text stringlengths 54 548k | label stringclasses 4
values | id_ stringlengths 32 32 |
|---|---|---|
Lore et al. {{cite:85f90c175c691ba37f61c0515b4933a5fd05f868}} proposed an auto-encoder architecture for denoising and brightening the low-light images.
Many Generative Adversarial Network (GAN) based networks were also developed for image enhancement. Chen et al. {{cite:dd69144c42557cbf936bc5a226fa90835583b12f}} proposed a method that uses a two-way GAN architecture. The network transforms the input image to an enhanced image with characteristics of a reference image. Ignatov et al. {{cite:207290c95643ba58eb610d5835eabdf212332691}}, proposed weakly supervised (no exact image pair) GAN based network that enhances images that are taken using mobile phones to DSLR quality images. The network used DPED dataset {{cite:0260032fe426e78e3cd143a7bdb0cdaaf9346f73}} along with many other unpaired HD images.
| m | f78ef463e73da2833925c1a53cccd63c |
Furthermore, our algorithm for the exactly sparse case (depicted as
Algorithm on page 5) is quite simple and has low big-Oh
constants. In particular, our preliminary implementation of a
variant of this algorithm is faster than FFTW, a highly efficient
implementation of the FFT, for {{formula:c51fccae-f8bc-4c6a-b17a-60ebb17a92c3}} and {{formula:c21a1b25-1602-4e5d-b6dc-c60a193264b8}} {{cite:48e0995af4674ef14bd906118faf16623062018f}}. In
contrast, for the same signal size, prior algorithms were faster than
FFTW only for {{formula:b0df3822-4abf-4594-ac2e-810c46e07472}} {{cite:a46529e50cc6cfd67e34378afda81b3bc83aded9}}.Note that both numbers
({{formula:e698c8e0-cd66-4668-ad65-8f20093bb48a}} and {{formula:43c96c8b-9182-4aae-aabf-0959b70224e2}} ) are for the exactly k-sparse case. The algorithm
in {{cite:a46529e50cc6cfd67e34378afda81b3bc83aded9}} can deal with the general case, but the
empirical runtimes are higher.
| r | e2bdf0a8caf6d3db442195141fbb084b |
Cooperation is much more widespread in nature than the Darwinian premise of `only the fittest survive' might suggest. Why and under which conditions cooperation thrives is, therefore, an evergreen subject across the social and natural sciences {{cite:edbfeb195aa0aa93f2c59fd5ba63c36add4e4c6e}}, {{cite:b7bb0e226d0a33510b94adf8f4432c5ce4b1515a}}, {{cite:d06e0ea8178e74a214644dbdc30e78e754b263cd}}, {{cite:c14f6717c5ee10efc58444d864e1487962d612b5}}, {{cite:9bba58291526d4197d34fbd517e4cac4c4fbf81b}}, {{cite:b17315f615ffcc00ece4fb5be69f818d0c69219e}}. Evolutionary game theory is traditionally used to formalise the problem mathematically with social dilemmas {{cite:9f087d248d865efb7ac7078c30fa5d37ed8361bd}}, {{cite:89387fc5d273c089f5f46af6b45576ed96990c76}}, {{cite:f8a7a15ca1f3bc2cc4e975e30124a582f0e04f66}}, and networks are commonly used as the backbone for the simulation of these games. We have here shown that the structure of the former in terms of the unique sequence of degrees predict the outcome in the latter, particularly in terms of the game parameters at which significant shifts in the level of cooperation can be expected. In particular, we have proposed a conjecture for cooperation transitions in arbitrary networks, including phase transitions from absorbing to mixed strategy phases.
| d | 96cfeea4d4d81b4fa314bfe62927a1f6 |
Comparison with semi-supervised approaches.
Here we compare DETReg to previous works that employ semi-supervised learning of both labeled and unlabeled data. Semi-supervised approaches like {{cite:1483799e25c339609e429450813f1b5e2b76260c}}, utilize {{formula:ffff307f-e959-45dc-95b6-1ad3c511601f}} of labeled data but also use the rest of the data without labels. We pretrain DETReg on the entire coco train2017 data without using labels, and finetune on random {{formula:cd87b928-926c-462a-b851-a4bf3d681174}} of train2017 data for {{formula:51302182-8933-4b7d-afa1-fe583acea4bb}} . After the pretraining stage, we finetune with a learning rate of {{formula:dcce45ac-cc8f-447b-8e24-d6ff7d450ee5}} . For {{formula:c85a5d49-b80a-4e7e-afbe-9fe894849ddb}} we finetune for 100 epochs and decay the learning rate in a factor of 10 after 40 epochs. For {{formula:fa30f2a8-0b3e-4908-a583-b154189420e5}} we finetune for 200 epochs and drop learning rate after 150 epochs. In each of the settings, we train 4 different models each one with a different random seed, and report the standard deviation of the result. The results in Table REF confirm the advantage of the DETReg pretraining stage, which results in improved performance for all settings.
{{table:8a339876-253a-4c1a-ba7d-60d793d1fea8}} | r | d9a91a4e59640b88103fa959296a4d08 |
At the start, we want to clarify a common misunderstanding about the role of adversarial pretraining. The goal of an adversarial approach is not limited to learning robust representation against potential attacks. Instead, it is also employed to improve the generalization accuracy in downstream tasks, especially when the adversarial model attacks on feature levels {{cite:69fb3d0bf34eeb4182ad55110a79c33461eac5f8}}, {{cite:cb8d4da40b1b236622bcab578a09ce90ab3426b8}}, {{cite:893e8449d55f579580a076c8fe17c5d992cd684a}}, {{cite:e8b875120f8cc72bb0732b6905e5c5a8ee7f5cd9}} rather than on raw inputs (e.g., image pixels) of individual instances (aka instance-wise attacks {{cite:7a02bff0fe736dc1ee649bc8e5ee8c88af68e854}}, {{cite:bf7af7f05cd0e3d2767d28d49ca8217dea503ef2}}, {{cite:e536fc1b9bb076efd228060d457a18ee5d4f32cf}}). When not attacking on the raw inputs, the adversarial pretraining often cares about whether the learned features are generalizable to future problems, aiming to avoid learning trivial solutions that merely use low-level features to bypass a pretext task. For example, easy negatives in contrastive learning could result in less discriminative features to distinguish between positive and negative samples for a query {{cite:e8b875120f8cc72bb0732b6905e5c5a8ee7f5cd9}}, {{cite:69fb3d0bf34eeb4182ad55110a79c33461eac5f8}}; in masked imaging modeling (MIM), the network may learn low-level features to reconstruct missing patches by simply using the similarity between locally correlated patches {{cite:1f497b5fa3181f1a2420cd49f48fb76843b0e5ce}}, {{cite:31322826dc8e25db6fdca7dc1c7843c3c6f0f6f3}} if the MIM objective is not sufficiently hard. In these cases, it is beneficial to explore adversarial approaches to improve the generalizability of learned representations. In other words, learning more generalizable representations through adversarial pretraining is an equally important goal in literature as learning robust representations against some presumptive attacks.
| i | 54d9fcaaf23c7820c9bc43394fd84a3e |
It is generally accepted that coexisting firing patterns induced by noise or chaos can arise in networks purely connected by electrical or chemical synapses together with synaptic weights and seems to depend on other factors such as time delay and networks topology{{cite:4b9ecf2db21db76c330b577a73390c3fe1d7f403}}, {{cite:4ff91aa257e9d54dbd3ce3c92d6a002b8e3560e1}}. Computational models show promise to identify some of underlying mechanisms to reproduce coexisting firing patterns. Early studies have reported that the inclusion of noise was regarded as an ad hoc mechanism required to produce interesting dynamics{{cite:b626629cc3c5ec22e48b0dd36e8d8139e0854cf9}}, {{cite:1dcaefeb026405c232ab84e1149bc81ed5b1a561}}, {{cite:4618e4c5eef68d3cafd26e5207063218d0352076}}, {{cite:a90aebfe33aed4cd36bf83344d6f807b724bd6a5}}. A common feature of the noise-driven multistable models is the existence of a subcritical Hopf bifurcation to generate multiple attractors. Then, noise produces the switching between them. In other words, for deterministic models attracted to stable or periodic solutions, noise is fundamental to avoid that dynamics become stuck in a state of equilibrium. Therefore, the ad hoc introduction of noise in a dynamical system near a bifurcation ensures stochastic switching between different attractors, endowing the simulation with the kind of various firing patterns seen in the empirical data.
| d | 6d1d95bdfa399beedebe3db55a8d9bbd |
One method uses computer graphics to augment camera images, superimposing virtual objects of the target category onto the image {{cite:8947262f4c011715c935215a02be3b60a16a0074}}, {{cite:e9829b11f04f6b3a8662fad2b3e2fd1b86c9176d}}. This approach is simpler than synthesizing entire scenes and improves generalization. A second method is to build scenes and objects using game engine technology {{cite:bcff2bfc272dc2e09a0847e576d55b4c72db9166}}, {{cite:3d5e9e7d4879482723c1afb7097eb314ac445d16}}. This approach is helpful for building scenes; its limitation concerns the image quality of rendering. Third, some authors use ray-tracing to produce high quality images {{cite:d5bd700199513eadf041bfa22238fa808989966b}}, {{cite:faf52faa497e368b496ce35b568ecbdbcd70a0c5}}. Ray-tracing has been combined with driving simulators to automate the creation of large numbers of realistic scenes {{cite:169f798256d40d0953e6dc346f0b946fa2acbb18}}. A limitation of ray-tracing is that the rendering time is long compared to game engines, real-time ray-tracing is a current undergoing research topic {{cite:a0b1ef986420b4032d5c9269ae9b1c7178856689}}, which can potentially be helpful.
| m | 0d71ec96f8a4ae51aebf0e45bb220d09 |
The mean filed Gross-Pitaevskii (GP) equation is not sufficient for the droplet; we need to consider the higher-order correction, popularly known as LHY {{cite:5204136065650ead1e193eaa7a6eb73471cd953d}} correction term.
The well established coupled GP equations for the 2D droplets is given by {{cite:87168e7b8bbaae7ca2a58c6bd234e4790253431d}}
{{formula:6869a79c-4b3d-44d9-add9-73c3c2130b9d}}
| m | 1e44cf2316231204d026005407c19c8d |
As asymptotic gauge transformations act on the two dimensional {{formula:44dda03e-350b-44b5-a498-48ec8d8fc462}} , it is natural to expect that the asymptotic symmetry {{formula:f8b296cf-1d35-41d6-87be-75142cb74fd6}} gets enhanced to a Kac-Moody symmetry as long as singular transformations are allowed, which was actually shown in {{cite:5ab07f426bda98d26d1aa022e38ebd249e36513a}} for Yang-Mills. However, it was found that the usual holomorphic and anti-holomorphic currents do not decouple, and hence is only possible to define a single holomorphic (or anti-holomorphic) current algebra.
| i | 2e1ef09d4460545e213a76fbc4950f27 |
While there exists work surveying ML-augmented CO methods {{cite:8f7deeac39e0f0416b6158d36615187967bbf263}}, {{cite:b0bd746794f7f8373cae4b817cec103579d173e8}}, the more modern end-to-end CO learning methods lack of a cohesive and critical analysis. The goal of this survey is to address this gap and provide a focused overview on the work to-date on end-to-end CO learning, provide a critical analysis, and pose a set of open questions and directions.
| i | a9e03b00a1dc037d1f4c96f26a1d59c5 |
Our approach to develop this controller is to first identify that the kinematic structure of any modular design can be represented as a graph, where modules are nodes and electromechanical connections between them are edges.
Fig. REF depicts the relationship between modular designs and graphical structures which we adopt in this work.
We introduce a modular control policy, represented by a graph linked to the physical kinematic graph, where a module is both hardware and a component in the policy controlling it.
Even though the designs all differ, they are made from the same set of components repeated within and between designs.
This observation presents an opportunity for dynamics and control information learned from one design to transfer to another.
Our learning architecture enables knowledge about the dynamics and control to be shared among the modules, providing better efficiency in learning than conventional neural networks applied to individual designs.
In this architecture, any assembly of modules has a robot-level policy consisting of the union of modular policy components.
The policy learns to automatically alter its behavior depending on the design by learning an inter-module communication protocol computed over the graph edges {{cite:17f9943ba47040985499e3192823c8ce25cfed94}}, {{cite:2dbdf8476e4f6c34aa34a21f18361a6438ca061d}}, {{cite:a61db8d615e525301e388d403c01fd7a1b214050}}.
| i | 87cb1eab444876100487ae60ee17c27e |
Over the past decades, there have been considerable developments in soft robots that attempt to bridge the gap between conventional machines with high performance but rigid components and biological organisms with remarkable versatility and adaptability {{cite:5abc719ead97ab1d272d8122b81c72cdca9fcbc2}}, {{cite:fab37c2db7f0cc304c4dbc6ae3ceebe82cacc6fe}}, {{cite:5045e2cfc45a27bf4d1f0612ac60b47890fa4a0d}}, {{cite:ad4e5ac2e0c91d7d3ccbda5f65aaa373f8e54dc1}}, {{cite:9cf6815dae95a823a61c7b0e9f312f7c773e9013}}, {{cite:a7d012b1455125143b21ff91e04049d943125eb4}}, {{cite:ed3e693810538fc9d74b940665cf31b4ce647256}}, {{cite:c94adf6cf107974ce510f57b6681aac60738a338}}.
