8f68d2a-612a-4c92-b06c-83048f0f49ac}} is concave, positive, and satisfies {{formula:69f2b82f-a158-4a54-8e3a-a2945491e930}} .
Assume w.l.o.g. that {{formula:34075adc-cdc9-4721-92eb-be92dfcdbf4d}} . The proof for the case {{formula:fab0539c-4938-4f6f-8c30-22db01637643}}
was explained in {{cite:38452e041be8a106fd10d13c6c6ab80ef733cc93}}: The chord
of the function {{formula:46b9a64e-cac8-4f49-81b7-3cfe2bdcb442}} from {{formula:6b31703e-b3be-4273-b6c3-24ccd57fdede}} to {{formula:bde148f4-5f61-4867-b6bc-de71192c54f1}} has maximum absolute slope
either at the extremes – either at {{formula:fa2cfe0e-138d-4b6a-bc6f-7cdaca7fa8d6}} or {{formula:581bf48e-2592-4601-9dd6-0a280ec14118}} . Then,
{{formula:d94d707f-2c8c-4e3a-bb34-52eb44413467}}
| r | 7693cdb00e869baa889e469e8c277cd2 |
(iv) We find that the proposed geometric alignment objective is beneficial even when the target sample size is limited or when the source model was obtained via robust training {{cite:b8f3ddb8fb38282e97d2cbfd9da752c14b4994f3}}.
| i | 3bd3cf09b81504465edf282d6412b63f |
The fake news classification is not limited to the linguistic analysis of documents. The studies analysing different kinds of attributes which may be of use in the same setting {{cite:f4b70daeeeadab8d4a501dfbddd485d6eb5b5be9}} are interesting. Amidst the typical methods that the study comprises, there are the analyses of the creator and reader of the message, as well as the contents of the document and its positioning within social media outlets being verified {{cite:d2cdb44ff7958659e5a0c3edde87a349d4b25cac}}. Another method which shows promise is analysing images; this approach concerns fake news in the form of video material {{cite:2f8867d755e0435d9d259b5519d00748cf9a2434}}. Similarly, the study by {{cite:f06916c7b0edc1b01f9a35573cd4ef56fa00c9cb}} is evenly thought-provoking; it proposes to divide the methods of linguistic and social analyses. The former group of models encompasses the semantic, rhetorical, discourse and simple probabilistic recognition ones. The latter set comprises analysing how the person conveying the message behaves in social media and what context their entries are building. Then, {{cite:73a3db55deac3def05ac9fbe2bd41654087ee040}} has based the design of the recognition models on the behavior of the authors, where the background of the post (both the posted and forwarded ones) does depend on their bodies, and at the same time refers to other texts. Diverse representations of data were analysed by {{cite:95a4ebdd261def6ed76a982a10b3e7b8cc2ef571}}, whilst {{cite:0625bb67d1d7b9697c497b44000c796e889612d8}} has examined various variants of stylometric metrics.
| m | 485f582b7c6f98c4fc8d2e4c0a989a84 |
A rigid body {{cite:4797bb41e2011c647c1851ed820ba5fb0b3ed481}}, {{cite:0e4fe96057098a3f5485987e83f45624fbc3b64d}}, {{cite:07b2a660f66e35004d52676f9d690ccd3a1bcd30}} that conducts Brownian motion can translate and rotate in space. Most interestingly, in scenarios where the particle is screwlike {{cite:f41efc01efe460077c41945b0cfb3abb66875ad3}}, {{cite:1e311f53650595026d81327d22677f5d61f09a7b}}, L-shaped {{cite:3f600dbc296a21f67db8ec11e21067f1d4f8237a}}, biaxial {{cite:4d1d5f89ce6084a4fb70ef4173c73f00479316e1}} and ellipsoidal {{cite:ffbf074e3fd0603e040013889183815e8bb9146a}}, etc., translation and rotation can couple {{cite:f41efc01efe460077c41945b0cfb3abb66875ad3}}, {{cite:1e311f53650595026d81327d22677f5d61f09a7b}}, {{cite:93c00dac568515dab259ae3613069dacc15412ab}}, {{cite:a75f03e79c79324bb912d2353f50dae39e3bd727}}, {{cite:98bfc5b8eb0444e0f7b981c087a8eca03b99fb37}}, {{cite:c9f0696c2a0eca0759ac5e12838ede0c812f15a2}}, {{cite:ffbf074e3fd0603e040013889183815e8bb9146a}}, leading to a rich class of trajectory patterns, e.g., helical motion {{cite:4d1d5f89ce6084a4fb70ef4173c73f00479316e1}}, circular motion {{cite:3f600dbc296a21f67db8ec11e21067f1d4f8237a}}. In recent years, apart from exploring these novel dynamic behaviors arising from rigid-body Brownian motion, there were general models built, such as Brownian Dynamics {{cite:4805f21f7b2797d4a5b1ef4fdd408a5f2594c13b}}, {{cite:0e4fe96057098a3f5485987e83f45624fbc3b64d}}, Stokesian Dynamics {{cite:a33db5451361e746d1e416abfa3a5347fe0d1659}}, {{cite:4ad59cc5120c8fc27eb0969fdab4233dc3bcfd57}}, Fluctuating Hydrodynamics {{cite:3610cfdcdc64fffb7ca754c581556ae3fc8623c6}} and the Langevin dynamics of arbitrarily shaped particle {{cite:93c00dac568515dab259ae3613069dacc15412ab}}, {{cite:07b2a660f66e35004d52676f9d690ccd3a1bcd30}}. Usually in these models the effects of non-stochastic factors such as hydrodynamic interactions and particle geometry enter into the displacement equations as resistance tensor {{cite:93c00dac568515dab259ae3613069dacc15412ab}} or equivalently the mobility tensor {{cite:07b2a660f66e35004d52676f9d690ccd3a1bcd30}}, while stochasticity is contained in force and torque terms. Subsequently, the trajectory of a tracking point (TP) on the body, e.g., the center of mass (CoM), or the center of friction (CoF, also known as center of hydrodynamic stress {{cite:a75f03e79c79324bb912d2353f50dae39e3bd727}}) is generated. Hence the particle is still represented by a zero volume TP rather than a finite volume body.
| i | eadf53a4eedddd5a541c40261492ba9f |
where it is assumed each {{formula:5165d49c-6d8c-4186-b8e1-9378ac7cac5a}} has some structural redundancies for example sparsity or low-rank. These methods have a subtle yet fundamental difference from CL, as in CL the structural assumptions which are exploited to form the CL sketch arise from the model or distribution itself, while in these streaming methods the structural assumptions come directly from the data. Moreover, several passes of the data may be required to reduce the low-rank approximation error {{cite:e62a711d8498b5ec2571a4175f4e95911db18a8a}}.
| m | c5d9463fab485d8f1a3c1d5a9af47f26 |
The simulation data presented in the following sections are also used in {{cite:b403083282d8e6d64ec9a46c14048a4bba9eb047}} and {{cite:c4e432d77004675df362f05d9007c305b60e0209}} (2015a). While {{cite:b403083282d8e6d64ec9a46c14048a4bba9eb047}} use the technique of Principal Component Analysis (PCA) in order to statistically analyze the turbulent flows in the spectral data cubes, {{cite:c4e432d77004675df362f05d9007c305b60e0209}} (2015a) evaluate the slopes of the structure functions of 2D projected centroid velocities and compute Fourier spectra. Here, we summarize the most important aspects of our hydrodynamical simulations and the radiative transfer post-processing.
| m | 6d86e38a50dd949d19fcbdbe4b3c00ba |
This work is not free of limitations. In contrast to randomly exploring the search space of CNN architectures (e.g., through automated hyperparameter optimization), we decided to strategically analyse in two directions: model-centric and data-centric approaches. For the latter, we decided to try four different variations, yet, want to highlight that a manifold of other data-centric approaches are appealing. For instance, to improve the DeepSMOTE results, a further increase of the capacity of the encoder and decoder architecture could potentially improve the degree of details of generated images. Alternatively, generative adversarial networks (GANs) {{cite:e793429f1fb244703f8042297a88b742aaca7e7b}} might be a valuable approach to achieve a higher degree of details. Similarly, we did not evaluate all feasible combinations of our data-centric and model-centric approaches which delimits the generalizability of our findings to some extent. Furthermore, we need to diagnose our models to verify their robustness, for which we intend to use several XAI techniques {{cite:fd61b09fad96b3325ce88232c00d527d0839de6b}} in subsequent studies.
| d | edc23a0d020717fbcdecd5329d6586cf |
In contrast to the existing one-dimensional (1D) time-domain or frequency-domain modulation techniques, such as orthogonal frequency-division multiplexing (OFDM), OTFS is a two-dimensional (2D) modulation scheme, where the information symbols are carried over 2D localized pulses defined in the delay-Doppler (DD) domain.
In fact, each information symbol multiplexed in the DD domain is spread across the whole time-frequency (TF) domain, allowing the potentials of exploiting the full diversity to achieve reliable communications against various channel impairments.
In addition, the channel in the DD domain exhibits various beneficial properties, such as separability, stability, and possibly sparsity {{cite:e3f6b11575945cf8c1df403da5394365b53c0dbe}}, which can be exploited for efficient channel estimation and data detection.
In particular, processing the received signals in the Doppler domain of wireless channels allows us to separate the propagation paths experiencing an identical delay.
Besides, OTFS modulation can transform a time-variant channel in the TF domain into a 2D quasi-time-invariant channel in the DD domain and thus relieves the impacts caused by channel aging {{cite:692638117050da34443f0d4e98abbcb9c12161fa}}, {{cite:c9219659083cf54c9aced10896fbc792b95d1c87}}.
Additionally, the DD domain effective channel is generally sparse in an open-space rural propagation environment with a limited number of moving scatters.