The merits of soft robots are generally accomplished by deforming partial or all of the compliant robotic bodies via approaches such as pneumatic and hydraulic actuation {{cite:086fb34621eb9a3b0c19c3e0d839dfec16d79354}}, {{cite:fc9ae8803b56dec3999a10e824e06f1cffac20ba}}, {{cite:4f75a6cbb1f23fc594cd9a2793821956cac228e6}}, thermal stimulation {{cite:81e6b75a2aa40f182695e3086021517e4564912b}}, {{cite:1e2986027a19170f16ce4b42ec1309c57d5fce7e}}, solvent swelling {{cite:a11dc245115bbfefba113d7134d13792baf44a57}}, {{cite:2a733a7c5e5eabed37c81b90150f21c7ec093076}}, and application of magnetic and electric fields {{cite:7f3823453cf6e599c05fb721c0dc25a302ec849a}}, {{cite:8020bb922b3403e2d1f5f70f917d9960c2e93170}}, {{cite:26f760d67a5946c029d3eedf834dc69dd98ce1e0}}. The deformation mechanisms can be roughly classified into two types: one deforms with a rate monotonically related to the input energy, but the performance is limited by the power and capacity of the input; the other exploits elastic structural instabilities to uncouple the input and output by gradually storing elastic energy before releasing it suddenly {{cite:83c6d1b68bbcebabd1030f93c8832cd1dffb44d1}}, {{cite:e108a8516b49c6e654cfff716f70eaf5104b21ee}}, {{cite:cce4f5abb2166cf4a288fca52b45a5710eca99d7}}, {{cite:e9841d1ee74da74b2556c2d5c3855f705237045a}}, {{cite:4e35675f935249a1f7f54903064bb014fb727d7c}}, {{cite:0a4b1c9b9a2b7d13fb89dcbfa1bfd1c4d30984cd}}, {{cite:4180e0bfdf5db1cb92bad555d0157d4f544b73f0}}. The latter is ideal for applications that require high-rate motion and fast energy release.
| i | 8f550db81f7ed63c5ac4c968cef58d74 |
Furthermore, there are examples where all the models may all fail to produce some geometric structures at object-level, as illustrated in fig:comp-ps-vgg (sixth row).
In this situation, we still need to find more informative statistics.
One may for example consider to incorporate a second layer of wavelet transform as in the wavelet scattering transform {{cite:a853a000e2bc6b1a0a65bfa1110b06d8f36c4167}}, {{cite:704a08320d8d8311f39a427ab7ea4c9650c3c476}}.
Another line of research is to introduce other kinds of losses
(such as to encourage image smoothness) in order to improve the VGG model
{{cite:3692d1c627e84f576fea2dd1463003958380b2bf}}, {{cite:9f0e126bf2be27044487f469e8a8cb1fdc91f534}}.
These losses are complementary, and could thus also be added to our models.
Integrating these models with learning-based approaches is another promising direction
{{cite:5457fed1b4c52004897e8ecbee702019c4d15e2a}}, {{cite:e5a5bfd27f7f9754350195731931d0e99c63c06f}}, {{cite:a88bfe8c84d1ea8305068eee058df1682da31038}}, {{cite:24e889a4f42b2980ae3f7a1d859eb870df8c646c}}.
| d | 219c793c51543faa65193dd08fc10174 |
Our task is to find all central extensions of an algebra {{formula:39232fae-278f-4cc3-afc2-a7caaab467f3}} by
a space {{formula:c937a044-79fa-4c38-8f67-8e1882ae8f8d}} . In order to solve the isomorphism problem we need to study the
action of {{formula:9f6c2bd9-6b8d-465a-a82a-5f9b7f2a8c1d}} on {{formula:0804b421-1c99-49c7-8bf8-17d33297def5}} . To do that, let us fix a basis {{formula:f9568b47-51d7-4200-8d38-aacc6fa394b0}} of {{formula:b37c9c50-7f90-417c-b8b8-241e2988718a}} , and {{formula:b2807cae-da65-481c-a042-e1957c21267d}} . Then {{formula:78438d18-de93-4fd8-9d5b-28105fc2796f}} can be uniquely
written as {{formula:a18ab5b4-063d-468a-b302-3a11ecd11fd6}} , where {{formula:07a20052-4ee7-42ae-ba66-22d1f4eeca5f}} . Moreover, {{formula:9827add2-7e28-430d-91a3-996b83b63951}} . Furthermore, {{formula:64dd8d11-6913-4d5e-95a6-0a88f7934d9f}} if and only if all {{formula:297eab90-9015-4833-a5ce-cc56739523fc}} .
It is not difficult to prove (see {{cite:fd581a912dadd84e34bffd6b7d18d56c7ec90235}}) that given a weakly associative algebra {{formula:04000ed4-2deb-4e88-b949-d6548f41f401}} , if we write as
above {{formula:9cbcd619-bc68-4f29-a359-88b381c9a688}} and
{{formula:a561a342-1420-4811-925e-c61d14585383}} , then {{formula:cb8fd671-34ba-4b83-bb91-80acb4b05ca0}} has an
annihilator component if and only if {{formula:ba263fbf-cbdb-47fc-b8de-05cd94157481}} are linearly
dependent in {{formula:f087943e-9d80-4728-b75d-194fefb4f456}} .
| m | 3893d1c57d8c58820ac68558bc5402ce |
Finally, we estimate the bitrate of our model. Instead of encoding the image pixels directly, we can operate on their latent representation. Compression, however, is a source of error since discarded data cannot be easily recovered. If the latent space encoding becomes part of the training process, we can learn what information to discard and what to preserve. Traditionally, the latent space is quantized and then entropy coded. Yet, such an approach does not allow for easy gradient flow. Therefore, we follow the approach of {{cite:61a83ce08b38bd0563846cf423368ede890d2d95}} and simulate the quantization by adding uniform noise {{formula:9b4b7f1c-2f94-49cb-b612-25f27a6ff09c}} to the latent representation {{formula:437ca02e-5220-4779-8461-67c06cb39a66}} during training:
| m | 7eb3dbde2909880bbd6a520369e39fb3 |
Brain dynamics is often inherently variable and unstable, consisting of sequences of transient spatiotemporal patterns{{cite:fceb56f53b5f9bc8c67630cf82912bc03a6dd7ed}}, {{cite:2ff76f7803e0da1cc0125b8c86841f79ac890508}}. These squences of transients are a hallmark of metastable dynamics that are neither entrely stable nor completely unstable. Our results pay a possible way to uncover the underlying mechanisms of generating metastable dynamics, such as chimera-likeness, that may mediate perception and cognition.
| d | 4065496aaff8af2fa0b27bbe12f898bf |
Uemura {{formula:f646b1d1-4ec0-469b-825c-3905cbde1843}} {{formula:4638f174-d1bd-4112-8b1e-fbae4fab1959}} . {{cite:2ae7f46b6a4a3f5ab0964ff36e938237c71d4184}}, {{cite:e66ddf03e9d787d465951af830534a3708b57246}}, {{cite:baa4e063b993ab23d57dac5dc402993889ea1cae}} classified the superconductor in conventional and unconventional on behalf of the ratio of superconducting temperature to the Fermi temperature. If this ratio value falls between 0.01 {{formula:bf135e73-b730-401d-a2a3-5004c4eb1e88}} {{formula:5669720c-ee89-4af3-98af-11dd779ecbcf}} /{{formula:8eb7800e-3865-405c-be56-ceb662977e62}} {{formula:6766acab-c6ea-43e8-bec3-5a95177f6d47}} 0.1, then a superconductor is considered as an unconventional one, and heavy fermion superconductor, high {{formula:b76bc22f-76f4-47a5-8999-bbc1327681f0}} superconductor, organic superconductor, Fe-based superconductor lie inside this band. To calculating the {{formula:0891fd0b-275a-46fa-bddb-67d38e10d539}} value for both HEA samples, the expression used is as follows:
{{formula:2a062689-5f52-4a04-b414-37463b47a57a}} = {{formula:7cf00e21-e8d1-4fa3-9fcb-f703654918f0}}
where {{formula:c3098195-78bc-46ef-b046-663a66054dd1}} , {{formula:a4458565-c936-44c0-9965-ae7686c13416}} , and {{formula:d0f979b7-d5ff-476f-8c58-dcaf9d0012a2}} are the effective mass, Boltzmann constant, and carrier density, respectively. The estimated value of {{formula:77029c5b-51a8-4db7-a278-3a7a11597abc}} for (HfNb){{formula:4336d951-0e3d-4790-bae4-9bd3d97d373b}} (MoReRu){{formula:35c6b6e9-44e4-47f6-b144-6c660d4b96cf}} and (ZrNb){{formula:8ef02daa-1375-48d8-b605-c7874e36942e}} (MoReRu){{formula:8afe07d2-57f2-42a6-ac5a-7e1a4b89a24b}} are 21520(1480) {{formula:ee24e0ac-7196-4d69-a486-bf4a5c301e3b}} and 18589(1445) {{formula:f36bfe00-810b-4e8c-ac59-e087a3bb3b13}} . The {{formula:6ea42e0e-4f16-4275-890e-658ffb3c6248}} which are far from the boundary of the unconventional superconductor like the other noncentrosymmetric and unconventional superconductors.
{{table:3cf7e7a7-89a8-4f40-9066-4c5921e89320}} | d | 62f2af0e674588962075c75a53c412d4 |
One line of work that aims at learning sequences of skills
that are compatible is
Hierarchical Reinforcement Learning (HRL) {{cite:eb4a91d1cad22d328bd2e3677569364f5e300c72}}, {{cite:3cfa998804561d8246a3f408b06164bca2aefbba}}, {{cite:c6674881689b3951391e0adfd55a954e8132ed74}}.
In principle, hierarchical agents should be able to transform a task into a sequence of subtasks that they solve sequentially.
However, to date, existing hierarchical agents have mostly been applied to learn navigation or reaching tasks where learned skills do not interact with each other.
It is unclear how sensitive hierarchical agents are to possible interactions between learned skills.
In this paper, we investigate another approach by reformulating the agent's subtasks and the corresponding reward signals. Similar to {{cite:9607576689eb30d2a00f6fa4756ca900de8c62b1}}, we train an agent such that it is motivated to control a particular component of the environment state representation while minimally affecting other components. Such an agent can learn to control components independently from other components, thus making the learned skills compatible with each other.
| i | c85f89d40a4e28d109233eba9750373c |
If a critical configuration is reachable,
i.e. if the above algorithm accepts, then {{formula:448d35a4-5134-4669-a2c1-89799dabae8b}} does not satisfy the second sub-problem, otherwise it does satisfy it. Therefore, deciding the second sub-problem is in co-NP.
Thus, the {{formula:a8c11d24-a8a8-404f-b820-ed5a92cb8902}} -time survivability problem is in a class containing both NP and co-NP, e.g., {{formula:3291de85-9127-4ad2-b919-d369e12a6294}} of the polynomial hierarchy ({{formula:4f71bf3c-1d10-40cf-8327-894823144bdb}} ) {{cite:1166cafe96431d69e34c73febfc479767d91cb9b}}.
{{formula:ae0b79ad-09b3-47a9-a0a4-8093f368b219}}
| r | ff8f0f26ef3defea5bf552600a9216d1 |
LV-ViT. We use the LV-ViT-Small {{cite:a30a728acfaea675c5d2188dffcc04006248e86b}} as the main architecture in experiments.
According to the design of LV-ViT {{cite:a30a728acfaea675c5d2188dffcc04006248e86b}}, the token labeling technique is introduced to improve the model performance of the model. This method takes advantage of all tokens to compute the training loss in a dense manner instead of computing loss on the additional learnable class token {{cite:a30a728acfaea675c5d2188dffcc04006248e86b}}. To improve the recognition on the token level, each token in the LV-ViT is assigned with an additional label as location-specific supervision. Therefore, as Figure REF (a) shows, we apply the same sparse mask on those labels as it on image patches to correct the computation of training loss, because the number of tokens is reduced when input image patches are applied with the sparse mask.
| m | 68359e8e779972c0a3bf90071a21f186 |
Several previous works have shown the advantage of using the graph structure in recommender systems, but they all have restrictions in different aspects. According to {{cite:36bf54e91184fb52075be1ca5eace9b8359de5b3}}, {{cite:666e46d922f44427c7abd47a80b8190d0b2f2488}}, the graph structure is capable of incorporating collaborative information explicitly. By taking user-item interactions as a bipartite graph, these models exploit the high-order connectivity of users and items and encode collaborative information into the graph structure. However, all these models are only suitable for static scenarios. The advantages of sequential dependency and time information are wasted in them. Moreover, SR-GNN {{cite:7220df35e44813e34ee2ee596f9e7d6549444d54}} proves the superiority of graph structures over sequences in the dynamic recommendation, but it fails to incorporate the evolving of items. To deal with these problems, we leverage dynamic graphs to model the evolutionary process of dynamic recommender systems.
| i | b796294af5f6ba294b21b020154031e1 |
Humans have always been the benchmark for the performance of most machine learning tasks. As deep neural networks have become more successful with several NLP and Computer Vision tasks, they get closer to beating this benchmark. However, despite their success, neural networks have several limitations in terms of distribution shifts{{cite:e0061061ded920703c6e999c64edd578d2af2495}}, adversarial perturbations {{cite:fc496be186481fab63c73fc7f2ef03dcd0bc9994}}, crowding {{cite:50a2413131697f2228216f8720df669434f5b6ad}} etc.
In particular, neural networks are extremely vulnerable to adversarial perturbations, which are small, carefully calculated perturbations that lead to gross misclassification. Since these perturbations are usually imperceptible to humans, the failure of neural networks in this area is an imminent issue. The most popular way to beat adversarial attacks is to train the network with these perturbations so that the network can learn better and avoid misclassification. This kind of data augmentation is called adversarial training and is computationally extremely expensive{{cite:52aa2133f501a74cac317d08e3be11be7d69ba05}}. Furthermore, all adversarial defenses, including adversarial training, suffer from an inherent trade-off between robustness and simple classification tasks{{cite:52aa2133f501a74cac317d08e3be11be7d69ba05}}, {{cite:49bc8b0845242036958d2740435b66d429391610}}, {{cite:8bfe09f985d25720247258ff76c1b1d94b23aed1}}.
However, humans do not have any such trade-off and excel in most computer vision tasks. Additionally, humans do not require any additional data to be robust to perturbations. Elsayed et al.(2018) {{cite:422c281ab4ada259537c8e5315753b7e80d48137}} have shown that some components of the human vision could be responsible for the inherent robustness to these adversarial perturbations. Through our work, we investigate such bio-inspired models, which incorporate different components from human vision in DCNNs. We benchmark the robustness of bio-inspired models against adversarial attacks and more real-world common corruptions. Here, we find that bio-inspired models without any data augmentation tend to be surprisingly robust to small amounts of noise and have the best performance on common corruptions{{cite:e0061061ded920703c6e999c64edd578d2af2495}}. We also find that bio-inspired models tend to use a mix of low and mid-frequency band information, making it possible to classify more complex corruptions and be adversarially robust. Furthermore, of the the bio-inspired models analyzed, we find feedback is essential in ensuring good performance on common corruptions, showing future scope to be added to larger models.
Code: https://github.com/code-Assasin/BIIR
| i | b2d3f665d5cbd32dcb952595a284809d |
Regardless of whether trapping occurs, comets seem a promising mechanism for delivering dust to the hot-emission region near stars. Sun-grazing comets exist in the Solar System, and there is evidence for extrasolar equivalents (`Falling Evaporating Bodies' or FEBs; {{cite:891337e57f801d6e2d47920498a6b895e60be581}}, {{cite:a501bdf03dd59a00161bc4262e42cb802e844908}}). A tentative relation between hot-dust detections and circumstellar gas indicative of cometary activity has also been suggested {{cite:200bcac1c82310e637cea15b5d69eb4c89fe025a}}, and stochastic comet infall rates would naturally explain the NIR variability seen in at least one system {{cite:6326d92b9773cfde47f012789be8b92dff89e3e3}}, {{cite:52fe60c7e0081f4a12a1af508a4a5f164da0ffee}}. There are several known mechanisms capable of producing star-grazing comets including direct injection of Oort-cloud-like comets {{cite:6d753c17f84a846873d33e0c3995fdb74a7215ee}}, inward scattering of material by chains of planets {{cite:5199af75087f17be5e2814fd08a7573f9cf06c12}}, resonant driving of debris eccentricities by moderately eccentric planets {{cite:5579707659d544dbdb319f5ca362cf377f98e0d4}}, and secular driving of debris eccentricities by highly eccentric perturbers {{cite:baa82417f1eaaf93f0929dd47ba38d5d3f3169d7}}. Comets also appear to be the dominant source of warm zodiacal dust in the Solar System ({{cite:87c9f6c2729bcf8465085bb04b317ddcf4cb5c1e}} and refs. therein).
| i | ab39b8af90c91fc03fcd9b6df61e9889 |
Many of the coefficients presented here have been obtained with different methods.
The expectation values of the stress-energy tensor and the currents for free massless
Dirac field evaluated here are in agreement with the exact results in the case of pure
rotation and vanishing chemical potentials of {{cite:b907ff9c20d9bbb0bb9a9aa7b90a1eda83d0dea5}}.
The CVE and AVE have been discussed at length over the past decade
(see {{cite:a67e59dbbd9e8697c8445186ca2ad6ae2920d4a8}}, {{cite:8e018d22fa159facccf991666244b167b22bd4e6}}, {{cite:64225d4272f2e63b8a57f8bf4df71910ed687834}}), and several calculations of
{{formula:68302135-1906-45df-a965-3210f98f6605}} (REF ) and {{formula:e35c32fb-103e-43f3-960f-154ed255b3b7}} (REF ) were presented in literature.
| d | a41431658228b5ff9749f7a21d457877 |
Semantic Segmentation.
We evaluate semantic segmentation performance on the ADE20K {{cite:9aa5789d4af08d8dc89683d90db8700020f5a323}} dataset. Following prior works {{cite:e523cd417d2c9bacc571bf1cbeb65ffebd5d9906}}, {{cite:a404fb87ab37136bfd3b144715ffe5eaefb9da21}}, we initialize the UperNet framework {{cite:72f90e70172f0a272b68bb9c6b5affbb323b9914}} using our pretrained model and finetune the segmentation model end-to-end.
The model is optimized using AdamW for 80k iterations with an initial learning rate of 1e-4 and a batch size of 16.
We set the weight decay to 0.05 and layer-wise learning rate decay to 0.85.
The results are shown in Table REF . Our method is able to outperform competitive representation learning baselines such as DINO, MoCo-v3 and BEiT.
Our model is even comparable to MAE, which is trained with a much heavier schedule (300 epochs vs. 1600 epochs).
By scaling the training data to the larger ImageNet-22K dataset while keeping the total number of iterations unchanged, our model performance improves by 0.5 mIoU, surpassing all prior arts by a significant margin. This indicates that our model scales well with data.
{{table:35e02290-3f1d-4ba3-be5a-bd454932a634}} | r | fbf72ceccda1705d2470d7e4b15bbcba |
MLP-Mixer {{cite:a9b825dcd1ef1cce25216abfb1c0acc372ccbe14}} is a simple yet powerful model initially designed to capture spatial location features for image data, and also found useful when processing temporal signal data.
ConvLSTM {{cite:3513e8071fc10fd89c3f0e9799d68989546330f9}} reports an excellent and consistent performance in capturing spatio-temporal correlations. It replaces the fully-connected layers commonly used in the traditional LSTM with convolutinal layers.
Bi-LSTM {{cite:0d81125cdaa4030847224d16891bbf1c707b7951}} has a structure that contains two stacked unidirectional LSTMs for both forward and backward pass. It is heavily used in processing spatial-temporal features.
Transformer {{cite:ecb2d18c53a74309d92b93e16a9eeefb2bc86ecb}} is one of the most prevalent models in recent years. Apart from its great success in the natural language processing (NLP) field, it also reports a promising results in uncovering highly nonlinear and dynamical spatial-temporal dependencies.
| m | fee09d2219d997e5738e3a6d2f5e0781 |
Our method is inspired by the F-Principle. In the research of deep learning, the F-Principle expresses an opinion that the low-frequency elements are related to the generalization of natural training while high-frequency elements are related to the robustness under the perturbations. We utilize the Fourier heat map {{cite:1ca1dff8284970f63726e18303fd1c18e6fe80f8}} to validate our robustness in the frequency domain. By biasing the high frequency towards the low frequency, robustness can be guaranteed with a low error rate. In the sub-experiment of different aggregation strategies of wavelets, we find that discarding the high-frequency elements absolutely leads to the degradation of robustness compared to Wavelet Average Pooling. In summary, our adversarial wavelet training method can be regarded as a positive example of the F-Principle.
| d | df773da8ae70b3252dd01e5244c65253 |
Inspired from equality of opportunity proposed by Hardt et al. {{cite:887c1bec75d542fa471612bd49d039f377da07c3}},
we define the fairness of {{formula:0e0ca156-3578-4829-a0ef-e25105602dbe}} to be the maximum difference of the availability, i.e., the fairness of {{formula:ddbd37d7-4d47-4c95-aed8-58ee0d51fd77}} is
{{formula:dbb7a8ae-7596-4007-8a91-ac07b897ef39}}
| r | bbba438df5c01779ceb2760dbac08a0f |
where {{formula:4bdbbb52-3dc7-4b0e-9c9c-aa177d759740}} and {{formula:94c47416-0ae2-4605-aad4-bb999c53eff1}} denotes the maximum and minimum entries of {{formula:9a0a9e63-ee53-4fc0-83ab-492651a4ec44}} , respectively.
The structural similarity (SSIM) index {{cite:c9ae7c35754c969a20e3455b03832323488df075}} is used to measure the quality of the recovered and true images
for image quality assessment:
{{formula:9fd01b9e-e8a3-4339-8338-34de5c0fb699}}
| r | a2a807ead3a99a98558b96fe832d1f3f |
Very recently, it has become possible to dissect further the contributions of different neurons to processing, using pid {{cite:60c1b6211f804839b6b6167ed8851758511c6337}}.
pid enables us to disentangle for a target neuron {{formula:2fee1c1d-57d3-4ca6-b99f-d99190eb5520}} , how much unique information it obtains from its own past activity {{formula:6e344aad-7f04-4e4a-aa94-71991f060f55}} , or the past activity of a second neuron {{formula:ceb6f063-34be-4640-a69e-21206b2eb764}} ; and how much information is redundant or even synergistic from the two (fig:pidschematic). Synergistic information is that part of information that can only be computed if both input variables are known, whereas redundant information can be obtained from one or the other.
| r | 2cd1e5c66a70074592f004dc7393d504 |
The method proposed in {{cite:9a611d5de06c9874b30e88fbaaaf98487bf268e5}} overcomes this problem by applying a preallocation of channels.
Preallocation means, that before the combinatorial auction, a non-exclusive assignment of channels to tenants is performed (i.e. a channel is potentially assigned to multiple tenants as well), and each tenant considers only the subsets of channels preallocated to it in the bidding process. In other words, the search space of the CA optimization problem is constrained. In this setup, the first critical task of the preallocation process is to limit the computational demands of the consecutive CA, by limiting the number of bids submitted. The second task is naturally to ensure an appropriate preallocation in the sense, that optimum achieved by the CA later on the limited bid set has to be efficient. The method proposed in {{cite:9a611d5de06c9874b30e88fbaaaf98487bf268e5}} uses the many-to-many version of the famous Gale-Shapley (GS) algorithm {{cite:30e359a23b318d0437a774b744df8aeef92e2f27}}. The paper {{cite:f2858bb85b04f148f3171abd370ac2fe80b20896}} argues, that using the many-to-may version of the GS for (M2MGS) preallocation provides slightly better results compared to the more straightforward simple approaches of distance-based preallocation. However, the paper {{cite:f2858bb85b04f148f3171abd370ac2fe80b20896}} considers only the total resulting capacity of the system as a performance measure for comparison. Furthermore, it does not assume the potential blockade of channels in the simulation setup, which may easily arise, due to e.g. obstacles in the operational area. In addition, this very initial comparison of these two preallocation methods does not analyze, how the proposed algorithms scale up for problems of increased size.
| i | bddba0437441a39aaf0031c1208c70bc |
The overall architecture of the proposed model is shown in Fig. REF . Fastspeech2 {{cite:1b0e7bf07f738789df6c87e7557d5e3605628b49}} is used to generate a mel-spectrogram from the phoneme sequence. We propose a speech mixer {{formula:f86dd304-035f-4ce2-8c3d-be67c6b60b12}} to generate pseudo-labels {{formula:518737f5-1f00-4424-8b54-c073660a64ca}} reflecting intermediate emotion intensities in a variance adapter. The speech mixer {{formula:dcbfe9b1-852f-452a-8374-7e09149fc686}} generates an intermediate low-level elements like pitch {{formula:72ed6308-d48e-4a2f-acbf-764217e16ad4}} , duration {{formula:6071d5ec-e977-4f4d-9d0b-3fe859ea66e8}} , and energy {{formula:f8a34c5d-99b9-4052-8819-ccc1f0a39eba}} . Also, discriminators {{formula:f2b5e9d7-c747-4351-b6f2-87ac512bbab8}} is applied to the predicted elements for improving naturalness.
| m | 2921836674d17c976004460648673082 |
Recently, gravitational waves from binary black holes have been reported by LIGO and Virgo Collaborations {{cite:691d6a364841294bde103e202fabe31a3417825d}}
and the shadow of a supermassive black hole candidate at center of a giant elliptical galaxy M87 has been reported by
Event Horizon Telescope Collaboration {{cite:a3648f1bf540e3841cca355c8b03df7f9fe99c2c}}.
Phenomena in strong gravitational fields will be more important in general relativity and astrophysics than before.