As a result, the coupling relationship between data symbols and the channel in the DD domain is much simpler than the counterpart in the TF domain, which is crucial for accurate channel acquisition and efficient detection algorithm design.
| i | e547f22ae33b6006f7a118090918dbb5 |
Tumor dynamics modeling is one of the most common model-informed drug development (MIDD) approaches in oncology {{cite:7509d302b6a3a5136fc5022a7aa4733d4d64f603}}, which could enable significant insights into the heterogeneity of inter-/intratumor and cancer evolution. Understanding the drug resistance evolution of tumors could be the key to achieving personalized treatment and precision oncology, which are vital quests in cancer research. Mathematical modeling is one of the most dominant approaches to quantify such evolution, which has been studied since 1970s{{cite:2ca1bc0f24fc6da67d632144f90bbbfbb2b2d810}}. It is worth emphasizing that these methodologies laid the foundations on classical physics. Modern society has witnessed an emergence of Artificial Intelligence (AI) and Quantum Computing hardware, which gives rise to Quantum Machine Intelligence. On the one hand, AI has penetrated every aspect of life: it can be fit in palms of hands as Machine Translation AI{{cite:90d97f248c38cd2d12309893ecd5f8441154badb}}, it can auto-operate on the street as self-driving vehicles, or it can approach efficiently (but not yet solved) the problem of protein folding{{cite:ea88213a066440120497e7bd96b817cdf5b0f5ae}}. On the other hand, the recent advances in Quantum Computing (QC) enable Quantum Simulation (QS) on conventional computers empowered by Central Process Units (CPUs), Graphical Processing Units (GPUs), or Tensor Processing Units (TPUs). As a result, Quantum Machine Intelligence or Quantum Machine Learning (QML) can be studied within computations of classical computers in the hope of creating algorithms that can make reasoning with powerful calculations of Quantum computers. These algorithms are so-called Noisy intermediate-scale Quantum (NISQ) algorithms{{cite:171254ab899c73841daf7d2c5efda3c9c7212920}}, in which quantum processors deliver the algorithmic solutions in the NISQ era. The surge of Quantum Machine Intelligence enables computational tools that offer us a completely new insight into cancer evolution via quantum mechanics.
| i | 04f7f1d5ec6d814e696a9156a234d968 |
Once it has been well established that a flux of high-energy neutrinos of cosmic origin exist {{cite:f5711156feb5aacba22c61aa5efd66ac763a85a9}}, the next step is to disentangle it and identify which are the sources. The most recent all-sky searches performed by ANTARES {{cite:2da7dbd26a933657561f3c0e031a2db6f4b7aa12}} and IceCube {{cite:ecc4312eaeccf6e771feff76f4cfd6dd275b4aea}} did not reveal any significant detection above the discovery threshold of 5{{formula:958dd2a1-4f71-4a72-9c6e-3b46f6652830}} , with the excess near the galaxy NGC 1068 observed by IceCube being the most interesting spot with a post-trial significance of 2.9{{formula:df883fe1-a8f7-442b-bec3-92d5e2896673}} .
This type of high-energy searches benefit from multimessenger astronomy thanks to including the sky coordinates and timing information from potential cosmic messenger counterparts. That was how the first evidence of a cosmic neutrino source, TXS0506+056, was found.
| r | f9bdf2a0568225f9f6d52d8c63ddbd8f |
We evaluate following baselines to compare the performance of our proposed model, BloomNet: VDCNN {{cite:ebde1d7445dc97c48e8748956bf752ca56caf9e1}}, LSTM {{cite:536cddc9052abe605d9ed42b44142e8180991a72}}, HAN{{cite:4cd8b7b2262e01f14a487b34931754a0a52b4d3b}}, CNN{{cite:18d710942d10b635a6c73d419caadfef885482b8}}, RCNN {{cite:df080d51d18105f6336b236ee492e93abfc72322}}, Seq2Seq-Attention {{cite:4ea586df0a1afe3b8c7875617b0dace2c740ef31}}, Self-Attention {{cite:2471c868e68a4fbaa0bc623421921aaa09cf2878}}, Random Forest {{cite:294394684ce2dd329465879a0d5a3352653db899}}, DistilRoBERTa {{cite:3466f5fc2f47db1b51fdeb93d3f57cdbe9fcad13}}, and RoBERTa {{cite:3466f5fc2f47db1b51fdeb93d3f57cdbe9fcad13}}. We find that BloomNet outperforms all the considered baselines on the two datasets, demonstrating its higher performance for text classification. Table REF reports the performance of BloomNet and all the baselines.
| r | cd447e74b9b6dfb054a4b016c004de51 |
Size and {{formula:d684f977-eb0c-4dd3-b100-58cda9f8f9f3}} -perfectness of the output: As stated earlier, size of a matching plays a crucial role in real-world scenarios.
In {{cite:2bcc4405d80231f83579ac3f800afcfd8ea0ae0d}} the authors mention that the administrators of the Scottish Foundation Allocation Scheme were interested in larger size matchings at the
expense of allowing blocking pairs.
In applications like school-choice {{cite:907d49f70d0245e6d0cde8397f1eb6c860d71e59}} every child
must find a school. In case of matching sailors to billets in the US Navy {{cite:38881a6e97f58587b91b06b2b3aeaabefdc1fe40}}, {{cite:7dfd974d9d2411d10b48a67140df8e81e7f92ac2}},
every sailor must be assigned to some billet, apart from some additional constraints.
The {{formula:d8c72af8-0747-416a-8f70-52f935f14cd2}} -perfectness requirement is imposed in the two-round school choice problem studied in {{cite:9cad4bba46e941609aa6a59b87eec0f31be9d734}}. Round-1 is
the standard stable matching problem; whereas
in round-2, their goal is to match all agents in a particular set derived from the matching of round-1. We formalize this connection in Section where we generalize the stable extensions from {{cite:9cad4bba46e941609aa6a59b87eec0f31be9d734}}.
In {{cite:9cad4bba46e941609aa6a59b87eec0f31be9d734}} they consider a variant of {{formula:eca54ffc-3c04-4790-89eb-059213b966b8}} problem (Problem 33, Section 7)
and state that their problem is {{formula:e5d43ab2-c1d4-47cd-aaac-927bd50ef12a}} -hard. However,
they do not investigate the problem in detail.
| r | 71c33567e61b6b85534b35e4b743850a |
For each {{formula:efb0c367-dbfa-45be-b27c-9409147bb6c9}} , NMF is solved for a series of random initial guesses (typically, 1,000 or more) for {{formula:cf2537dc-7bad-4f37-b275-7204dbf1b837}} and {{formula:e27cca75-5fd4-4e10-a429-cef775acec4a}} matrices.
The least value of {{formula:b37e1ab7-53a5-4f68-bc9c-ce658df68673}} for a given {{formula:212df181-49d7-41b0-b785-ddc361936b99}} is assumed as the best value for the reconstruction error.
After completing the NMF process, the columns of the 1,000 estimated {{formula:67f3b3b5-432a-4695-806e-a9f39803a2db}} matrices are clustered into {{formula:4e313831-6834-4e2b-aabe-8f02cb216e68}} clusters using a customized {{formula:69075ae4-9b13-45b1-88ae-adf262f5497e}} -means clustering.
Alternatively, we cluster the rows of the 1,000 estimated {{formula:2122c77c-952a-4662-b44b-ace587f8b7b6}} matrices.
Typically, we cluster the smaller matrix.
However, {{formula:88a60d49-5dd3-474c-88a2-4f7819acb5ae}} is also unknown in the {{formula:7a149c6b-79c0-49e4-99a4-2b33d946a2f0}} -means clustering.
The algorithm consecutively examines a specified {{formula:b03aa475-72bf-4722-af8d-548657e57b04}} by obtaining 1,000 {{formula:a90158a9-318f-4435-a499-c6e0bdbcb1b8}} matrices for each feature.
During clustering, the similarity between two clusters is assessed using the Silhouette width/value {{cite:fc8749e89889d0b602901fa4afe268ea0c80d333}}, which is essentially the cosine norm:
(p,q) = 1 - i=1n pi qii=1n pi2 i=1n qi2,
where {{formula:7b3f2de1-22a1-4306-bec6-d3ec08c78374}} and {{formula:8b676c8c-9fc2-4a7d-b545-e31a7f6a9a91}} are components of vectors {{formula:ee555fd9-6d70-47e2-b121-11e4731016a4}} and {{formula:995500b1-c2e8-402b-9004-8b2de6e44d1f}} .
The Silhouette value quantifies how similar an object is to its own cluster compared to other clusters and varies from {{formula:83e92e95-3bf2-4923-9c59-c3914b5bffe5}} to {{formula:08850225-6d01-4551-8148-d59295286f5e}} ; high values indicate that the object is well matched to its own cluster and poorly matched to neighboring clusters.
The combination of reconstruction error ({{formula:ef92e7d3-1d21-4b6c-9a22-e4f56d6976c7}} ) and the Silhouette value are used to determine the optimal number of hidden signatures.
If {{formula:6df578b1-bcf2-4594-b911-7110bf9563c7}} is low, the Silhouette value will be high, but so maybe {{formula:aca91acb-f931-4cfe-993d-3fd11f564c41}} because of under-fitting.
For high {{formula:09fddafa-6910-43ac-a614-8edca426a546}} , the Silhouette value will be low and the solution may be over-fit.
So, the best estimate for {{formula:37ad85cb-9763-4d13-a931-7ac8cade3cf9}} is a number that optimizes both {{formula:3ae39ace-f890-46a8-a8d3-7fe4daff57d2}} and the Silhouette value.
| m | 864bfef29634fa681f59b1c5e4763ec1 |
In this paper, we have shown how to classify the SMEFT operators based on CP symmetry by means of the Hilbert series techniques.
We successfully reproduced the same enumerations as those by {{cite:4acd581d8baed25426df954c427f822447dd12f6}}, {{cite:fcbba872ce16d3599f5930c8e2d7d75bc70b986b}} for dimension-six operators and pointed out a misidentification by {{cite:5f694d6c5111106c4b7d55111d6400dc197a91b7}} for dimension-eight operators.
Our Form code can output these results in a few seconds and can list higher-dimensional operators quickly.
Our method can be easily applied to other EFT theories besides the SMEFT, such as QCD EFT or the SMEFT with gravity.
| d | 40c432e1e46bb5d3f9af9a78a9d0b1ed |
Non-rigid structures and various shape make it more challenging for monocular 3D detection to accurately detect “Pedestrian” and “Cyclist”. Most previous methods {{cite:6bfae597b3331c2ade4e5f11cd8e5f6a602e706e}}, {{cite:07bf9198cf6c1160ec33931e35d059f504e9c25c}}, {{cite:55eaed45a7bff1b4487e6a761c58ee47cf07faed}} fail to demonstrate these two categories results, however, we report results on these two categories in Table REF to show the generalization of our PCT. Following {{cite:27eff2385bae6a1d8d7aefb3225d1010d465fb43}}, we demonstrate {{formula:93262f53-3241-4b38-8b91-b7a809fc6079}} 3D object detection results on KITTI validation set at IoU = 0.5. As illustrated in Table REF , our method still outperforms the base method PatchNet, benefiting from the more accurate localization and complementary global context information. Besides, we also achieves a better performance compared with state-of-the-art pixel-based methods {{cite:27eff2385bae6a1d8d7aefb3225d1010d465fb43}}, {{cite:62386f4b49a2b5c5b6d29948943369b30d9b4cd8}}.
| r | eccc8c4aeb2b74b60e93e8af17721d1d |
Good side effects. Deep classifiers would inevitably misclassify some inputs when tested on natural corruption data{{cite:203c6744700c8b6fa597dda0590940cbb68f6e76}}. Although no intentional perturbations are exerted on these samples, our MixDefense would still identify these misclassified samples as AEs. This is because there exist contradictions between the semantic meaning of the samples and the predicted labels, while the SP-defense layer ConraNet is designed to detect AEs with small perturbations based on such contradictions. See Section for more details.