Black holes and the other compact objects such as wormhole have unstable (stable) circular light orbits called photon sphere
(antiphoton sphere) {{cite:d7b76b69219ae7f0751ce0884fce31f96117e643}}, {{cite:0b945bb45a78485b8c36ee8f4feba17922f727cd}}, {{cite:2a4b1930d8c8d7521abe86c919ef4a02413eb444}} because of their strong gravitational fields.
There are many researches on both of the theoretical and observational aspects of (anti)photon spheres {{cite:728b7e9f46278b346c00bfc33df2b7a035602368}}, {{cite:4232b65ece68f0fad5cbfd422a2c4a44b7d2ff20}}, {{cite:64fe2f1eddb0d4fbd07423b2d6d25b7726a36ae0}}, {{cite:c7fe65c2288b0d01883e41beac6547df1ddc411c}}, {{cite:e3aeca6e5c074efa83008517c19fe2e5ee6af2e8}}, {{cite:eb995c1e6f87cb17fc828c9945db2a5bf6d184c9}}, {{cite:6c73c1870209b347846705a7f890da4afa6ca338}}
and their generalized surfaces {{cite:0b945bb45a78485b8c36ee8f4feba17922f727cd}}, {{cite:8def04ca169c8852c1f4483505e5f70a157fbb71}}.
Stable circular light orbits may lead to instability of ultracompact objects because of the slow decay of linear waves {{cite:8dcc92291d80715a444dc15519648167cc84ce18}}, {{cite:f470bc6fbbd472297eb14d189fad4f9e658b923a}}, {{cite:434e687f32a420a8679e6418ff39ea01299779ae}}.
| i | dc3d136c7e228498f78c1bf7b971b447 |
As shown in {{cite:77c8c4393f12d927247849fdabd126a377066d4f}}, to power an X-ray plateau, the dissipation of the spin-down wind must track the spin-down rate, i.e.,
the magnetic energy of the spin-down wind should be almost released in X-ray emission totally, so we can approximately equal the left and right sides of equations (REF ) and (REF ), respectively.
Therefore, the uncertain parameters on the right side of equation (REF ) can be connected to observable quantities {{cite:fe532c801cd2cf2b4bbfe5943ed2a5f004452b98}},
i.e.,
{{formula:81e10d20-0d7b-4604-aac0-fab2a8720de9}}
| m | ebc5efe61f91653ea34f5d37a762e153 |
The cloud envelopes in this system are ionised by a superposition of many physical processes, including photoionisation from young stars and other mechanisms which give rise to LINER-like emission, the best candidates being photoionisation by cosmic rays, conduction from the hot ICM, X-ray photoionisation, turbulent mixing layers or collisional heating. AGN emission dominates only the BCG core, see Figs. REF and REF . This is in accordance with the findings of {{cite:69c6f11246358b820232ff2c852263b0638d316f}}, whose analysis of the CO spectral line energy distribution in the M1931 BCG also reveals evidence for multiple gas excitation mechanisms including SF, interaction between the molecular gas and the ICM and AGN emission. According to {{cite:ea58faf9985f2329986d9e9ec96ce7acdc0cf8f8}}, such multiphase-star forming BCGs also tend to have unusually luminous vibrational and rotational {{formula:ac2faec9-2f31-4f88-bd86-709a58b9a877}} lines, which are by far more luminous than in normal, star formation objects. The emission in such galaxies is not fit by {{formula:e1ea4135-f046-4cf0-a0c3-47cfaef358db}} gas at a single excitation temperature, hinting to the fact that there must be a mix of excitation mechanisms at play, in accordance with our findings. The spectral decomposition method of {{cite:fe86ae355d534f6c3212fd36a9a85491d52fda1e}} allows us to infer the fractional contribution of each mechanism to the total luminosity of the emission lines. Ionisation from star formation has a fractional contribution to the luminosity of the emission lines of about 50-60%, whereas AGN emission accounts only for {{formula:a16d59a2-49fc-45c4-87e9-a54cea8821ed}} 10%. LINER-like emission accounts for 30-40% of the energetics needed to ionise the gas. The occurrence of such "composite" emission is typical for cool-core BCG. For example, {{cite:9d6e9193fdb1e3d09a67477b3505b8cefd5a671d}}, {{cite:1a4a32bdd996afc4964eb1355940f03c0b3d7fab}}, {{cite:16b03e383d30e07bc0789870edf2b41d9cefbd5c}} also studied (cool-core) BCGs based on IFU data and infer composite emission for their systems in the BPT diagram and concluded that the systems are ionised by a superposition of many physical processes.
| d | 5ba9f5d7bdcbfb8ad4e8e4f580c4344b |
Experiment Setup & Implementation Details.
We used two-stream I3D {{cite:d6473854efc3e16e56764ce8bb451c981b99653f}} pretrained on Kinetics to extract the video features on THUMOS14, following {{cite:7df0e2a22d4067ac20971365ab16f2bc4e6783e0}}, {{cite:39d82c6c293988382ebc62921cff3c30f639f030}}. We fed 16 consecutive frames as the input to I3D, used a sliding window with stride 4 and extracted 1024-D features before the last fully connected layer. The two-stream features were further concatenated (2048-D) as the input to our model. mAP@{{formula:669f5455-ec40-451f-b9ff-86885c83088a}} :{{formula:e2b52c3f-a80e-4c07-9d37-59943bd0dcf6}} :{{formula:12697a19-e2c8-4e0f-aac1-c36e9412e460}} was used to evaluate our model. Our model was trained for 50 epochs with a linear warmup of 5 epochs. The initial learning rate was 1e-4 and a cosine learning rate decay is used. The mini-batch size was 2, and a weight decay of 1e-4 was used. A window size of 19 was chosen for local self-attention based on our ablation. We also combined external classification scores from UntrimmedNet {{cite:1061a45ff053ce97f86f2e2635f10e6126dea2ad}} following previous works {{cite:39d82c6c293988382ebc62921cff3c30f639f030}}, {{cite:28406126acbb1722ec9e5e06437981efd424f1ac}}, {{cite:200fdd98e629ca7ed3c8a91424249a162fc64ad3}}, {{cite:8374fcbd990c0aff3d29c8a0446fb9685c3f7c4b}}. We refer the readers to {{cite:98949f04ec927b5de4a3a178403b8bb186f04e3a}} (Appendix E) for a detailed description of the score fusion strategy.
| r | 426a487fb8fc7e42047826bdd7a6cb35 |
Comparing to the existing work, the proposed method has the following characteristics to tackle the challenges mentioned in Section :
(1) Many existing unsupervised models are only trained with normal samples {{cite:4ff62cbada3074658e1d1f753b1aaf521076bcac}}, {{cite:48713872217b6f84678b9f18a7cd385100fe3a3a}}, {{cite:804412580762ab840e7f625212b1f9da4552a569}}, {{cite:623975011b94bf78dd199a13abafc08fbafc79e3}}, {{cite:7889fb6408055594e27513b0bd745f7798c0c24d}}. Although some methods can deal with the contaminated data {{cite:64ee1efae192c37c84d85ac659ebed525671ad88}}, {{cite:e7b10220a1b472098ead4a4a6ee96a2c0cae0ca6}}, {{cite:35e987815cb443ceea4d328d5dad447fd3ee3a90}}, they make efforts to retain the obvious-normal data and filter out the suspected abnormal data. Our method leverages both normal and abnormal samples in the datasets and attempts to enhance the detection capabilities in the transition field. Specifically, our score-guided regularization utilizes the obvious data to enlarge the score disparity between normal and abnormal data, and guide the training of the representation learner and the scoring network. By doing so, the proposed scoring network can improve the ability of the representation learner and the robustness of the entire detection method.
(2) The existing unsupervised models define anomaly scores with selected appropriate metrics, such as Euclidean distance for distance-based methods {{cite:48713872217b6f84678b9f18a7cd385100fe3a3a}}, {{cite:cc92f2f4651042289376f0dfa35814cfb3189ba6}}, {{cite:623975011b94bf78dd199a13abafc08fbafc79e3}} or cosine similarity for reconstruction-based methods {{cite:64ee1efae192c37c84d85ac659ebed525671ad88}}, {{cite:bb7478d119ef75520a9b9d24a0d6dec1bbed85ee}}, {{cite:35e987815cb443ceea4d328d5dad447fd3ee3a90}}. The anomaly scores are not directly optimized in an end-to-end fashion. Devnet {{cite:daf5078e6f6d080dc1a610c8c78d16ad832333d7}} is the first weakly supervised method to achieve end-to-end anomaly score learning. However, Devnet guides model training with labels that are missing in the unsupervised settings and treats the unknown anomalies as normal data. In contrast, our scoring network can be incorporated into existing unsupervised methods without additional assumptions, expanding their ability to handle contaminated data sets and directly optimize anomaly scores.
| m | 71273a14736c622f3cad40413495c933 |
To demonstrate the effectiveness of HL-Net in exploring heterophilic information under occlusion scenarios, we propose to calculate two different R@{{formula:5e2d34d9-6014-4990-a737-9d2aef98aa1a}} metrics for the SGCLS task. More specifically, we decompose the SGCLS task into two subtasks, namely C-SGCLS and S-SGCLS. The former determines the SGCLS performance on triplets where at least one object is heavily occluded by others, , IoU{{formula:2c35cf46-03e1-4f35-942c-03c706704b9f}} 0.5. Otherwise, we term the subtask as S-SGCLS. As shown in Table REF , HL-Net outperforms all state-of-the-art methods on both tasks. In particular, when compared with MOTIFS {{cite:b28049014de964e058da9de8c9b97c4f3a4a8fb3}} using the same ResNeXt-101-FPN backbone, HL-Net achieves an advantage of 5.7{{formula:fec333b7-703a-4083-8fde-792c7c6428c0}} and 1.8{{formula:05d178fb-b5a3-419f-8f22-7088dc552dba}} at R@100 for C-SGCLS and S-SGCLS, respectively.
{{table:667b7b79-acd4-42ac-99d3-c0409b12b96d}} | m | f9c2ac8215a191609f2591ce6d8a9752 |
Fermilab has recently released a measurement of the muon anomalous magnetic moment {{cite:3b8055e3d800f442b91a2d324357b7e4b5420116}}, confirming the previous BNL E871 result {{cite:a9dcc862540138beb624fa1a9b748cf76c746cb4}} and thus the disagreement with the theory expectation. Compared to the most recent theory determination {{cite:fc0badc68830d0bd4e499affe564fa72f7ea3a9d}}, based on {{cite:ad3704b476e64acfd741d8d5aaa3ad1a8362229d}}, {{cite:90dd763d41a69b16cfbb31ff6c48701922b7b215}}, {{cite:02856ba0230348ae2262c236a6fb7443e2b37cc0}}, {{cite:2320ca1f1bff8810728a7af2c9434a1371895e06}}, {{cite:9c133428d18de221a0d6584f2d8a1093267991de}}, {{cite:e97f597ba2c69d4e3e0a550ae62c4dbe82d7b036}}, {{cite:1dc87e001dd96faf83ceef9203d0c1167042da06}}, {{cite:e34b7003f01409bb130675ad2ab3216bf97b33d5}}, {{cite:668b1310716721e9c69ae7518a8cb64bec59ab6b}}, {{cite:9fc0b00db1d63458f6dd0c0c8828d4ce42435d41}}, {{cite:7ea8df1038e33647fa25d2f52ea84a0e37d7366a}}, {{cite:604d23edab2694fec3d1dfc51278473e58eafe66}}, {{cite:72395ccd54cd699f5502c0fdf526ab507744cc9d}}, {{cite:09177bed588e75b3ac245a2fdc99b017b4acc743}}, {{cite:8a840298a2a7ed587fe041963b2a7d4515fff564}}, {{cite:a1700b90804654b965490bf9455dbe898399347d}}, {{cite:5c0fe0f317ba75eeab947f59b9dedc30922542a4}}, {{cite:4746eeff89b641c0f360f9819d2c3596e91f4d04}}, {{cite:54a060fbd77746bb361564e42552b0dea6c3b601}}, {{cite:05e2770ce08a649b516a52bad4e59585506920b9}}, the discrepancy is now pushed beyond four standard deviations. The Fermilab result is based on an analysis of the first data run, with a precision comparable to the one of the BNL number. In a matter of few years the precision will be improved by a factor four, thereby reaching the projected 0.14ppm.
| i | d0af2a8ba5baa731b6623cd0fd5b9e09 |
The role of the spontaneous breaking of the PQ symmetry can simply be played by a
new Higgs boson (a PQ breaking scalar), via an assumed double-well type potential, analogously to
the standard model {{cite:32b4c55e8437c91aa16ab015a147b02ab679f740}}, {{cite:a2789cf947016a2cee2672b2dd1f8c2c757e2873}}, {{cite:e9b99aac12ee2b60b630d6e2042316722f5a301a}}, {{cite:a01cee3b3b83e1f0ac0d28c3a90fef6a60e8b122}}.
However, this kind of simple and perturbative scenario inevitably suffers
from a hierarchy problem or instability for the PQ breaking scale ({{formula:df5abf9a-a85d-44bd-a7ee-8f5aeeb7b986}} )
against power corrections arising from the Planck scale,
similarly to the gauge hierarchy problem caused by the Higgs boson in the standard model Moreover,
quantum gravity would be expected to
break any global symmetry, which could destabilize
the axion relaxation mechanism
{{cite:bd2214948721d459dc8683a0e9eb2b405d09425c}}, {{cite:1edc8ee035fbd6add6f969da6daa4759546148d7}}, {{cite:7b604198d1f6627c7353d36504fe40eaee95a793}}, {{cite:7a2d270361565e03601a69cbdcfbb1f08fc2c105}}.
Later we will give comments on this issue in terms of the present framework..
| i | e6da962782952909b9c9b49d48ff94c2 |
Effect of the caption/embedding generation tools.
We investigate if the selection of the caption generation tool and embedding generation tool affects the attack performance.
To this end, we leverage pretrained CLIP to generate embeddings for generated captions, generated images, and query images, and ClipCap {{cite:cee9e0771c3f70564726bd16bc07d7230566083d}}, a state-of-the-art image captioning model, to generate caption for each query image.
As illustrated in figure:accclipcap, the attack performance is consistent with when we use BLIP as the caption and embedding generation tool.
Concretely, Attack IV and Attack II-S still achieve the best performance, with semantic-level attacks still performing better than pixel-level attacks.
Hence, we conclude that the choice of embedding and caption generation tools has negligible effect on the attack performance.
{{figure:34563f56-175c-41db-8d08-4809cbcf8980}}{{figure:09e89024-5672-4c64-97ad-1fddd0ed30fb}}{{figure:e3d1cdf6-a058-4bc8-983b-27d0c92d99b1}} | r | f0f415f0a9066621b3a11bb5fd68da44 |
Here, we define neuromorphic computing (NC) systems as non–von Neumann information-processing systems, the structure and function of which either emulate or simulate the neuronal dynamics of brains—especially of somas, but sometimes also synapses, dendrites, and axons—typically in the form of spiking neural networks (SNNs) {{cite:0dbc4067f7b15d98728dfdcedb2be08486cef724}}, {{cite:3d6065769773fd681e38a7c9e243d84601cf8f4f}}, {{cite:ae416f338f23d255a3bda2c7a1a5bfd1d12811d3}}.
NC systems open up new algorithmic spaces—through asynchronous massive parallelism, sparse, event-driven activity, and co-location of memory and processing {{cite:14a7c8f1714b5d720d4886f5aaf5c1ff13adb373}}—and, in terms of energy-usage and latency, offer superior solutions to a range of brain-like computational problems {{cite:4eab47c695ce542587303834449fd141d8aca06c}}, {{cite:d733ce5a5650d13b59927371d0716e402fbb8d7c}}, {{cite:8714d4af0fa89e61b82d420ef41d235662b1f369}}, {{cite:15f07b3ecf8b78ca1461c2b4643bc6eab4e6883e}}, {{cite:0abe3875412f51a553a5ada359c2fe5742c92d91}}.
Furthermore, beyond cognitive applications, SNNs and NC systems have also demonstrated potential for applications such as
graph algorithms,
constrained optimization,
random walks,
partial-differential-equation solving,
signal processing,
and algorithm composition {{cite:bf602a6dec77dbe780d91380bb7fdc6b37c6ee1a}}.
Consequently, there is a growing interest for NC technology within application domains such as
automotive technology,
digitized industrial production and monitoring,
mobile devices,
robotics,
biosensing (such as brain–machine interfaces and wearables),
prosthetics,
telecommunications-network (5G/6G) optimization,
and space technology.
| i | 4e37395b4573ce8ef153ccbafe2d8fd7 |
The overall results are listed in REF , which shows that our methods are on par with or exceeding previous state-of-the-art methods. Notably, our DUBD-B shows much better performances than other non-blind ones. Interestingly, DUBD-B and DUBD-NB show negligible performance gaps, even though the estimation is not quite perfect. In other words, our method is robust or not very sensitive to the conditional estimations. ATDNet {{cite:29c76179f513bf6d401b7263995ea7f3a5dfc517}} blind model shows comparable results to ours or sometimes shows the best results among the others. But the applicable range of noise-level of ATDNet is narrower than ours. Also, an interesting fact is that CBM3D {{cite:1649a2cdccb82dfdc3fb0a0abe64638959c57c49}}, which is a conventional approach exploiting non-local self-similarity with 3D transformed Wiener filter, shows quite great results better than most of the CNN-based methods in Urban100 {{cite:a3bdb6103e00eb29dfa7c8ae264439a27276fcef}} set that have many recurrent patterns.
| r | 7ad0c8c222ed693bf24de37e6133e4c5 |
The proposed two-phase flow solver is implemented in the CutFEM library {{cite:83849eb3e15da6b2eeb48429fa73f701b2d41e27}} based on FEniCS {{cite:e0c73860c0b92a582a657cc7e33ad68757082a25}}, {{cite:80d387893bad655e599abf7b45ee5c1ddc6a4040}}.
| r | f9f5d47da88e46a7f8438129abc6e526 |
We report results in Table REF for BPTT, SAB, UORO, RTRL on the Kodak dataset. All learning approaches can be seen consistently yielding lower distortion reconstruction as compared to JPEG, JP2, GOOG{{cite:251dd7c7cda6e48bccbb36f2b7a7ad7090111477}}, and E2E {{cite:fd4631ab899f8bd85b1323152b064d3c1ddc4d3a}}. Worthy of note, when using SAB, we noticed that there was a tendency to memorize the previous patch pattern for a longer duration compared to BPTT and other approaches {{cite:9e7ffbc0534d71ee0ca5c3b8198a35c92ba6afdf}}.
In terms of PSNR (on Kodak), we achieve nearly a {{formula:9a67928c-b4cb-40b1-b2ad-abbbcc83ad02}} decibel (dB) gain (with LSTM-JP2-SAB) over JPEG and a {{formula:fd589f39-4bea-400b-868f-b66b655fac0f}} dB gain (with LSTM-JP2-SAB) over JP2 when using SAB. With respect to GOOG, our LSTM-JP2-SAB estimator yields a gain of {{formula:1d6c66d9-f1fe-4ebc-952d-b2d433ce1198}} dB and when compared with LSTM-JP2-BPTT, our LSTM-JP2-SAB achieves a {{formula:0409cf06-1f5d-4c0b-bb7b-5aeb124c8c80}} dB improvement. The results, across all benchmark test-sets, for all metrics (PSNR, SSIM, and MS-SSIM), show that decoders learned with SAB and iterative refinement generate images with lower distortion and higher perceptual quality (as indicated by SSIM & MS-SSIM). One observation is that, according to our experiments, UORO and RTRL, despite having low PSNRs, have reasonably good perceptual quality (SSIM and {{formula:8e2808c3-c3af-45e5-a479-95d7eb1a7427}} – these two values are the same for both algorithms) as seen in Table REF ). In addition, note that, besides BPTT, all approaches require smaller {{formula:3d03a0ef-8db1-4c01-93fb-53f2a8e9e303}} .
| r | 97d78b483d9fe62b6172632f0c7cfd48 |
We compare MiFGD with publicly available implementations for QST reconstruction.
Two common techniques for QST, included in the qiskit-ignis distribution {{cite:f228afcd94a4cdc0488bf27223b67ac83912886a}}, are:
{{formula:3bd93c5c-1caa-475b-b70b-8572e7b58dc4}} the CVXPY fitter method, that uses the CVXPY convex optimization package {{cite:752a622484d437107784ecb091ecb59dc78d3bb1}}, {{cite:9420f07f72a0ec3c24f1df3e937e6a3391e3b749}}; and {{formula:16ee7cfa-33c8-4ad4-b57e-16d166e5d79f}} the lstsq method, that uses least-squares fitting {{cite:28c716ea4f10e1d5b4b849f000eaa164a9899214}}.
Both methods solve the full tomography problemIn Ref. {{cite:cc805507302eb3e053ee505a442c9a5058d3ecf9}} it was sown that the minimization program (REF ) yields a robust estimation of low-rank states in the compressed sensing. Thus, one can use CVXPY fitter method to solve (REF ) with {{formula:c323ea89-f179-419e-ab04-ef1d856de830}} Pauli expectation value to obtain a robust reconstruction of {{formula:c1b4ed29-a5c7-415f-8347-08e4307e4c9d}} . according to the following expression:
{{formula:59c1b37e-370a-4a05-af28-a2ef821be0ee}}
| m | db09cbc433700682e4cd7de26e6b4403 |
For the neural networks, we standardize the numerical covariates and encode the categorical covariates by entity embeddings {{cite:caf284f9102acbd2c912286927e7462a710ae5d3}} half the size of the number of categories. For the Classical Cox (Linear) regression, we one-hot encoded the categorical covariates (dummy variables), and in RSF we simply passed the covariates without any transformations.
The networks are standard multi-layer perceptrons with the same number of nodes in every layer, ReLU activations, and batch normalization between layers. We used dropout, normalized decoupled weight decay {{cite:c732752f1672fa1fc5a236d47325df48db53faff}}, and early stopping for regularization. SGD was performed by AdamWR {{cite:c732752f1672fa1fc5a236d47325df48db53faff}} with an initial cycle length of one epoch, and we double the cycle length after each cycle. Learning rates were found using the methods proposed by {{cite:e5d5189f43162b3b3b7c9ec9c2a24ad03040dc1b}}.
| m | 1d96f808f206e9bb0430eebecea28646 |
To jointly reason across agents in space, and their interaction, we employ aggregation mechanisms used widely in previous research works {{cite:6322f8c5dc04b05a558a198be0a6f02ec173ccd3}}, {{cite:b65aa07535108a0250006be2e57d098cbd63eed1}}, {{cite:80996340f3bbb79c5f7dfe624f8695671a0e94cd}}, {{cite:e62eb83d7f8eb4f9ca275589cc1a22d981911921}}. We use the social pooling from {{cite:6322f8c5dc04b05a558a198be0a6f02ec173ccd3}}, attention similar to {{cite:b65aa07535108a0250006be2e57d098cbd63eed1}} and a simple concatenation of hidden states of {{formula:bcaf417d-1a1a-4692-b4de-9daf56cd6d91}} nearest neighbours. The aggregation vector is computed using one of the three mechanisms per agent, and concatenated to its latent space.
| m | 61588f0fda9b64e13e82ae8fcba0e3d5 |
Remark 3.3
(a) Our proposed Algorithm requires, at each iteration, only one
projection onto the feasible set {{formula:9a5a8ba5-e039-4eaf-8965-0e59b72facf5}} and another projection onto the half-space {{formula:86edadf0-42da-43cb-936b-af0d6960be59}} (which has a closed form solution, {{cite:e8335df65828af879b74d6613a004342e9c86afd}}) and this is numerically less expensive than the twice computation of projection onto {{formula:5b82af57-1913-4c16-b8ed-f011cd50c25d}} per iteration in extragradient method {{cite:1a25dbc662ebec9621c3258ed8d34977dbe21eba}}.
| m | 70402d3e4032cb255b3b9bec7ed27735 |
Decision tree is one of the most commonly known model in machine learning for sublinear nearest neighbour search. However, training a tree-based classifier using templates directly are problematic due to insufficient training samples and noisy feature dimensions. In recent works on 3D object pose estimation using tree-based classifier, both {{cite:1e5b22f60ca0fe3fe7dca8dad096319d9ec5b27e}}, {{cite:12edf635c5b09000bb59cde2e8c367952f42db51}} use local feature-based representations instead of holistic template to alleviate the overfitting issue. However, this approach requires an additional geometric verification stage, Hough voting in {{cite:1e5b22f60ca0fe3fe7dca8dad096319d9ec5b27e}} and RANSAC-based optimisation in {{cite:12edf635c5b09000bb59cde2e8c367952f42db51}}, that likely to drag the process out too long to meet the requirement of real-time applications.
| m | 8d9bce74b47f4e46b2f7f07ab84eaf9d |
To improve the flexibility to complex geometries and reduce the cost to evaluation residual (REF ),
rather than the commonly utilized orthogonal polynomials {{cite:3efafdb19974d5d71c10875bef78800ade56462f}}, {{cite:4cd7f258045199dae98f0c53f4cb77727a47e796}}, a set of compactly supported test functions
{{formula:93df3080-7244-42ca-94d0-584da04ca2c5}}
| m | 00388dd30a7920a12a79ae2a18632c38 |
Figure REF
plots a 95% confidence interval on mean log-regret versus the budget used thus far. We perform experiments using a single budget in each problem; the goal
is to minimize regret at the point when the budget is fully exhausted at the right-hand edge of each plot. To focus attention on these budgets, we plot results over the range from 20% to 100% of this overall budget.
The results for {{formula:11671d28-48e2-48b6-ac2a-203c8270ba84}} look-ahead steps are deferred Section to the supplementary material to improve readability and also because they are outperformed by the {{formula:7ebb08c5-c2bd-493f-ba55-d165a5552422}} counterparts for most problems.