Though wrongly recognizing these misclassified samples as AEs, rejecting such samples would not decrease the accuracy of the target classifier and may even improve the classification accuracy {{cite:9d374395dd1e8090ec5c48720cdeb8be098dfc7b}}.
| d | 1f0020d78aa69faef543cf630373835d |
it can be used, with adaptive procedures, to reduce the number of degrees of freedom where not needed because in certain parts of the domain the error is already under control;
it can be used to generate a hierarchy of (nested) coarser grids starting from a fine mesh of a complex physical domain of interest, in order to employ them in multigrid solvers {{cite:0493a4ede2cdb4b4d23bc44eb008974ac23c99c7}}, {{cite:3e65d002bea2b3a18322407f66e0f2f83df91827}}, {{cite:54c71ecf9d18bf05df86c78638871fa61c9929a3}}, {{cite:f5646e3000d40be003b034ed6d3190fae27e00fb}}, {{cite:e0247c81ce4ea112d9f52c22fb67d9f7feca92ef}}, {{cite:8b11331a4281f73a26bea87c7fbb51b0685e9dba}}, {{cite:5372654f376ed1b770395d6aced6a1e4f150d100}} to accelerate the converge of iterative algebraic;
it can be employed together with domain decomposition techniques {{cite:496eb040e4b48826e2375c00acbebd4eb4bc8f10}}, {{cite:9d538f741f08c426b56f6aab6def9b779738ba87}}, {{cite:71877f840d50bc7c2900cbd2b29c5e527b4e8108}}, {{cite:6ac3def1058f2e8e99be5b0bb1bd511450c4e3f1}} to obtain a meaningful decomposition of the domain, starting from a fine discretization.
| i | 62ad27f2b4a9bf12be4bbd5d1b2b288d |
We present a new problem setting: user-guided modification of the geometric rules encoded in a pre-trained generative model. To edit a model, we ask the users to move a handful of control points as warping examples. Exisiting GAN inversion {{cite:da6632296427b5840157724806360a5bd96812e9}} and few-shot GAN adaptation {{cite:99fe02a8c0010991295cb4bd8aa314e1939dc2c3}} methods fail to tackle this new problem: the geometric changes can go beyond the learned distribution, placing the goal out of reach of GAN inversion; and given extremely few user inputs, it is difficult to adapt a GAN to the new domain. To address these issues, we directly optimize the model weights with a reconstruction loss between original samples and their warped versions. We introduce an augmentation scheme based on style-mixing {{cite:b2858512d641e724c72d34bf60dec29cf02b5a53}} to improve our model's generalization to unseen samples, and we show that we can update only a small subset of networks weights to rewrite the geometric rules. Moreover, by a simple linear combination of edited model's weights, we observe that the edited models can be composed to build a model with combined shape-changing effects. Based on this finding, we present an interactive interface that allows users to build new models by composing different models.
| i | 87519e5a7278db8c336a16f9e16b1eef |
First of all,
We show that the pattern of linear correlation of accuracy claimed by {{cite:492fb2fa634d0fcadd07a26ff901b23e1a8ad306}} does not necessarily occur even when the correlation patterns we observed, such as diminishing returns, are probit transformed.
The results of the probit transform of the scatter plots shown in Appendix REF are shown, following the method of {{cite:492fb2fa634d0fcadd07a26ff901b23e1a8ad306}}.
This may be due to the fact that our experimental setup uses a larger number of data sets and takes IRM and other factors into account.
| r | ea9ff7c3655bac27e11e489e4e5c3e5e |
Second, a widely used initialization for gradient descent is to randomly and independently draw the entries of the initial gradient descent estimate from a Gaussian distribution, cf. Xavier initialization in deep neural networks {{cite:8c3515534b4605ccdc52b9e3decab32bcfdaa53e}}. Compared to the initializations we considered, numerical simulations suggest that random initialization converges faster but exhibits a more complicated behavior which heavily depends on the initialization variance {{formula:a4803938-2fac-4663-a8cf-1a507ac65122}} , i.e. {{formula:5386100a-23bc-4d69-826d-6a8e5239a68d}} where {{formula:c5d399b8-23d1-489f-bf4b-265c8d038d35}} are i.i.d. with {{formula:1e83bfac-77a2-4783-b6f9-7402cafe80e0}} and {{formula:c1654134-f70a-47ad-bab3-6f153217266f}} . In our experiment we take {{formula:b595ac57-47ed-478e-bd03-1508279b3fd9}} to be Gaussian. For instance, Figure REF shows that, for {{formula:1adffef5-85f1-4823-ae0c-881fb75f0935}} large, the order of eigenvalue approximation is not determined by the eigenvalue magnitudes, thus indicating that the above described phenomena of (effective) rank approximation do not hold in this case; in contrast, for {{formula:eb518aa5-7d84-4f77-871a-e003340300c7}} small, we recognize the dynamics to be similar to the one of our perturbed initialization. We assume that those observations can be rigorously stated and proved in appropriate probabilistic settings, but for now we will leave it to future work.
| d | 1953013172a411a70f8e8e11148d3790 |
Recently, StyleGAN architectures {{cite:646db17238341109b56b66a5e0dc3b309aba5f5b}}, {{cite:cb58cbecd364ab3d2235f979b1bc75cc1fb38d8b}}, {{cite:92ec9894fd19285a8ef3957a3415f4aa7c2efc3d}} and their variants have been verified effective on various facial generative tasks, including face attributes editing {{cite:4cb12433924036e8cde5e58702feea3300990d97}}, {{cite:faedb112efe2f5a8a5f7eaa66f6afe5e91755438}}, {{cite:e9a580101d3399a845a1687ce2ede0fcdd8a7d79}}, {{cite:f94361cd5d9c9b1a61670b09c98ebd1c324b96d9}}, face enhancement {{cite:9a7034e38eef389c4e51b1dc5e1d13800c29c07b}}, {{cite:998d9241b32d153185e8904b82ef070e28220224}}, and even face reenactment {{cite:8ef0a59ff582f4ae1dabf8267148f44b144ac94a}}, {{cite:544a298bca68a5d4dbde714af810716cea3ec48d}}. It is owing to style-based generator's strong expressibility and its advantages in latent space manipulation. But the exploration of such architectures in face swapping {{cite:c540c9fe6269c754cbe939fd33d48bf950b27d1c}}, {{cite:c2406b72658b06012dfa81becd1c5c3dfff75683}} is still insufficient.
Specifically, the lighting conditions are greatly condemned in Zhu et al. {{cite:c540c9fe6269c754cbe939fd33d48bf950b27d1c}} due to the limited distribution covered by the fixed generator.
The structure of their feature blending procedure is also designed in a hand-crafted and layer-specific manner, which requires complicated human tuning. Concurrently, Xu et al. {{cite:c2406b72658b06012dfa81becd1c5c3dfff75683}} aggregate the StyleGAN2 features with another designed encoder and decoder. Thus, a natural question arises: can we avoid tedious layer-by-layer structure design by adopting a versatile style-based generator {{cite:646db17238341109b56b66a5e0dc3b309aba5f5b}}, {{cite:cb58cbecd364ab3d2235f979b1bc75cc1fb38d8b}} with only minimal modifications?
| i | 58530ba358fd67b7c7a3bb73bf8d7475 |
In this section we present the analysis of the
data-vector {{formula:307f0540-3db6-437e-99ab-0df35b1ba9f2}} ,
where all statistics are in redshift space
and include projection and coordinate-distortion effects. This set-up thus fully matches an analysis of a realistic
galaxy survey such as BOSS {{cite:b555aff5b09233ce767db349e88d861fcc12a8e5}}. As for the power spectrum, we expect that the addition of redshift-space distortions (particularly the fingers-of-God effect {{cite:f7537e369db79600b3e63aab9adee57b59a55d91}}, hereafter FoG), will reduce the non-linear scale, thus it is likely that {{formula:9fe84176-4b2d-45c7-b593-f1b18f9c9f0e}} , and the constraining power of the bispectrum monopoole, will decrease.
| r | f2acdcb058d7ac194f81545f56716709 |
Recent progress made it possible to fabricate 2D structures composed of almost any layers on top of each other {{cite:31b1fd2fde1418be688aad2c3c2bb26f2f371243}}, {{cite:4c4a808030515e3bf57371f25d78e55ad7baef83}}, {{cite:0faa4d4192072de18d7ac7d96d7fd83f5f4e445c}}. These structures display a higher variety of properties than natural materials, can be tailored to have predefined physical properties, and open a way to design high-performance devices based on unusual interface physics. Motivated by the latest advances on fabrication of four-layer vdW heterosystems (typically graphene/bilayer CrI{{formula:7527a091-a963-4c53-998c-df244f35710a}} /graphene {{cite:6a90deb1ff9db224a4a3d0f2539f409df6fa8b9b}} and WSe{{formula:f6a761cf-30c5-47f1-a70a-94cbc58d0ffc}} /bilayer graphene (BLG)/WSe{{formula:168e8bd9-7fa5-41f3-9182-95bc539a0386}} {{cite:ff03b6a89e1ed804600bd7e6051695b55eee1472}}) and a BLG-based device that allows for local bias control {{cite:47fcc7661a2d0c6bbcd5fb6b2dc2ff2a2506bacb}}, we propose an artificial four-layer LAF device [Fig. 1(a)] with two gated terminals linked by the central scattering region to realize a fully electrically controlled spin valve. Its operation mechanism is similar to that of existing giant MR devices {{cite:739297e5117a3873722271b7050ce508efffd074}}, {{cite:6ec1bbe1e22800674069f61150acf765389f17e6}}, {{cite:9cedf95922bcc2736f76e4203a98e9683eb5e13b}}, {{cite:f51c3bfe6f5e082819f3f4801c8220cf7a4f5984}}, {{cite:6b412e81198c246932fd93f82c7b37975020b72b}}, {{cite:a7aa8fbd27352f5bb10dfe82f1aecc30c6a17808}}, {{cite:6a90deb1ff9db224a4a3d0f2539f409df6fa8b9b}}, {{cite:a757a6af72e6ea32b4c74736b1a8bd243c7cc697}},
with the difference that it does not require magnetic field: By altering the mutual bias orientation in the two terminals, from
P to AP, the MR of in-plane transport varies significantly. Due to the magneto-electric coupling [Fig. 1(b)], the spin valve effect is transferred to the nonmagnetic vdW layers. We start with
a phenomenological model, and further use the density functional theory
(DFT) calculations to demonstrate the concept with a material platform{{formula:4edfa6c3-a79a-463d-94d5-b99e18309b35}} BLG encapsulated by two single layers of CGT.