All implementations use BoTorch {{cite:760ed1c609f4eba3d833c978c49967f46a017e33}}. The objective and log-cost functions are modeled using independent GPs. Additional details and runtimes can also be found in Section of the supplementary material. An implementation of our algorithms and numerical experiments can be found at https://github.com/RaulAstudillo06/BudgetedBO.
| r | e5c5c4765e40b53d237812f0ae396ca7 |
The revival of the conformal bootstrap program {{cite:7a9fb14272ff2dcceed428c4c729e37378c79d30}}, has provided new tools to study non-perturbative physics. The numerical techniques introduced in {{cite:7a9fb14272ff2dcceed428c4c729e37378c79d30}} have given a wealth of results, with the most impressive being the high-precision estimates of the critical exponents of the {{formula:89af4bf8-4482-4359-8b50-ac8225d2d8b9}} Ising model {{cite:a6990f838e3db5b29dd0b2d440721f98d33b9dac}}, {{cite:c3c5eca4f54d1eadaff67bc3f2efd1091e82c399}}, {{cite:a5226f999356edce2c81c94caeffd93016db6c95}}, {{cite:afa8fbe258545d6819c84fd83844040dabc28474}}, {{cite:2ec531ad414b3ed95b67f864e762d576b16d26c5}}.
In a parallel line of development, analytic approaches to the bootstrap have also been explored, and recent progress has given access to the spectrum of conformal field theories (CFTs) at large spin by means of the lightcone limit {{cite:dd48e58a108b3588838a4a114236476c4ed6f554}}, {{cite:a9b99dd65de06f190f37989c560b518c9b64a5d4}}.
These two methods were combined in {{cite:8df180e02f06e2ebe009ea3cae10f92cc93d2cb6}}, {{cite:402c65a534938d9eb331a43bca084f3a3c512aa7}}, where knowledge of
operator dimensions and operator product expansion (OPE) coefficients, obtained numerically for the Ising model, was used to derive analytic approximations for the CFT data at large spin. Remarkably, the analytic results obtained matched the numerical data down to spin two.
| i | b8ca284b9e93f90993f52e0e1c8a1504 |
has smaller tensor rank than {{formula:6fd9030e-45b9-43e9-8866-281384554409}} .
The concept of representing
a function on a transformed coordinate system
has proven to be a successful technique for a wide
range of applications, e.g., {{cite:c298c859017214c835a5075f511196aeba632c33}}, {{cite:4f4265eff73aa23b0279c9a5278effe81b8f69db}}, {{cite:da7825aff850dac703c15eafc874f6575b576d69}}.
However, we are unaware of any work that has
used coordinate transformations for the purpose
of tensor rank reduction.
To illustrate the effects of coordinate transformations
on tensor rank, in Figure REF
we show that a simple two-dimensional rotation can increase the rank
of fully separated (i.e., rank one) Gaussian function significantly.
Vice versa, the inverse rotation can transform a rotated Gaussian with
high tensor rank into a rank one function. Similar results
hold of course in higher dimensions.
| i | 8e083b566c0883c1e7d2d14292742b03 |
Several recent literatures have dedicated to alleviating these issues by involving some localization priors of downstream tasks in advance. For example, the works of {{cite:1008431812e51846b7b637ca681789323783064f}}, {{cite:d6c41487dace8efa8f67cf3749df5aead231ecf3}} explored pixel-level consistency between two augmented views, while some other works proposed to match the representation of a set of pre-defined bounding-boxes {{cite:5b1f9437126d4f5979d539369cbe6b55d0c0e972}}, {{cite:c2d5006cdc2e6e7f00c6e66a4d06e0cc88889598}} or pre-computed masks {{cite:af0a04a37c3002d585e75b41ec9a08182229b1fc}} between the two views. Despite the improved performance on dense prediction tasks, these methods still suffer from several drawbacks, , they rely on the prior from a specific downstream task and fail to generalize to other tasks. Specifically, there is an undesirable trend that these methods perform worse on the classification task than their instance discrimination counterparts, since they are
delicatedly tailored for dense predictions and emphasis on the local feature learning.
{{figure:47d524c7-b753-4a52-aadc-ffb0858f98fb}} | i | 9a444a39c4b4d3ed6ffc8f662b648ba5 |
In the proposed method, we construct a correction model using Transformer, and incorporate a Chinese pre-trained model developed by {{cite:235d3a47cfabaa4a74cd32a4c4023e430400b94d}} in two ways as described in the following sections.
| m | 8f6221b7d1001dd341a10ffd288d3a66 |
where {{formula:b5a4cf1a-9db4-4720-9919-83df5226c205}} . In that paper, the main tools used are the variational methods combined with genus theory. The results in this paper were extended in {{cite:a25658a74bf4429e5eed1189fe7864a971b4e672}} including more nonlinearities. For further results
about multiplicity of normalized solutions for nonlinear Schrödinger equations and systems, see {{cite:9cc305debd7d7578cf8f58662ef8c0dd206592db}}, {{cite:d451b2d38adc803946b24e3ca8c66283e022afc4}}, {{cite:c9c39ab31eada214e61c0127811fea76dc25026f}}, {{cite:9168ee93b1e65b744138b163c28c5814aa65f982}}, {{cite:ccc7aca86d5cf81263baceb202046a5d1ec350d4}}, {{cite:ec1a1b4217ca95a6014d76a030a5e0484c27d1ad}} {{cite:b309adccf1feab687c62acc1dbb1a5fc195cd005}}, {{cite:0678ff2592b91e667992bd8032864f7f2e5424d0}}, {{cite:a8f554a9a93251461ef438637681e9d62e090bcf}} and references therein.
| i | b674b9228aa3d4caa4d110a35d9c09c2 |
Our results show that pure AdS{{formula:4a111fee-2b37-4a31-9fb7-3bb84946503a}} maximal supergravity saturates the most general non-perturbative bootstrap bounds in the large {{formula:0087d306-4710-409c-ba03-fe38a53fdca2}} regime, while CFTs with string/M-theory duals lie in the allowed region. This suggests that to study the latter theories, one needs to disallow the existence of the pure AdS{{formula:c1841330-9970-4402-9efb-2cd7501683aa}} theory by either looking at mixed correlator with other single trace operators {{cite:11c7107da5d8bf8f820cac76a63585bd88b13429}}, {{cite:9a235912d2c17d339d419bc00e151f8111e35968}}, or imposing theory specific constraints like supersymmetric localization {{cite:9d2b65d473881653a6f1ccee5ea927bb60f8a9a5}}. Indeed, in 3d one can strengthen these general bootstrap bounds by inputting the OPE coefficients of the {{formula:820ba693-8012-4cea-bf27-155d6e9c5be8}} and {{formula:627f3b52-a08e-4293-bdf7-00fc777b2ae4}} multiplets for the {{formula:e2d37207-7530-4bc8-80b9-d8c42cb9cd9c}} ABJM theory for {{formula:67b33c82-7642-4b14-9564-008b93041bcd}} , as computed to all orders in {{formula:da0a961d-a256-47c3-954a-2f2a7ca33f22}} using localization in {{cite:b0db0b2dc5662919ab7aab49178cfe02f525fa08}}, in which case the 1-loop data for the dual {{formula:c3e88755-6dac-4275-918a-a376c3d4200d}} theories then saturates the bounds {{cite:e12d84d5c2090b635a2c5b52557b6bba5580f37b}}, {{cite:edd518206594553dc68e1ef8f55604c17c304d7c}}. In 4d, one can input the two localization inputs for {{formula:b4418a93-d76f-4551-887c-afe4083d48dc}} SYM derived in {{cite:1d9b356b3f67b356ab20715809b792e44b909b48}}, {{cite:e40cc90ebce7294b45c86c6f76fb064d49b803b8}}, which are a function of the complexified coupling {{formula:b8b17542-1a9a-479e-8a7f-c6e0f31d6376}} , in which case the the bounds in {{cite:6596c59fd527f755601e44a2a10ee8f915bed981}} match 4-loop weak coupling results {{cite:201f148b12fae1ed4e8016dbb0f7b448830136ce}} in the appropriate regime, and exclude the general bootstrap bounds shown here for all {{formula:78352166-31ac-4485-81ce-640f7108f4fc}} . In 6d there is no localization, but for correlators of single trace operators other than the stress tensor one can input nontrivial OPE coefficients given by the protected 2d chiral algebra {{cite:e9023684338fea3fb381a41b0c5223c56d758fe9}}, {{cite:26225a1b94e99286b7ca9f8c91cbd988ed874e1b}} for the {{formula:96c25019-1a41-4dad-9655-9054740c1bb8}} or {{formula:4e0377f2-c4a1-4f9e-ba74-5fc4b646dcc8}} theories
| d | 93ba3cbc637e518b99f33dc549153e49 |
We also note that stars with different ages have significant overlap in their distributions of [X/Fe], {{formula:5367fcfe-9def-48e1-8d42-a19e68f5a3d4}} , and current R, but that the peak of those distributions are distinct from each other. For younger stars, the peak of the radial distributions of stars is at larger R, and the abundance distributions exhibit progressive enrichment. There is an overlap in the [X/Fe] KDEs, which is manifested in the gentle slopes in the age-[X/Fe] trends. The element N, whose yields from SNe II have a progenitor metallicity-dependence, has the most distinct [N/Fe] distributions as a function of age (i.e., [N/Fe] distributions overlap the least at varying age bins), and therefore also the steepest age-[X/Fe] trend. On the other hand, O, whose yields from stellar winds have a progenitor metallicity-dependence, has abundance KDEs that overlap for different ages, and therefore exhibits the flattest trend with age. We therefore find that for the simulated Milky Way-like galaxy m12i, the direction of the age-[X/Fe] trends is a reflection of the chemical evolution prescriptions in the simulations, and the scatter around these relations is a reflection of the mode of star formation, with higher scatter at earlier times when star formation was more bursty and chaotic, and lower scatter at later times when the disk is more stable. Meanwhile, the magnitude of the slopes in the age-[X/Fe] relations is a convolution of the effects from the chemical evolution prescription as well as the star formation history. With the next generation of spectroscopic surveys dedicated to densely mapping stars in the Milky Way such as SDSS-V {{cite:1e177aeba057d7b20042e53fc6bfc8bda8b42b84}}, WEAVE {{cite:2b280c6af33ae3f24ad9edc3f20c4ff138cd259b}}, and 4MOST {{cite:a09985662a05daa1f2bb7890fecf00e1362c7cf1}}, as well as novel ways to derive ages for these stars, we can use the {{formula:f73760e2-3434-4dc6-b24a-c4de11ecf6f6}} in the age-[X/Fe] trend as another diagnostic in understanding the evolution of the Milky Way.
| d | f7bcf96ff4a928692a6224eb67c0a2d2 |
At first, we treated the angular momentum as the control parameter. Similar to the treatment of the photon {{cite:656148da438dc765476362cc1f5f99329ea64d63}}, {{cite:14179953e04e3e4a14f015504bc72915faaabfa4}}, we started with the time-like geodesics and constructed a vector {{formula:e8db7e9b-f8df-4521-8106-06ea11c8a329}} by using the effective potential. The TCO exactly locates at the zero point of {{formula:d0b8254c-85ed-48ad-b164-9d25527e8a2e}} . Thus one can endow them with a local topological charge, the winding number, for the TCO. Moreover, we studied the asymptotic behaviors of {{formula:786fa73a-04f3-4a09-aad0-022c42eda6e0}} near the boundaries in the {{formula:a8c04627-5f16-4bcd-9980-33b99bee66e9}} -{{formula:c9813bd5-d846-45c8-a1ab-8e2b938f28da}} plane. Via examining the axis limit, horizon limit, and asymptotic limit, we found the angular momentum does not affect the asymptotic behaviors of the vector. This strongly implies that the topological approach can be applied to the TCOs, just like that for the light ring.
| d | be659f064073d6543fb5e55c9ebafae7 |
where {{formula:82335437-55d4-459f-848d-caddc75124c9}} is the expectation operator.
The OMLSA adopts the IMCRA {{cite:67b1e670200e6fec9d58b37017f2242a41b931e9}} to estimate the noise power spectrum.
After the estimation, this method reconstructs the enhanced speech signal by minimizing the mean-square error of the log-spectral amplitude.
Under the constraint of {{formula:4427378a-d59a-4e01-95d9-6eae562050f5}} assuming greater values than a threshold {{formula:ed1454a9-2042-43fe-83b3-9fe08fb9ce91}} when speech is absent it is shown {{cite:09d87ed7985d396c7fafec2370864efeecc8638c}} that
{{formula:9fc06dd3-aae5-4076-abad-b00ac6908b40}}
| m | d50944ae4384b1a73ab0e918b88251f5 |
Figures REF (a) and (b) demonstrate
how individual factors in Eq. (REF )
behave as a function of the creation beam cross sections {{formula:f9254e03-da5c-4c3a-ba27-978896f62a55}}
with respect to a fixed inducing beam cross section {{formula:3c865243-3af5-4c59-bf04-c31aaa34325c}} , represented by the yellow dashed line.
The assumed ALP mass was {{formula:03c58f03-43ac-4c36-b0bf-c007872c2f8f}} eV and {{formula:3190b479-8f79-421a-a62f-ef4ffd8dba50}} eV, respectively.
Individual factors are explained with different colors
inside the figures and have individual
normalizations of {{formula:91df6277-6267-4153-abdc-9f7675d1c59f}} mm{{formula:63c59f70-1846-4a0e-98c0-0ca6e33e1d3c}} ,
which is close to the actual value of the maximum creation beam cross sections used in this work.
What is important is the behavior of {{formula:b4224b4f-d41c-4d27-91b1-86c83d701873}} in the region {{formula:f867d46e-9e2c-4373-8bfd-9779faadd6e9}} ,
shown with the red curves.
Depending on the ALP mass,
{{formula:5ef90b18-e70a-4567-b0d6-34f2bb66ee31}} behaves differently, and the {{formula:984439e5-fa33-419e-b05a-f46828b2ccb1}} dependence
on {{formula:3177ecdd-d225-4631-b720-a433721a9b8c}} obviously deviates from the constant behavior
seen when optical-element aFWM contributes as the sole background source.
Therefore, if {{formula:e25db416-7245-47e7-8bf1-f67fb73184c6}} exhibits a significant non-constant behavior for {{formula:23f364f6-c736-449c-8a46-123548abd921}} while in a low-pressure environment, then, in principle,
we can determine the mass scale of ALPs from
the creation beam cross-section dependence {{cite:865c74f2cdea0f3d81822102b5ddf4e18ccba9be}}
in addition to testing the existence of ALPs.
| m | 6d6415bc9583e876af6066e8e1a7681c |
The deep neural network (DNN) {{cite:585d91003df4812c5589186dd77fe31ba563e283}} has become the most powerful machine learning method, which has been successfully applied in computer vision, natural language processing, and many other fields.
One of the major problems of DNN is its robustness against noises, outliers, and adversaries. There exist vast literature on improving the robustness of DNNs {{cite:5e9a64242a19cc2dc5defe965a72c94ecae51302}}, {{cite:57d3dc052ba98c53389ee91069dc9bbfdb578323}}, {{cite:1870a43d3afc32b69713f127713e2d69b06b3eed}}, {{cite:6d48c6b9a67ae05b83910a4721e2497a6bf97d52}}.
In this paper, we present a new approach by changing the autoencoder from an un-supervised learning model into a classifier.
| i | fe9c9978b3646dde156ef4bfeaecab76 |
It turned out that RNNs are simultaneously capable of both information import and information processing only in the low-density, i.e. sparse, part of the classical edge of chaos. Remarkably, this region of the phase space corresponds to the connectivity statistics known from the brain, in particular the cerebral cortex {{cite:ae1de13f6c91e3e2e992a48ec06983989fc42046}}, {{cite:20f4b7e9dcc7ca3bf7bfec2d18a42797303745a6}}, {{cite:8c476cb6a0eb76d5939066f83697ce4ad97e420c}}. In line with previous findings, i.e. that sparsity prevents RNNs from overfitting {{cite:047f018543386941b188b27552e143d61d2b7448}}, {{cite:666ec6d8d58c9e8e7525b4d3fac2937c5a6c3dd6}} and is optimal for information storage {{cite:ff1823584129e611cce075f81f3f9296561f261f}}, we therefore hypothesize that cortical connectivity is optimized for both information import and processing. In addition, it seems plausible that there might be distinct networks in the brain that are either specialized to import and to represent information, or to process information and perform computations.
| d | 970843f6f62f766b23af0f1c93409c90 |
Network Slimming on PreResNet-164: We have reproduced the results by using the same pretraining, pruning and finetuning strategy is used by {{cite:b6508bf02d55eadf6a22a6ba572f370e198c4c2a}} and the same pretraining and finetuning strategy is used for our results in order to do fair comparison of pruning algorithms.
| m | 12501f2cd14b357dcea4a8e302cbab3c |
Recent works demonstrate that for semantic classification tasks, contrastive pre-training approaches result in better transfer performance compared to non-contrastive pretext tasks {{cite:c3c2506610976ab61d764ac472d2c4516595794a}}, {{cite:793d03f0ca223c67f838bc336d17db1e67eae4f4}}.
In this section, we investigate if the same trend holds for pixel-wise regression tasks.
In Fig. REF , we show transfer performance results for colorization and jigsaw as compared to contrastive methods.
For classification tasks, we replicate the common observation that contrastive methods provide a significant improvement compared to non-contrastive methods.
However, we again observe that the difference is less pronounced on pixel-wise regression tasks, echoing the conclusion of Section REF .
| m | 1f87d3e63f9628c19dd9a6110d0f0bde |
fig:bananaconditioningcomparison shows that using this form of Dual Conditioning outlined in eq:fastcondit we are able to recover a very similar posterior to the offline solution. To build a batch for BO and AL applications, we follow the Kriging Believer algorithm (see Alg. 1 on p. 17 in {{cite:b4b3f72def93abe904bc5417c8c0013b5856d6b0}}): we fantasize data points given from an acquisition function {{formula:cfa49fe8-ec70-47b6-8fea-ee9c65993d21}} and then use eq:fastcondit to add the fantasized points. We outline our algorithm in alg:BO/AL and optimize {{formula:37c400c8-dcdc-496a-9109-a76356fc6077}} using {{cite:d921cb6e2863f920182f767d20181151a16a0272}}. We keep inducing inputs {{formula:d8da1347-199d-440e-8634-006d8f6943b3}} and hyperparameters {{formula:8c185fd9-1042-45f1-8312-8a2f90a02d1b}} fixed within this inner loop of optimizing the acquisition function, but update them in the outer loop when we obtain new experimental data.
| m | dd6e73b1e2e2af50ee4eea908c1a66a1 |
In this section, we present our approach for relation modeling in spatio-temporal action localization. Firstly, we introduce our overall pipeline for this task. Then relation modeling module is presented with transformer-based architectures to capture the relations among persons in the spatial and temporal dimensions. Furthermore, we adapt memory bank for storing person features along the temporal context of video clips to model long-range relations. Different strategies of online or offline maintaining memory bank on the AVA-Kinetics Crossover are studied. Finally, we investigate learning approaches for the lone-tailed category distribution in the AVA-Kinetics dataset {{cite:adcd6779aec8cd0a9d26b92bfd6b8d05da9538da}}, {{cite:632d81892f1f4e2763aa08797f0e76932efd89d6}}.
| m | 38c3ae717d361f4568bab8854fba0d44 |
To investigate the effect of rotation in a phenomenologically interesting case of charmonium one would have to consider rotation of the spherical symmetric potential, which has to be dealt with numerically. Nevertheless, we can make a ballpark estimate using (). The binding energy of {{formula:7892a5e4-828d-4c96-8586-a2495cf582fe}} is {{formula:3263141c-aa13-4ae5-9635-5ed3dbcb799d}} GeV. The reduced mass is {{formula:75ad9546-33a0-488e-8011-f1d11b0e4e25}} GeV which implies that {{formula:53a42ebf-d6c1-44bd-afb6-b99c3578fe16}} . Using the values {{formula:b0e2a58c-5ec5-45c5-be58-7bff5236fe31}} and {{formula:9d3fd10b-c8ed-4e19-a422-9e160ef1220f}} listed at the end of Sec. we derive that {{formula:4f1cc55d-755b-4b40-a010-47f91da39b6b}} and {{formula:fdde9e32-0748-4d1d-a465-8ae28ef7f581}} dissociate into the continuum when they get swirled by a vortex rotating with angular velocity {{formula:b281af95-502f-4abb-b748-9e2a05592804}} fm{{formula:61baa73e-6db2-4bc3-8bc0-c028dab6e1eb}} and {{formula:ee45dbc7-4d87-4e7d-a19c-3a27a825feb0}} fm{{formula:d98cd643-fa0e-4773-b33d-8f0c4b9c4091}} respectively. These estimates of {{formula:389151b9-cb60-4522-ba18-2a94ca3f9cb4}} are somewhat larger than the vorticity achievable at a (relatively) low energy relativistic heavy ion collisions {{cite:005369b09d68b53782e85e3cefee65168946f941}}. We stress that at the above values of vorticity, the bound states completely dissolve, however the effect of rotation — the decrease of the binding energy with {{formula:439b0434-69f0-432a-88a7-9cb9f23fc37c}} — is essential even at sub-critical angular velocities. Needless to say that these estimates should be taken only as a motivation for further investigation with more accurate quarkonium models.
| d | 9f2d33d9d1f9020391c725a46a0cba43 |
A somewhat related idea is to use the sequence of linear approximations to maximum likelihood estimators provided by the Fisher scoring algorithm {{cite:abcc3262eab186ffcf219a21a76d623221458cad}}. For generalized linear models the associated derived response and weights based on the score equation have a convenient canonical form and the solutions to the iteratively reweighted least squares equations eventually coincide with the maximum likelihood estimates. See {{cite:a9b4926f6cdbc1274df7987864239953186ef752}} for an extensive discussion.
| d | f26ee549bf9dd93b3725f5c100b468a6 |
There is a large body of literature on the evolution and ecology of migration and dispersal {{cite:149f9d4f4fad44edcc3546c257cd9a4f47efefb6}}, {{cite:7a883407a608071f2b891fc33589d7bf4fdfc89c}}, {{cite:603ada1d405a966472e582867863926297e7c221}}, {{cite:4e0d3eb6efa23ba1fc06c6af23537f3afa1963db}}, {{cite:d84d842bb14b572b5dc547992f6afa0bdbad9b48}},
especially for population structures formed by islands (also called patches, demes, or metapopulations) {{cite:16f00dc5639bc94a6281a109c4b5ab3425fa7955}}, {{cite:8777ca3240fc7573027db4cbab25510a71125d13}}, {{cite:5985e7c88f8e6f2089d9a1076ff431a69edb4f35}}.
Our framework is a generalization of this approach in the same way that evolutionary graph theory is a generalization of the
vast literature on evolution and ecology in spatially structured populations {{cite:872018d62f635f149088029777d5023c779c2621}}, {{cite:3cb3ec4d72e6f450a64396a783d749e9705c78e1}}.
The framework is flexible, allowing us to study both simple and arbitrarily complex population structures of any population size.
As such, it facilitates a discovery of new phenomena.
| i | f2cb5e1ced16108b949134ec4890cd52 |
We have proposed three `fairness through awareness' bias mitigation strategies. The first strategy (Stratified batch sampling) ensures that in each batch all protected groups are equally represented. The second strategy (Fair meta-learning for segmentation) aims to make the model more aware of the protected attribute(s) by simultaneously training a protected attribute classifier. The last strategy (Protected group models) aims to train independent segmentation models for each protected group. Our results show that all three strategies improve fairness, with lower SD and SER values compared to the baseline model. These results are in keeping with previous works where the strategy of `fairness through unawareness' has shown to be insufficient in preventing bias in several problems {{cite:7bfbcaff06ce0b3b11ab36d226f689f4d5e9d732}}, {{cite:679870aacf9d285d30d584f80ac129a92b960232}}.
| d | dfa1ae6728bfdbba831957aa51d8595f |
In theory, the {{formula:6da3687d-8553-47d9-97b0-cf9f36adc770}} mass spectrum was widely discussed within various quark models, for example,
the relativized quark model {{cite:b2127104a07ba8239fa18e05b8f58b758d07efb4}}, {{cite:cd5ca7e5090d2f88ba3f20e8bde54ca3663767ca}},
the nonrelativistic covariant oscillator quark model {{cite:1e8b4f5c7f393256aac2af35d701b23bbda704c2}},
the QCD-motivated relativistic quark model {{cite:c2b8a4cb27d063b0fb14a62022e7b482ccbca8df}},
the nonrelativistic constituent quark model constrained
in the study of the {{formula:7124b60e-998c-4682-ae9e-6986c6ca2406}} phenomenology and the baryon spectrum {{cite:190f276981b55572e83a9cbd226b16c3a650dba8}},
the nonrelativistic constituent quark potential model {{cite:8a963b9a45310d4c4fb0094c4877d8eacffece50}},
the extended Nambu-Jona-Lasinio quark model {{cite:fbead0192ebcd2fe88f92ef7e7961713413967cb}}, {{cite:eb38f5d5bc8ca0668a16c6deba81bd8b2cbeef86}},
the approach of Regge trajectories {{cite:6d202faa2405ef13d933a01988e9feb0712fcc91}}, ,
the modified relativized quark model {{cite:513fbb270d7f52322c1de54bade5807883ac6de6}}, the framework of
the Bethe-Salpeter equation {{cite:f267e0eeaf143e100bf0b291f1f52e7cb8ef565d}}, {{cite:c9de0f947a1751ac1f8067dcd1c4ea0f56319972}}, {{cite:f5dfdf0a7376fc9942cf1ca5b7b789be7460b180}}, and so on.
Furthermore, the strong decay properties of the strangeonia
were studied within the pseudoscalar emission model {{cite:b2127104a07ba8239fa18e05b8f58b758d07efb4}},
the flux-tube breaking model {{cite:299d47ef12edf725aa9359d952fc28977c9af98d}}, {{cite:f2f1d936180000eb568307226b69aaa21661c26b}},
the {{formula:371dc691-c50c-4dc8-b28d-fa0062b69529}} model {{cite:c7d11c4d3d40689970017d452adf685e1b1f8924}}, the corrected {{formula:517f95d4-1ead-43db-ae65-e2c1c66cbaa9}} model {{cite:1c5830ebde061ad0f7a05335c635eecc47ec9871}},
the framework of relativistic quark model {{cite:f267e0eeaf143e100bf0b291f1f52e7cb8ef565d}}, and so on.