| i | b38ec9ee639c1ea4c9300bea9e805da4 |
Network Architecture: The used network for pixel-wise segmentation is oriented on the multi-slice variant of the 3D U-Net architecture of {{cite:81dd0887274a498c3cded2453c8ec79fd3d59b3e}} proposed by {{cite:519494b33fc74b603b2ff708707ff79beb880230}}. The network is modified in the way that a 3D input volume with eight slices in the {{formula:a9d4e533-5c7b-4678-bc25-a0113a953a8f}} dimension is cropped out of the full volume: four slices below and above the corresponding data slice are extracted. Therefore only partially labelled volumes can be used for our training. Each 3D convolution throughout the network is followed by a Group Normalization layer with a group number of two and a leaky rectified linear unit. By processing the volume through the network the {{formula:95772c88-e30e-4982-a079-252f88bf0e19}} dimension is reduced to a size of 1, which enables a supervision with individual 2D segmentations.
| m | b41061c7ba7cc3e9f94a805fdd5e2005 |
In this paper, we formulate view synthesis as rendering a sparsely observed light field.
The 4D light field {{cite:4fce872e9ef26da434fd212af7742464466ca4b0}}, which measures the radiance along rays in empty space, is often used for view synthesis {{cite:4fce872e9ef26da434fd212af7742464466ca4b0}}, {{cite:0739f32aad302d00d6775c95d3bbb56073348898}}, {{cite:231834fe97bf474465c71859033afa81e2ec4e37}}. Rendering a novel view from a densely sampled light field can be achieved with signal processing techniques (, interpolation) and without any model of the 3D geometry, but no such straightforward method exists with sparse light fields. From sparse images, rendering often utilizes additional 3D geometric constraints, such as predicted depth maps {{cite:9bba233ac504c7c6a6a4ceee3e22bb1a01420d19}}, {{cite:231834fe97bf474465c71859033afa81e2ec4e37}}, but performance is sensitive to accurate depth estimates which are difficult to obtain for non-Lambertian surfaces.
| i | ff6b2035f9530ba0559ae3d9e969294f |
{{cite:3fea1d8448f267dc3575fc06c2635518f14e343a}} convincingly showed that SN 2010lp displayed a nebular spectrum closely resembling that of SN 1991bg about 200 days after maximum. The main difference was the absence of the narrow [Fe iii] lines. These were replaced by two narrow (full width at half maximum {{formula:57386406-8f22-4a73-83bc-183377af5bc9}} km s{{formula:95042845-1ff6-4d2d-a2a7-d78c139a8c08}} ) emissions which correspond to [O i] 6300, 6363 Å, with a blue- and a red-shifted component. {{cite:f2b20b8e45af62a0f722604c0f30cd1233fa9c79}} showed that the higher ionization [Fe iii] lines in SN 1991bg are emitted from a small volume, presumably at the centre of the ejecta, bounded by a velocity of {{formula:87359d64-fcfd-4f7c-b6d3-4c001844d122}} km s{{formula:109849e6-1e53-4172-8593-067745881fcd}} . A low central density was likely to be responsible for the high ionization.
| r | fa43a5c4d436812ee0d029c914a29df1 |
While our search only ever considers solutions that lie directly on the constraint boundary, other approaches {{cite:1adedea44089524f2f56695404f0352d6577900f}}, {{cite:835b12041a8c424de7f1093d102690d87aea2560}}, {{cite:acf857444a6f7fe815fdbadee45d4b21065f15f1}} typically have to evaluate many candidates for which the constraint is either violated or not tight (in which case the solution cannot be optimal) in order to make progress. Additionally, the DP solving always ensures proper utilization of the layer-wise speedups. For illustration, a layer with poor acceleration will not be pruned much for most sensitivity values. Similarly, due to the quadratic nature of our errors, most layers will only be pruned to very high sparsities, where speedup curves typically flatten, if this is really necessary to reach the overall speedup target. As a consequence, non-key layers will often have large ranges of acceptable sensitivity scores that all lead to a very similar sparsity choice. This makes finding a good profile in terms of {{formula:cfc2e7fa-ca36-4929-b46c-f896ab685e78}} much easier than by direct optimization in the huge space of all {{formula:5173610f-3eec-4b2c-82f2-34bee756f708}} possible layer-wise sparsity configurations.For example, the architecture search space in {{cite:1adedea44089524f2f56695404f0352d6577900f}} is {{formula:6a770be7-8c1c-4c70-bddd-83b048a32d6a}} whereas ours is {{formula:99878c12-ac89-453f-8a9d-b3c4680b1629}} for the ResNet50 model.
| d | 930dc765ad60d04c7cf6e968a137f775 |
A critical assumption of several commonly used statistical tests is that every data point within a dataset is independent. In other words, every data point would have to be equally different from every other data point for any two randomly selected data points. “Pseudoreplication” refers to studies in which individual data points may be more similar to some data points than to others within the dataset, and yet the statistical tests performed treated the data points as independent replicates {{cite:a140d841fa92be99c05c3aad800b6badc7aaa090}}. While pseudoreplication was first extensively reported on in ecological studies {{cite:a140d841fa92be99c05c3aad800b6badc7aaa090}}, {{cite:af55b7df72beb6b7fe495538d40a22b243a5de1b}}, it has since been identified as a common problem in other fields, including neuroscience {{cite:4f31ef8163c43f14f2b9fd970815ecd5ad090678}}. While resampling methods including bootstrapping were originally suggested as tools by which one could overcome the pseudoreplication problem {{cite:66b4907b928f68e1c4ba2a6c01822d8087a7edad}}, the bootstrap was argued to have a conservative bias resulting in larger error bars than necessary {{cite:5b65628428924163f2fcaa5294757f70a42211ed}}, {{cite:146f408d59bd5936ccc4ec6f4e65ddef243cddc4}}. Since then, however, several versions of the bootstrap algorithm have been developed to apply to hierarchical data and have been found to be unbiased and more robust for calculation of uncertainty in clustered data than other statistical methods {{cite:ff2f77cc95cacc3a8ae5ab91a6cab51afd3bf4f2}}, {{cite:7dcacf2e43ddc01cdf9ce63046d503cd3770bdc5}}, {{cite:9a6c1adf10bb19354b98a214ba61dc0110e2f8ef}}, {{cite:271a1a3687efb5c1558c235598a3008134565dc3}}, {{cite:95660ffabeb5a3b22fe0483e0729f884e29aea2d}}, {{cite:2b60193d26b00444d38ada34bc49e82cb5126020}}. In order to test the bootstrap for any potential bias in a typical example we might encounter in neuroscience, we performed simulations to quantify differences in mean firing rates between two groups of neurons when there was no difference between the groups. We illustrated that the bootstrap produced a false positive rate significantly below the expected 5% (Fig. REF b) and had larger error bars (Fig. REF c) than other statistical methods when the data were independent. However, when the independence between data points was abolished by introducing a hierarchical structure, the bootstrap was not statistically different from the expected 5% false positive rate (green bars in Fig. REF b) and the error bars computed were similar to those computed using summarized statistics (red and green bars in Fig. REF c) demonstrating that the hierarchical bootstrap is robust to bias for applications in neuroscience.
| d | 0c4f2ab75983e9a6598b1b30ec001f99 |
For other two-stage action localization methods (, CBR {{cite:e4a913a34f3616c302eb950d6e47fcc6ec3f9ddd}}, R-C3D {{cite:dd147844b69ea2fd48d4d2061e10f6364eb63c55}}) that do not rely on the two-stream pipeline such as SSN, we only insert one GCM into them.
Specifically, GCM takes the original proposal features {{formula:0e958eaa-8d40-4650-8f33-bc9964e10a79}} as input and outputs the relation-aware features, which are further processed by two individual FC layers for predicting the action classification and boundary regression, respectively. Formally, the action localization process can be formulated as
{{formula:b9b13906-8f4e-4b3f-b954-f076c30ea4b3}}
| m | 0d138bafc9c69f5ce9697da48af3098c |
where {{formula:f80f33a8-796e-458a-b458-0416fd9eb0b1}} is the encoder network and {{formula:268a101d-0365-466e-8c2e-6067366a5bf5}} is the projection network. Rather than directly using the output of {{formula:520dd2dd-9d66-47cb-aff6-b4f76237b19a}} for downstream tasks, we follow previous works {{cite:96e391f7fbe6f337ac638766e12ebdba54e88137}}, {{cite:f3bede54d749104acfe3f0732ff2975c1121b822}}, {{cite:3e6346fc505d4b29ebfb9f93b94d4ec2c19add33}}, {{cite:3c155f602c73a76bdd044079c7f360f49518bf98}}, {{cite:62868c450638175bf97775d222ce3878efd29de1}}, {{cite:5488136b15587b82ca0de95be2decdacc49e5f9a}}, {{cite:c87c6e2d017848de5204c4b237f141934abf1cee}} and use the output of {{formula:fb30f84f-5551-4a23-9a63-836fa7935a06}} instead.
| m | 23f24e690c1a926c1b9aa65edb64b3e7 |
Our pose estimation method is based on a previous method {{cite:4285f9155c0f3bad870bcf2ee97fac5d7280f98d}}. The method is trained on the BlenderProc {{cite:3f180a32b9925686b44f168a7f0df664bf4be326}} dataset created for the BOP challenge {{cite:c96d611ccdf1c4193a6229809399efde58e705e2}}. This method use domain randomization based on the Kinect sensor. In this paper, the domain randomization is optimized during the training without knowledge of the scene.
| m | 660b705bf26276041d9da925363d306f |
Following the idea of `entanglement wedge reconstruction' {{cite:cdd4dae09f9fe65a2bae11f0ef53a7d834a2233f}}, we can interpret the phase transition of {{formula:8aa27ed5-e34f-4c98-826c-c7376e19db70}} in terms of bulk reconstruction from the CFT. As a function of {{formula:f75f37bb-f012-4cd6-ad30-9682abbe3433}} the entanglement wedge, the domain of dependence of {{formula:f5ff9555-2536-4d84-88e4-7282f38cfe63}} , include the black-hole interior starting from {{formula:601c9229-723f-425f-befb-c2d6b407582e}} (figure REF ). We learn that it requires operators from half the boundary area to locally describe the black hole interior (to the extent such a description is possible). Note that this is possible only for a pure state and not for the thermal density matrix.
There are different ideas for explicit boundary-to-interior mappings {{cite:c709bacb3dbcdc3773cdecd04e9ffa209a740110}}, {{cite:1f4ef35adddede5d589c1221496a750f47a78392}}.
It would be interesting to see if such mappings are possible in the state ensemble context (or they are too state-dependent to survive the average).
| d | 6e93883f5807e2aa5a815cff006f8517 |
The standard approach to reinforcement learning typically assumes
that the environment is a fully-observable Markov Decision Process
(MDP) {{cite:a9e090d7c02ef549729ad715006757e5ec6e4d68}}. Many state-of-the-art applications of reinforcement
learning to large state-action spaces are achieved by parametrizing
the policy with a large neural network, either directly (e.g. with
deep deterministic policy gradients {{cite:ed51589b14bc190f52158e398659c50f36cdc8cf}}) or indirectly
(e.g. deep Q-networks {{cite:0e5560c91914f95e3020732b39e78dc8b3a03d42}}). These approaches have
yielded superhuman performance on numerous domains including most
notably the Atari 2600 video games {{cite:00bc873255855efddee9c42935cbaa9bf4fe2ebb}} and the
board game Go {{cite:4e767de99d6d989a057d7d47deabb7d65a37bc48}}. This performance is due in
large part to the scalability of deep neural networks; given sufficient
experience and number of layers, coupled with careful optimization,
a deep network can learn useful abstract features from high-dimensional
input. These algorithms are however restricted in the class of environments
that they can plausibly solve, due to the finite capacity of the network
architecture and the modelling assumptions that are typically made,
e.g. that the optimal policy can be well-approximated by a function
of a fully-observable state vector.
| i | e25fe9a15d7b3f6bf81d0d3a4073f15e |
Lyapunov function method {{cite:ad8e52d17c088f0bb21c21b0ed6ef80b56d8f63a}} is the main tool for stability analysis of nonlinear systems. The following theorem presentst a sufficient condition for hyperexponential stability.
| m | 7fa731b1b5f22ca12c8d01a0e4a97c6e |
Deep learning has achieved remarkable success in diverse applications {{cite:eb7ab32d2732fd4f9de0fb446e8dfa0134fb8ec0}}, {{cite:e45726f6d5e714f8cb4e012aaf9a7e7871aa14dd}} including computer vision and natural language processing, but its use in real-world engineering fields with small data is limited. For the TSR problem, a neural network can fit well along the loading paths but often fails to obey physical laws, due to a lack of experimental data and understanding of the inherent mechanism. To this end, we seek a physics-informed approach that enables us to encode physical laws as prior information into deep learning models, which can mitigate the issue caused by a lack of data {{cite:079cb0a58db48ee8af40c78930d0e55ed192ba8d}}, {{cite:2332396346337da7fda615656f163cd702e462e4}}. Recent advances in physics-informed neural networks (PINN) {{cite:1008b2b2836f8140f04394c6f29857bb33797392}}, {{cite:88632c2d9c11fd04b4403c0320641f87af7264cc}}, {{cite:722e63a100539cb8b2161ac0735878cabbfd736a}} that have been used in a wide range of engineering applications including fluid mechanics {{cite:94e211c0999e2db52e7f5f94ebdec58b9b06c1ac}}, {{cite:0688d7d3f6fcd75331c849ff6f620e52d0579591}}, bio-medical engineering {{cite:28f1e6df6a10b0bae207bbe39923c827b9aacd93}}, nanophotonics {{cite:f6089cba29de94cebd614a36928a32da62dc7f05}}, {{cite:837781ee1fe12efd950494e5c078b162cc06b223}} and computational materials science{{cite:16f773352bb694809f69ace6b1199130d9534d4b}}, {{cite:1c4ae3237a46970174b0df2e81a8c5e975703bc5}}, may bring an opportunity to address this challenge. PINN aims at solving supervised learning tasks while respecting any given law of physics described by general nonlinear PDE. The trained neural networks represent a class of data-efficient approximators that naturally encode underlying physical laws as prior information. This important feature of PINN enables solving inverse problems with limited data observations {{cite:ed80c13b6f2ffc0d2f45500301a94d173a69ebc9}}. However, it is a non-trivial task to simply use the PINN for the TSR problem because there are three challenges: (1) the law of physics hidden in TSR is complicated and can not be explicitly described by PDE governing equations; (2) the thermodynamic consistency in TSR is more abstract and difficult to be extracted than typical PDE-based governing equations, such as boundary conditions in PINN; and (3) the prior information from thermodynamic consistency needs to be formulated and implemented into neural network models in a rigorous and data-driven approach.
To address these challenges, our core contributions in this paper can be summarized as follows:
| i | 214d2d86d30bc5f404a2826ff3d7ad69 |
Training attempts to minimize the cost function {{formula:dbcb558f-f274-4ee8-9809-20f2f81c0772}} with respect to the network weights {{formula:51fc5911-402b-4049-be51-c42750305475}} . This form of training is often called “end-to-end”; that is, we do not train the network {{formula:35ae4137-efb4-4d2c-b574-55e63f064105}} that replaces {{formula:51110e10-d590-4807-908d-3d24809cd9a2}} in isolation, but rather on the quality of the resulting estimate {{formula:70253932-8cb0-49e4-9180-d65000e15a10}} , which depends on the forward model {{formula:85af8793-10c3-431d-b877-aeb896cb030c}} . Above we assume that all instances of {{formula:472c4de0-a86a-4c9b-a6e5-745df4b8ad8d}} have identical weights {{formula:6c68997b-7f7b-4e58-8ecd-bb166e61573e}} , although other works explore variants where the {{formula:04559cff-ab0d-44c2-a0c0-ab8ec698c195}} has iteration dependent weights {{cite:5c784e263f6187a56d6874b274c5307d50f9324f}}.
| m | 74c7adaeb95bbfcd7df346849f3e78fe |
The calculations were performed using ab initio density functional theory (DFT) in conjunction with
projector augmented wave (PAW) potentials {{cite:675aa4b866c11caf29498c7fea23999c8915e621}} and Perdew-Burke Ernzerted (PBE) generalized
gradient approximation (GGA) to the electronic exchange and correlation {{cite:8b849c561707a7cfd749f7d2bba380a53a07c63c}}, as implemented
in the Quantum Espresso (QE) package {{cite:594047962d177e33c0593c806639726edf927957}}.
The K-point mesh in the lateral directions was always 9{{formula:9d256391-10dd-4e49-ab72-efd1a5dcf2c1}} 9{{formula:17ac6123-c9ba-4196-8e96-cf467f19deef}} 1 while,
the 8{{formula:5c595f52-f300-4026-ae90-85bc7233680f}} 1{{formula:ee56027e-89ec-47a3-8f3b-71c71307135c}} 1 grid K-points was used for armchair nanoribbons
and 1{{formula:fe2a270f-77ac-4a2f-8816-b273517ede2a}} 12{{formula:989bcd52-1a94-4bcb-a996-3345039d6d43}} 1 for zigzag nanoribbons.
A kinetic energy cutoff of 500 eV was used for the plane-wave expansion.
Test calculation with the bigger number of K-points and higher cutoff gave essentially the same results. Bader analysis was used to
calculate charge transfer between Ga and S under different strain by Critic2 {{cite:cd355936dfe81539db1b162723558b3938e10fc1}}.
| m | b971e8685f06c17e88ca218752526a04 |
Graded-Feature Multilabel-Learning Network (GMNet) {{cite:b4160759a8f655002f8431c923d076e848544672}} includes two encoders for feature extraction and three grading decoding stages to restore original resolution. The proposed model employed ResNet-50 {{cite:6be384319e8f472cbac48567f158661f5df2fac7}} as the backbone of the encoders. The fully connected and average pooling layers of ResNet are removed as they may result in the loss of spatial information and details. GMNet divides multilevel features into senior, intermediate, and junior grades. The features extracted from the ResNet's last three layers, in which the visual receptive fields are enlarged, are selected as senior features. Besides, the features from the first layer, which have more detailed information, are selected as junior features. Moreover, GMNet introduces two fusion modules, the shallow feature fusion module (SFFM) and the deep feature fusion module (DFFM), to use the junior, intermediate, and senior features. SFFM fuses the features from the first two layers of the encoders, whereas DFFM accomplishes the fusion operation for the last three layers. Finally, semantic, binary, and boundary loss functions are used to find the optimum parameters of the model.
| m | 406a4a129dfde4fe8412f96616274a1e |
(Part REF )
We show that 2 implies 1.
By Lemma , {{formula:4a03926e-7efb-4b0a-922a-cb82740628bb}} uniformly on {{formula:d6a81b2a-0f59-4322-9408-fa7e32c828c9}} .
Consequently, {{formula:a5fd8f08-d309-4464-9357-5f41c04e34dc}} is continuous, as the uniform limit of continuous functions {{cite:b645ecac2ed25c4884f1bde40da78a63de0d73d2}}.
In particular, {{formula:3e9f8ad6-140e-454b-a76d-a6365b50e9bc}} is continuous at {{formula:d43cd241-fa72-440f-b7ec-b8acdb686599}} , since {{formula:39f33ca9-2ff7-4304-8bc4-bdc461a3fe01}} is an interior point of {{formula:f79427dc-fc80-4e18-a2d5-4d98081ffd18}} .
For any {{formula:e16b2f41-3315-40e3-aa8a-53ff12c31adb}} ,
{{formula:f1f0726e-dd6d-477a-b7ef-efa3fd24f282}}
| r | 55300454d061dcfd034e3735289f923b |
In this paper we pursue this line of research and study the fine-grained and parameterized complexities of {{formula:3299077c-0181-407c-ac4d-4342ea1f2922}} -COLORING problems and, more generally, of #BINARY-CSP{{formula:5f7dc757-7ac7-430f-96b4-6137a6e3e647}} with a particular focus on the parameter twin-width, recently introduced and now widely investigated {{cite:ec793b0a99193365701d3d3ae3c832239d9c5d6d}}, {{cite:fff59e2032c3de3ce32a349924ae39ae927262dc}}, {{cite:e6cda36c88309ac0f6c4cb8d00c663b3074527a7}}, {{cite:9742832785176ddc715519f43741db0645aa109c}}, {{cite:dde5a0e5b9f43cbbd50dae6a46b21fd94abb8d86}}, {{cite:7ba8aa227ccb67ee17c5bc7ef17493497d66c87b}}, {{cite:130ac9efdd71c5f899fe67721f4a16a7762bccda}}, {{cite:e951a8a1d7e9fce0aac7a70626b26092ec17323d}}, {{cite:fcc6a9520a2b53ae3d23924b209b0a1c920c2722}}, {{cite:b327365ee35a4c6e81aa87c7f3e158860e466f5d}}, {{cite:a0067e51bbd03ede09a1cd62e3cfa261f69d4bfd}}, {{cite:158b4725fdd2f752dd47d6222fc6e7c1088118be}}. This parameter, as well as the underlying contraction sequences, give information on whether two vertices are “similar” (in the sense that they have almost the same neighborhoods) in order to treat them simultaneously and thus reduce the computation time. A major achievement of twin-width is that deciding whether a graph {{formula:1d318e8c-688d-4d59-a568-af66d10c6ae1}} is a model of a closed first-order formula {{formula:178c83e3-89bc-48a4-8a9b-4b4eed6352aa}} is FPT when parameterized by the twin-width of {{formula:52c0f49e-d6e3-4351-89a5-e421298ecf70}} and the length of {{formula:d3a7414f-44ef-4e06-9062-e1906c7fe16b}} {{cite:b327365ee35a4c6e81aa87c7f3e158860e466f5d}}. Also, the {{formula:df74e078-8338-45e1-9726-d87b0724c9b8}} -INDEPENDENT SET problem is FPT when parameterized by {{formula:4f6ec57f-1d70-4f8e-9554-d070c96c223c}} and twin-width {{cite:9742832785176ddc715519f43741db0645aa109c}}. In {{cite:e951a8a1d7e9fce0aac7a70626b26092ec17323d}} Bonnet et al. also gave a proof that the {{formula:ebc5e586-edfd-40bf-b1e9-4620ce4fcaf0}} -COLORING problem was FPT when parameterized by component twin-width. Specifically, the parameter component twin-width appears to be the most relevant and natural parameter in the setting of homomorphism problems. In fact, many of the algorithms in {{cite:9742832785176ddc715519f43741db0645aa109c}} that are FPT when parameterized by twin-width, such as the one solving {{formula:56cb5851-dee3-45c5-b043-12e37e95196a}} -IND-SET, can be seen as an optimization of a more natural FPT algorithm parameterized by component twin-width.
| i | 04d4720feacca8bcf8082fccb02c4978 |
Finally, we acknowledge the promising results that recent large-scale language models have produced in terms of generalization and their (acquired) ability for few-shot learning {{cite:726ddd873de1ecb4edf4f827285b98270c68be71}}, {{cite:7f6d534f80c5e3d6858231ad549c874e66183309}}.
They are evidence for the possibility that human-level generalization may be achieved by scaling existing approaches using orders of magnitude more data and network parameters.
However, we remain pessimistic as to whether similar results can be obtained on less structured domains, such as when learning from raw perceptual data.
As we have argued throughout this work, the fundamental lack of a suitable mechanism for dynamic information binding precludes the emergence of the modular building blocks needed for acquiring a compositional understanding of the world.
| d | 9259500c963480756aacca49aa3d39e5 |
The solar atmosphere displays a wide variety of spicules with different temperatures and velocities. It has been suggested that type II spicules are a major source of coronal mass and energy {{cite:9df503edea96ab350fc3362b46f4baf45e9506b1}}, {{cite:ded10dd3b8bf7d6554809f41f4ac8660b7bc7296}}, {{cite:0a5471c73dbd20fcbf6f6b3d92fed93ce8f01784}}. In this work, we numerically investigate the role of spicules in producing observed coronal emissions. In particular, we examine whether, in the absence of any external heating, the hot tips of the spicules and the shock-heated ambient plasma can explain the observed coronal emission. For this, we inject spicules with different temperatures and velocities into a coronal loop in static equilibrium. We choose a relatively cool equilibrium so that the loop does not itself produce appreciable emission in the absence of a spicule. Each of our injected spicules consists of a hot tip followed by a cold body. We consider three different temperatures for the hot tips, viz., 2, 1 and {{formula:901744bf-a415-496b-a211-7eba3305aa85}} MK, while the cold, dense chromospheric plasma that follows the tip has a temperature of {{formula:c6db4c4f-05fb-425e-8acd-73ceffe7bced}} MK. Six different simulations are run by injecting each of these spicules with an initial velocity of either 50 km s{{formula:2ac097d1-7f27-428c-84a3-481cc344f978}} or 150 km s{{formula:07f147f7-e832-4147-8c79-5138f517fcff}} (see Table REF ). We also have constructed spectral line profiles and estimated the spicule occurrence rate required to explain the observed intensities from the quiet Sun and active regions. Our main results are summarized as follows.
| d | 6ea1b16e6bc130e3b3db249bf37a9daa |
We use the Markov chainMonte Carlo (MCMC) method to constrain the parameters in equation(REF ). EMCEE {{cite:75de1d158250743a66dab5151aa26672826135b1}} python package is used. In order to execute the MCMC process, we need to provide the prior values first. In {{cite:7a34f159ff04302407e90e79285f20b00ba2e1c5}}, the best value and the 1{{formula:53ff4d14-0c7c-4d13-90da-e56b727981ab}} confidence level of the parameters are shown. But in our work, we take an SIS model of SGLS to calibrating these parameters, which will different. So, we take the prior interval that completely covering the value range of the value from {{cite:7a34f159ff04302407e90e79285f20b00ba2e1c5}}. The prior probability for parameters {{formula:f7f0866b-a197-4f35-b615-b034fd933b6c}} is the product of prior probability of each parameter. The prior probability is assumed to be uniform distributions: {{formula:cdc212bf-a714-424a-845d-5703910db0b1}} , {{formula:04d71839-5035-40a7-9d6a-e9855006d69f}} , {{formula:86915bb1-00ba-44e3-9025-dc4326c0403e}} , {{formula:44324751-8157-4d6b-aa29-d10b0add9481}} , {{formula:e30cfa3c-54f7-427d-934a-3915f0680ef8}} , {{formula:88122543-9381-49df-a0bf-5f7ff3c8875c}} , {{formula:2c0489a7-f3a0-429e-8496-c3f9cb7cfaa4}} . In Pantheon samples, the errors include both statistic and systematic deviation. The systematic error relatives to all data point and appears as a huge covariance matrix. In this work, part of SNe Ia are selected, only the statistic error is considered.
| m | 836f5be984d9bcfe84a1b54e2cdbae77 |
We present a comparison of the results of TeliNet, VGGNet-16 and the benchmark approaches. In the training set there are 153681 covid slices and 181991 non-covid slices. The validation set contains 35016 covid slices and 40517 non-covid slices.
Since the data sets are not balanced we use an F1 `macro' score metric to measure the efficacy of our approach just like the one used in the Benchmark paper {{cite:ae009402ba0a9aeff08d7b444c511e206d8a1b31}}. These results are shown in Table 1.
{{table:97912887-fd5a-4860-a8c1-2b3cffb61df7}} | r | ae2481421c93af6cfd7457d3abaac3a4 |
In Section , we give a short introduction to the concept of {{formula:f2caed32-205d-4a49-a395-0717419b4251}} -convergence. This includes convergence results for minima and minimizers of a sequence of functionals, which are presented in {{cite:da9fd046d0497c67c07acd6320274c1000427c17}} and {{cite:6a6718c6e8ba83f78f06b01e6018f5b0d8b99e8f}}.
We also recall a few basics from the theory of Banach spaces. In Section , we define the functionals in question and propose conditions that will allow the theory of {{formula:4a5f9be3-c158-4d49-a812-1b6682cc7d07}} -convergence to be applied to those functionals. We work in a Banach space setting and consider the norm, weak, and weak{{formula:09961a05-2dbf-4340-ba1f-09287c06c06b}} topologies to highlight the differences between the respective conditions in these topologies.
In Section we present examples to demonstrate how the theoretical results from the previous section apply to typical inverse problems settings like integral equations of first kind and a parameter identification problem for an elliptic boundary value problem. The main achievements of this article are the presentation of criteria for a family of Tikhonov functionals in a fairly general setting to satisfy {{formula:e8a8374a-2196-4bbe-ac6e-e86d6d0141ab}} -convergence to some limit functional as well as an infimal property as (REF ).
| r | 219d174ae8d1a2c68479c431c27cbb63 |
Furthermore, other properties of the materials can be calculated by the dielectric function, including the optical conductivity, absorption coefficient, energy-loss spectrum and optical reflectivity. They are given as follows {{cite:a470574cfcefddca91bdd7507a3269c139e86cc4}}, {{cite:0b4745f6207204d9beceadbe478b74cefb02cf22}}, {{cite:8d2cb52adcd593f8e68ca2c11ad6d2ebbe1fe013}}
{{formula:a4d6e089-ff63-4009-b301-ebd5298e2188}}
{{formula:7bc41333-8f59-4e5c-9684-ee1c726c623e}}
{{formula:bc063925-3689-4f83-8622-e5cfc639bfea}}
{{formula:6922b856-fc0f-47a1-aaeb-32fc119ff527}}
| m | dd548de8ae48790927c89981fea85f5c |
Quarkonium production in hadronic collisions is an ideal tool to access the gluon content of the nucleon {{cite:c11e080b258cd9321ddee0ab3faa7a1e79af1c44}}.
Collinear gluon parton distributions (PDFs), depending only on the longitudinal fraction of the nucleon momentum carried by the gluon, have already been extensively studied.
Recently, quarkonium production in {{formula:efc558bf-9c52-479c-b5d3-02a9b03404a9}} and {{formula:947f7f47-acd5-484c-b904-3b61166ed765}} collisions has been considered in the context of the transverse momentum dependent (TMD) gluon distributions {{cite:836e5ebe63d076f818269c2b6b211df702e66691}}, particularly for the unpolarised and linearly polarised TMD PDFs and for the gluon Sivers function (GSF). In this contribution we consider {{formula:ea88eb30-dec8-4df8-a665-c143819ddcc8}} leptoproduction at small-intermediate transverse momentum as an important tool for extracting complementary information on the GSF. To this end, we adopt a TMD phenomenological approach, the generalised parton model (GPM) {{cite:901d90037646d83f2304f6bf33d9f3544633c410}}, and the nonrelativistic QCD (NRQCD) effective theory for the quarkonium formation mechanism {{cite:707cfc2d1239ecf465a6ff7c0a6fb09caaadb9d3}}. The combined use of these two approaches is challenging, but it can be very helpful in better understanding both the complex three-dimensional structure of hadrons and the details of quarkonium formation.
We will evaluate in this framework both the unpolarised cross section and the Sivers single spin asymmetry for the process {{formula:86e90d2b-d03f-4ffe-bb6e-5a9c7f776f49}} , comparing our results with data from the H1 Collaboration (for the cross section) and with a single data point available from the COMPASS Collaboration for the Sivers asymmetry.
We will then give some estimates for the Sivers asymmetry in kinematical configurations suitable for the Electron Ion Collider (EIC).
| i | 15495d288a1e3a8e59fa8bb6a9c7b5c8 |
For more challenging cross-view cases,
{{formula:0b31c8a1-359c-4c22-90e2-33ddfc871a6f}} covers more dramatic view-changing factors than that within {{formula:3320cea4-8628-4238-9f60-6b09c4e75a1f}} ,
i.e., {{formula:f2c999a6-2b39-42fe-b67c-4cd75da00b1b}} ,
intuitively meaning that it is hard for GaitSSB to maintain its pre-training performance on this occasion.
But encouragingly,
Fig. REF shows that GaitSSB can perform still promising performance in the most challenging cross-view case,
such as {{formula:86e65b54-8752-4b41-9c8e-263b3f758587}} rank-1 ({{formula:4f657e89-fc17-453d-a99c-8a8f37ab3d0c}} rank-5) average accuracy for the most challenging {{formula:92f8b527-4177-4cba-bc1f-6046a6d948ea}} -{{formula:33c1a732-f4e6-4eef-8cb3-c7fecd2ab769}} (probe-gallery) view pair on the widely-used CASIA-B{{cite:c77efbde88e271cb526e0460254337077bf7d4e0}}.
This observation encourages us to explore the further understanding of the view-invariant feature learning in our GaitSSB.
Next, we start by analyzing the inter/intra-class distance.
| d | b3789a591e1f0dacb4667f36bfe6ff06 |
In this section, we perform extensive experiments with ASWL on VGG-16 {{cite:a96a34cdfb9aa13311ae807bce25187e68b3bc88}}, ResNet50 {{cite:6554ddf31a0f32528d4361ade999b2bf501661ec}}, and MobileNetV2 {{cite:3c312780b980839691e8e9488698d07681c31434}}. For VGG-16 and ResNet50, we simply replace the traditional convolutional layer and dense layer with our attention-based convolutional layer and dense layer. For MobileNetV2, we replace the {{formula:24ac1c7e-e02f-4475-9735-759b24c8ecb3}} point-wise convolutional layer with an attention-based convolutional layer but left the depth-wise layer uncompressed since 99% of the parameters and calculations are contained in point-wise convolutional layers. There are totally three hyperparameters in ASWL: the coefficients of sparsity and L2 regularizer ({{formula:f1a278c9-8149-409e-8c4e-dccb71a54de7}} and {{formula:2e5bec20-68af-4926-b7a2-8b32c8bbe509}} ), and the pruning factor {{formula:5a6932a9-fe4a-4512-ab1c-ad02b2babcc1}} . They are specified later in different experiments.
| r | 5b53826055ad0629fd831f0478f70584 |
In this paper, we investigate the Casimir-Polder interaction between an atom in an excited state and a Chern insulator {{cite:7872c036aec9042df11ed2110681e4be4c0cd5ad}}, {{cite:0a55f7a4c50fd759092bc6ec81f864a34fe20939}}, {{cite:80a296ea04aeb41bfc8002733655043b1281cf6d}}, accounting for the full frequency dispersion of the conductivity response of the Chern insulator.
Such an investigation is motivated by the existence of three broad types of dispersive behavior that the conductivity can exhibit depending on the frequency regime of the photon, which leads to qualitative differences in the spontaneous emission behavior of a neighboring excited atom. We study the resonant Casimir-Polder interaction corresponding to each of the frequency regimes, and find that the resonant Casimir-Polder force can be greatly enhanced if the excited atom is resonant at frequencies associated with a van Hove singularity of the Chern insulator. We consider the Chern insulator because it is one of the simpler examples of a nonreciprocal Hall surface, which enables one to investigate how the presence of the nonreciprocity makes possible the emergence of an effectively monotonically decaying and hence longer-ranged repulsive resonant Casimir-Polder force. Such a phenomenon is predicted for certain sign combinations of the Chern number and the angular momentum change associated with the atomic dipole transition, and does not occur if the surface is made of a reciprocal material.
| i | e6fa74fbdeedc7e98aa985ea44d59447 |
We solve numerically Eqs. (REF ) by split-operator technique {{cite:a51a20bfd848235837c65a68295c393221b6802c}} for trajectories of an atomic white dwarf corresponding to closed eccentric orbits. We consider the Bose-Fermi droplet consisted of {{formula:020aeb01-d294-42b5-ac9a-31d1db44b515}} Cs bosonic and {{formula:6c4fd42c-5063-4ac8-9408-c01f44abedca}} Li fermionic atoms. Such mixtures are studied experimentally nowadays {{cite:9a6a7635065a751df3219f9827481871b9e572ca}}, {{cite:acac6dce508db65049124e0b5995717bc06e581f}}, {{cite:fe8e133b3e91df4d3fffac1a47a5b174edd7fb61}}. The droplet, consisted of 1460 bosonic and 100 fermionic atoms, is initially located at apostron at some distance (about {{formula:ab37f9a6-3c86-431b-9be4-e1b1a66df96b}} ) from the artificial black hole, far away from the horizon. The initial numerical parameters are chosen in such a way that we can follow several revolutions of a white dwarf in an eccentric orbit (with periastron at about {{formula:c5a22dbb-e666-43d2-937b-abac0f1ecc9a}} ). Such a task is numerically manageable (i.e., is not too intensive with respect to the computational time) provided the white dwarf is not located extremely far from the black hole. This, on the other hand, means that the period of revolution of a white dwarf in our case remains short, in fact it is at most of the order of ten seconds. At each periastron passage a white dwarf is stripped off an approximately (at least within initial few passes) the same mass, which is about {{formula:ab8e1475-11d3-449f-b859-e5c27b5b0477}} of the white dwarf mass for a particular orbit discussed below.
| r | cec39b1c6819b255d56f598aa8687f07 |
We find the factorization (REF ) appealing as it shows that solution of the IFPP for arbitrary distributions, and in multi dimensions, is reduced to two standard and well-studied problems, i.e., constructing the Rosenblatt transformation (or CDF in one dimension) and designing a uniform map. It is therefore surprising, to us, that the factorization (REF ), and more generally the Rosenblatt transformation, appears to not be widely used in the study of chaotic iterated functions and the IFPP. Grossmann and Thomae {{cite:45e3aa33e15ffc5a2e62547d3bc84f860d4c1866}}, in one of the earliest studies of the IFPP, essentially derived the factorization (REF ) by introducing conjugate maps and establishing the relation (in their notation) that {{formula:eafb13d2-c653-4aa5-81f3-742b2b5daa5a}} where {{formula:1c46af83-4022-4425-8196-510c5a59e1b5}} is the invariant distribution and {{formula:9d35ca9c-9afa-4740-979c-d1b63c07c39d}} is the conjugating function; see {{cite:45e3aa33e15ffc5a2e62547d3bc84f860d4c1866}}. It is a small step to identify that {{formula:2ef89ef0-cc91-4ef5-a5f0-4a84770617d9}} is the IDF, generalized in multi-dimensions by the inverse Rosenblatt transformation. However, the connection was not made in {{cite:45e3aa33e15ffc5a2e62547d3bc84f860d4c1866}}, despite the Rosenblatt transformation having been already known, in statistics, for some decades {{cite:3c8684cf4ee1390a83a5d0b63fe3fe466ee22bf9}}.
| d | ea2e18e3daea37b1686a54ae5c0afe0e |
The recipe to restore the gauge invariance is given by the theory of polarization in crystalline solids {{cite:a379cd39ae902520e386db34cb68d3deb8ab4806}}, {{cite:2dc84a53a1ab8b081b0ef9443554f79cd3f4ecc4}}.
We first take the exponent from Eq. (REF ),
and then use {{formula:f5055835-b41e-451f-94b7-aa58c4221675}} ,
{{formula:4d0719b1-9169-4b9b-b10b-9706e891307a}}
| d | 18b2c957d1a726c764dbe4653db2ac35 |
Along with the experimental progress on manipulating the coupling among multipartite quantum systems, the tunable anisotropic triangular structure can be realized in various physical platforms, such as trapped ions {{cite:105feeada23ffc3013855e8f832de8d7fad32f38}}, cold atoms {{cite:d09010e877ab455006830875007bd8d28f7adb11}}, artificial-spin-ice heterosystem {{cite:f83099f59ca2c8cf96dff7a61479519966b83f87}}, superconducting systems {{cite:6937f550ce30d54d1e38e5d61f634f76ba689727}}, and so on. Since the MQC is an important resource for quantum computation and quantum simulation, it is desirable to realize the correlation modulation in a practical system.
| d | 6010493dedac5d40c355c6570d7b7786 |
Strassen's Theorem {{cite:ad045c665ba5ac5103a3e0bca074ab82e80320de}} characterising the existence of a probability measure on a product measurable space,
having fixed marginals and prescribed support, has enjoyed an illustrious history, both leading to new fruitful
research directions and
having significant applications.
Such joint probability measures, known as couplings of the pair of original measures,
are the starting point of the theory of optimal transport and
appear as a fundamental concept in the celebrated Monge-Kantorovich duality {{cite:97d788e24082399d72cd26244bae7f57ee673358}}, {{cite:48e4f58a8f5042d280d121d9224e9b802cf239a4}}.
They are also at the heart of Arveson's Null Set Theorem {{cite:ee2f3c21ff6327fadcc775f1526ac96e18360235}}, which formed the base of vast parts of
non-selfadjoint operator algebra theory and had a lasting impact on the study of invariant spaces
for collections of Hilbert space operators (see {{cite:4866cbeecbcc48d84ca78ad293fde1fbea079663}}).
Arveson's Null Set Theorem was given a quantitive formulation by Haydon and Shulman {{cite:8a8baa0fe4d6cd21675ab854ef709b6688cc251b}};
the quantifying parameters defined therein
were shown in {{cite:8a8baa0fe4d6cd21675ab854ef709b6688cc251b}} to be capacities in the sense of Choquet's capacitability theory
{{cite:7a5a3e4b1c60925b527b41c8a19e62696cbc5d94}}.
| i | 23d3e7a5670408fee0569030325c92e2 |
Recent developments in deep learning can be applied to modeling of image-to-image translation problems. Generative Adversarial Network (GAN) is a class of deep learning algorithms that is used for generative modeling {{cite:b9a42102d14e57fc6e5f691a58571c50438e0028}}. GANs formulate the generative modeling problem as zero-sum game between two networks. A GAN consists of a generator network that produces samples from a given input sample or noise and a discriminator network that tries to distinguish if the generated sample is from the real or fake data distribution. Although Convolutional Neural Networks (CNNs) may be used to perform image-to-image translations, many of them apply a stochastic approximation to minimize an objective function and require paired data {{cite:e74420648a20ac2bdd708b8c00f4eed38ab19237}}{{cite:05b4e9a6989b26e0e62ff70e98967708280e50ad}}. Alternatively, GANs try to achieve Nash equilibrium by generating a distribution that is close to the empirical one.
| i | da7eb5e0d6d9f3fba79a8f1a730b858d |
The spatial search algorithm by continuous-time quantum walks was introduced by Childs and Goldstone {{cite:1e0e3401a28f9c6df4792a36ba93fd3317e29f30}} after the successful definition of discrete-time spatial search on hypercubes {{cite:e67428443b353c7befaaf9ae3b1e6e326ef132a9}}. In the continuous-time case, the time evolution is driven by a Hamiltonian that is obtained from the adjacency matrix modified by a term that depends on the location of the marked vertex.
Experimental implementations of search algorithms by continuous-time quantum walk are described in {{cite:c7716a91517a25f4460a2b92d6d774e1ea027d1d}}, {{cite:d9668d99c485630262e0b3f932a516e241329c86}}, {{cite:7214a844ffc99ce78c7178f8d92cec2f5bded4c8}}, {{cite:32588d5a4cd7c39ff5020b0b87dbcf50733d5906}}.
| i | baa91641ff1bf12db9e8c1d26ac5f3ed |
Returning to the motivations for the TCC discussed iin Section II, we see that they are based on considering the evolution of fluctuations at the level of an effective field theory. Thus, a point of view that one could take is that the TCC does not necessarily imply that inflationary models with an energy scale which exceeds the bounds discussed in Section III are ruled out, but simply that they cannot be described at the level of an effective field theory. This is the view presented in {{cite:5e7c00e81a2775677c0ed8eceaf44b6d54aac7fe}}. In this aspect there is a connection with the question of whether de Sitter solutions can be consistent with string theory. Based on the swampland criteria and on the TCC, the answer is “no”. In fact, there are no-go theorems for the existence of de Sitter solutions at the level of an effective field theory (assuming that the internal manifold is time-independent) {{cite:1198f4efacd8e83becdced1f906abdf7de29faf8}} (There is, however, a lot of controversy on this issue. For an opposing point of view see e.g. {{cite:b9a87af3a651849d81f19c89624e080661ed0f1b}}.). However, going beyond an effective field theory approach it is possible to obtain de Sitter solutions, although they are unstable {{cite:480abe55462600789dce5cbdb68ce1f0778d427e}}, {{cite:6d618f6966b4ae0af383bb6d73631329bcd04857}}.
| d | 7d7687a05a51ab98b3712bd36f1a792e |
There is some correspondence of the dynamics of the model discussed in this paper and the quenched Kauffman NK model {{cite:94d9dc187cd72334bc6970ac05b3e6eaf96c0d60}}, {{cite:5b4dc9f59e47de0e45fa7790fd2f5719a39558bf}} of time evolution of networks. As we argued in {{cite:c921cf2d3500bf0658bd2f3c21e0846306346ff5}}, here the number K of incoming links which determine the current state of a node (here: of a link) evolves with the number of degrees of freedom (here: {{formula:36398836-2a90-40f7-8597-8ebc8c1805d3}} ) as a square root of this number (here {{formula:ba717dfc-6626-4708-b61d-a522f0958f72}} ). An important difference is that in our case, there is only one function (given by Eq. (1)) which determines the state of each link in a subsequent time, and not a random (fixed in the quenched model) set of these functions, different for each node. What is similar is the large number of steady states with minimal energy, which in our case is just the number of balanced states, varying with {{formula:dbcfc03b-2ebe-4c2e-9149-5e04cb3f51f0}} as {{formula:c94ae875-b979-40d5-a259-a03b3666eb5a}} . We add that the process of reaching the Heider balance, modeled by Eq. (1), has been termed as 'social mitosis' {{cite:3d43f945e0b31c7d34a1df2056b7f2dc9a3056e8}}.
Limit cycles in the Kauffman model {{cite:047014c0a548742508e8dead2c3e1d5d82baabab}}, {{cite:f924c6adcbe4630ab104827f8862e95c19b8048c}} are no less
important than fixed points and have
biological interpretation. Our results indicate that limit cycles can also occur when evolution is deterministic
and identical for all components of the system.
| d | 1e34f6f9a06f6a09522f99ad60674df9 |
Given the lack of a significant population of low luminosity star-forming dwarf galaxies in the M101 Group, we can place rough limits on the slope of the faint end of the star-forming luminosity function of galaxies in the group. A steep faint end slope would predict a high dwarf-to-giant ratio; the lack of these dwarfs in the M101 Group argues instead for a relatively shallow slope. By adopting a {{cite:1f90853d3d9a23074205287cd4952320c070bd62}} function with {{formula:0864115b-9cf7-43f8-ae09-8d15e86454f8}} {{cite:3d2d7d8f40cb239d5d262bf07f921abec2bb175a}} and varying the faint end slope, we can calculate how many faint star-forming objects we should detect within the group.
| d | abd0c2cbc201808614fb79225db14c9f |
Several noisy-reverberant conditions are used to evaluate the proposed techniques in terms of speech intelligibility and quality.
In order to compose the noisy-reverberant environment, two real reverberation rooms (LASP2 and Stairway) and two non-stationary acoustic noises (Babble and Cafeteria) are considered in the experiments.
The LASP2 room is part of the LASP_RIRAvailable at lasp.ime.eb.br. database with a reverberation time of {{formula:d479d5bc-1025-4b2c-9e93-bb9d5dd31477}} s.
Room Stairway is selected from the AIR database {{cite:6fbc99d5c9ea06495b52fc8ca9c6d629c346fa89}} and presents reverberation times of {{formula:b7f153d4-c983-4539-8644-4e4143cb233c}} s.
The Babble and Cafeteria additive background noises are selected, respectively, from the RSG-10 {{cite:ea5e9370a769f4a0f2edfc6d10bc2ed1e5d11700}} and DEMAND {{cite:b7843d0c9fd14ca3405537cd5ab7731c9d9f9235}} databases.
These noises are classified as non-stationary, with maximum INS of 40 and 15.
{{table:2a8f39e6-74ca-453e-a8a0-2f49e603e108}}{{table:09f2c43b-fc07-4f3b-8ad4-4e1b04c1b2c8}}{{figure:f58399f8-1764-4bab-81f3-4c098e8a812e}} | d | ce40a0068902f7803c931b4838f8fae7 |
is nonempty and bounded, where {{formula:bee2b481-2a0c-4b5e-b3f0-5999c54c6c97}} denotes the set of differentiable points of {{formula:224ec208-c3a7-4dca-98c0-cef9855477b9}} .
The set {{formula:a5b210af-4334-42df-9000-eb5cbb0e409b}} is called the B-subdifferential of {{formula:6a7e3b93-3e06-4947-bd9b-fb3b8ef3ccab}} in {{formula:166aabfa-3479-4a7c-bcd6-4adf4d4404df}} , its
convex hull gives the generalized Jacobian {{formula:987ed8b0-e177-4be9-ae3b-64faedd79ecd}} by Clarke {{cite:baf19ebe8a724ed341ea3236747230d66d7c89a1}}.
A point {{formula:5d6eba1b-e433-4b78-96ee-1530feffd47a}} is called BD-regular, if all elements in {{formula:db73af91-9a84-43d7-bdbf-99dea0a7b18a}} are
nonsingular. The nonsmooth Newton method
{{formula:b1782fae-23b1-4762-9ca6-513f6cf2abe5}}
| m | 7aa7fbf42a98e884b7742bcf86ce595b |
The equality {{formula:7f0d2a10-5f9e-4852-a473-2bbc2a033a4f}} was first proved
in {{cite:6e5b2d8f88e32777e5e89b27db7da4c18497beee}}. The equality {{formula:5b07baba-e033-4f98-b7f4-56573622440b}} is
{{cite:73d098bc58df8f85ba0a56189fb7e7af839bb5db}}.
| r | ca4d12f9f7c0b86bef3d5d96fea34159 |
Since the state controllabilty, a concept decribing whethere the state variables and space of the dynamical systems are controlled or not by the input variables, is put forward by R. Kalman, et al, in 1960's {{cite:9f48c5d378b2632c53f404c88c5de6784de679aa}}, the further studies on describing quantitatively the control capability of the input variables to the state space attract the attentions of many researchers in control theory field. Based on the analysis of the controllability Gramian matrix and the mobile analysis of the system eigenvalues, some pioneering works on quantizing control capability (defined for describing the control ability and efficiency in paper {{cite:aa826f94f284c09e58db3879a6a7bc7948070151}})are made {{cite:39a131fd5354909456c2c45ec30aa6deaa537ad8}}, {{cite:3552afbe9f4c4ab07856b5273d78146620190e7e}}, {{cite:63c95be52750dde777065cf91d3cd92c99478011}}, and {{cite:17e6e5324f09f448e6c1eec2e58dc332c06faf01}}. In recent years, more systematicness and deep-going works about that are got {{cite:222ce5da788cd9293c5b062c571206922f7be27d}}, {{cite:aa826f94f284c09e58db3879a6a7bc7948070151}}, {{cite:4a14808875e87057c6ff0bac24de9ff01359425b}}, and {{cite:0927f10d1d8113d4bd73f7e43f94fd0409ad420a}}, such that,
| i | c45bba9488b49a7c1045dba8b9e74ba0 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.