However, a systematic study of both the {{formula:43a03dd1-ef60-4f54-9f78-03b289836532}} mass spectrum and their decays by combining the
recent experimental progress is not found in the literature. An early review of
the status of the {{formula:12f29fdf-c279-4e4e-9840-52a226e70580}} spectrum can be found in Ref. {{cite:1cb4b0e4259a0fc9b3f1883240115ee5bb03ef68}}.
| i | 13349c555170e15c1e7a77d2a9527f90 |
A word of explanation behind topological properties of filters is required. Since there is a natural correspondence between subsets of the set of natural numbers and 0-1 sequences, one may treat filters on this set as subsets of the Cantor set {{formula:dcd8bfb9-07a0-475d-9bf8-8576ab17f334}} equipped with the product topology and consider properties such as being Borel, analytic (continuous image
of a Polish space), etc. that we shall freely do. For more information, we refer to {{cite:b49e699c6ae4d941cb277ded011850ad7878b8f2}}, where the the dual notion of an ideal is considered. All unexplained terminology from Descriptive Set Theory (such as the definitions of the {{formula:41c90750-46f9-48b6-af8b-e90cb5fe4f58}} , {{formula:159e83f6-1571-4ed7-b436-55269accebae}} , and {{formula:f609b9ab-2d51-4fe4-86f3-d303ba939f02}} -classes, {{formula:d04e0aae-d64f-4fa2-9ac2-e89dfff7ed24}} , {{formula:fed75c31-a870-4093-ac26-9008ead7d0c2}} , and of Baire-measurability) is in line with {{cite:b1b98311c29db709c8cb43b03a179c7f04374540}}.
| i | 8afe14b8659ce204705e7d318b8982a9 |
In this section, we evaluate the proposed DAP3D-Net for action parsing in videos on two datasets: NASA and HAU. Our method is implemented using Caffe {{cite:439d39242b06927b0d50e3bfc9150d036a137de6}} deep learning framework and we further modify various layers from {{cite:4836e8abf21e028b4b4e88ba82dfe3463bd69b09}} to specifically fit for our architecture. DAP3D-Net is trained on a workstation configured with GTX TITAN X GPU.
| r | 615a4076eed861aecb845785d4d5d361 |
Note that, for computing {{formula:f2ebabe9-f299-4b2c-97a6-66c7acb80973}} , we do not need to evaluate (the possibly expensive) functional {{formula:32af94ee-aac6-4d9d-9b73-978f27c94d8d}} , but instead use a quadratic approximation around the current iterate {{formula:7b15c9d4-430f-4d07-a2ad-5a7146277c62}} .
One of these variants is the TR-Newton-CG Steihaug method, presented in {{cite:bde5bce9950f1511843826c3b7ca742b2eefeee0}}, where CG approximations are used for the Hessian.
The variant will be used in the sequel of this thesis to mimic a truncated Newton line-search algorithm; cf. sec:exppaper2.
| m | fedd133053dd12a5790a8fa1e46e8fec |
Comparing with the constraints presented in Ref. {{cite:5f1e17432513bbde54aca0d9191a8d3affb9cd48}} for Planck 2015, we find that, while the impact of our improved treatment is clearly visible in the CMB power spectra (and will be relevant for future experiments), it has only a marginal impact on the constraints, and our bounds are in very good agreement with those derived in Ref. {{cite:5f1e17432513bbde54aca0d9191a8d3affb9cd48}}, which only included the leading order term in the daughter radiation hierarchyLet us note that the implementation of the BAO/{{formula:d56aa3e9-75a7-421a-9c5f-c834038118fd}} DR12 likelihood used in Ref. {{cite:5f1e17432513bbde54aca0d9191a8d3affb9cd48}} within the MontePython code had an issue that led to constraints on {{formula:7c1a6306-92d8-46b4-b61a-ed34a1fd2e9b}} that were somewhat milder than the true bounds. MontePython has since then been corrected, leading to an improvement on the constraints on the stable/long-lived ({{formula:98c1e796-e449-4541-845c-ee50a57af0d6}} km/s/Mpc) case by about {{formula:e3984685-f012-43cf-a970-c28c85e4fba9}} . However, we have verified that this bug had no impact in the short-lived case ({{formula:cd9dbef4-78c0-45c2-a2b5-24d8c586a91b}} km/s/Mpc)..
The bulk of the improvement is due to the newest Planck 2018 data and can be understood as follows.
As shown in Fig. REF , for the masses we consider, the main effect is an almost scale independent suppression of CMB lensing spectrum.
This suppression can be compensated for by increasing the primordial amplitude {{formula:f89e8a1a-9de2-46f5-a612-bebf2410ff59}} or by adjusting the matter density {{formula:785321ff-1ba5-4462-afd5-09fc57dd8e5b}} (see Ref. {{cite:f8d7a3c550714ccf4259bd9df32ce6c60dbb5574}} for a discussion of the correlation between {{formula:c4e2fd7a-42e8-4633-bef5-1602b2bf7d92}} ).
Due to the well-known degeneracy between {{formula:e6bf83a0-3360-4ce2-9490-35822d48627b}} and {{formula:8491a278-90ee-4969-8ae9-ebd5688e0438}} , Planck 2015 data, which was limited in polarization, were unable to place a tight constraint on {{formula:cc6d907c-d760-4e9a-9a60-c9ef2377b0b2}} , and thus the constraining power on the sum of neutrino mass and lifetime was limited.
The precise measurements of low-{{formula:ed9735f1-80b8-4335-87ff-5c5f31b1a8ff}} polarization from Planck 2018 leads to constraints on {{formula:5f51af14-0962-406c-b398-5008344f31ed}} that are tighter by a factor of two than those from Planck 2015.
As a result, parameters degenerate with {{formula:0e9400a0-b65a-4149-ab7b-efa9ebdb45c2}} such as {{formula:0e38737b-77f2-4c8d-aac2-577c9a98e6dc}} are now much better constrained.
Consequently, the constraints on the sum of neutrino mass and lifetime have significantly improved with Planck 2018 data.
To confirm this simple argument, we perform another MCMC run with Planck 2015 data and a tight gaussian prior on {{formula:0af8e96e-0a69-417e-888b-0b584c9eaa6b}} , chosen to match the optical depth to reionization reconstructed from Planck 2018.
Given that the constraints on {{formula:ff3546c8-ae5c-4b5c-92ed-126b27c9fa21}} are independent of {{formula:0b9ae6c1-5224-4815-931f-3eb5488fc877}} below {{formula:3ba50243-8441-49db-bf9c-c7226f151d7f}} , and the scaling above {{formula:51a5c309-1b94-432b-8897-ade82096e14e}} is monotonic, we focus on the parameter space {{formula:bbcc078d-9ce6-4540-82ea-505f2858f9c1}} to accelerate convergence.
Our results are presented in Fig. REF , where one can see that this simple prescription leads to constraints that are very similar to those from the full Planck 2018 data.
We attribute the remaining differences to the additional constraining power of Planck 2018 data on the parameters {{formula:9140e61e-8a7c-44e4-8a73-20fc1842f8bb}} and {{formula:2791bee5-edac-4c02-b538-d75e539b851e}} , which are mildly correlated with {{formula:acea0ff0-4f53-4f10-a09d-936bd2dafe2c}} (see Fig. REF , top panel).
Note that our constraints are a factor of two weaker than those advocated in Ref. {{cite:49ed065008b3f72da1fdbe309ca32dad8389a5c0}}, which performed a `model-independent' reconstruction of the neutrino mass as a function of redshift, but neglects the decay products.
As we show here, including details about the daughter radiation is necessary to accurately compute the effect of neutrino decays even in the non-relativistic regime.
Finally, as discussed in Refs. {{cite:f8d7a3c550714ccf4259bd9df32ce6c60dbb5574}}, {{cite:0f0d142873b34139b500d6672e5b14f319bde12b}}, a combination of CMB data with future tomographic measurements of the power spectrum by DESI {{cite:55006c97dc112c0ef58548219ebbe969007414fa}} or Euclid {{cite:921bd8895dd59950dafbd63fb4512a10a7a5460a}}, and an improved determination of the optical depth to reionization by 21-cm observations with SKA {{cite:616d46377f8def26aae71c40f3252ca6c4b9ca66}}, {{cite:75c00eebd71faebe3652ffd0085b51216ea6c62d}}, could greatly increase the sensitivity of cosmological probes to neutrino masses and lifetimes.
{{figure:93424472-a2c5-4ee3-8c7c-03adf0f9480c}} | r | 5d5041e808ebd0f1914662aa36ba6f21 |
The lifted form of the ILC's input variable {{formula:208ee551-4888-4d34-9f38-83d2749e3c73}} is denoted by {{formula:2e61997f-3382-4cc8-a549-cc08c5bfb17c}} , where {{formula:f2cd4b60-d15c-46b0-a129-46d0e34fdc2c}} denotes the trial index. To update the input trajectory {{formula:dcb743c8-b36e-4042-a483-ac6bffcffdaf}} , an ILC law as in (REF ),
is applied, where {{formula:e55fed91-56fd-4109-9208-562ccb9f51f4}} denotes the lifted form of the learning function and {{formula:5ab93309-27bf-4758-b606-af5500a2ed52}} denotes the lifted form of the Q-filter. To design both these transfer functions, the dynamics of the closed loop plant are linearised leading to the lifted form {{formula:a6542d95-9a8a-4ca8-8840-416c8dc007b8}} and the linear dynamics (REF ),
where {{formula:361522ca-90ed-43e1-8598-8eeecdf55d67}} denotes the pitch trajectory. The learning matrix and Q-filter are calculated using quadratic optimal design, see {{cite:c6d3832e5ab8a0e7ddf25e3f79cdd83c6ed58468}}, yielding an asymptotically stable, monotonically convergent ILC system with the convergence rates
{{formula:86024b27-61db-46d5-be35-a998a479b5af}}
| r | 29b1f916c42037f86c66a85e9842e945 |
FPN. Feature Pyramid Network (FPN) {{cite:77ccf37cda70b50dd59f76cd6f7d0eee28aeca3d}} is to explicitly address the problem of large scale span. The MFD task contains a large number of extremely small formulas, which bring great challenges to our model. As illustrated in Figure REF , for such a single extremely small character formula, their short sides are usually around 16 pixels. We observe that for any layer of FPN, the limit of the detector is 3 pixels. This means that if we use the default FPN (3-7), the short side of the formula needs to be at least 24 pixels to be detected. Obviously, there are many small embedded formulas that do not satisfy this condition. So we change the selection of FPN level to (2-6) so that our model can overcome this defect.
{{figure:b67e03cd-fca9-41c7-8524-bd69232d3bd9}} | m | 582c69e0b0ed3d697afa4432edd31b5c |
In this paper we have presented the discovery of the Giant Arc (GA): a
{{formula:bf7e1e8d-0e02-4edb-976c-247a743fd98f}} Gpc LSS at {{formula:a7cd6f94-a83f-4e60-aa57-993b48c3f47b}} , mapped by Mg II absorption systems in
the spectra of background quasars. The GA forms a large crescent shape on the
sky that appears almost symmetrical. However, deeper analysis reveals some
asymmetries in the GA, in the redshift and equivalent width (EW)
distributions. The GA spans {{formula:d1abf4ec-8a0b-4669-b037-9c15101ddc37}} Gpc on the sky and has a redshift depth
of {{formula:3a035e6a-8bec-49a6-99f1-1aeb494d0ae4}} Mpc (both proper sizes, present epoch) . Visually, we determine
the GA as a single unit, but using a Minimal Spanning Tree (MST) type
algorithm (see Section REF ) it splits into two portions: a
large portion (GA-main) and a small portion (GA-sub). We proposed in
Section REF that the two portions of the GA could in fact be
connected in reality, since potentially one more background probe could lead
to one more Mg II absorber that would connect the two portions. On its
own, GA-main is a statistically-significant clustering of Mg II
absorbers, with a membership of 44 Mg II absorbers, an MST-overdensity
of {{formula:27c7dadd-2dbb-4ebd-b388-055c2e32e36b}} , and a mass excess of {{formula:fe1873d5-df9b-4700-8c42-87cca02120fc}} . In these
respects, the GA is comparable to the Huge-LQG {{cite:c2d7706d73d0f24f397f18ad2397caeca060543f}}.
| d | 850a22048a4aa4e23987f522fb1f2b17 |
The canonical map {{formula:69462f79-3820-4720-ad08-d62b8429f483}} , the inverse limit of the bonding maps {{formula:444dedb0-ab74-455c-9156-59252cd17414}} , sends
a point {{formula:417302a7-6508-4064-8c7a-77ce9eb7aaea}} to the intersection with {{formula:f92fad07-9bb4-4f60-8a9f-cd142df56c49}} of the ray starting in {{formula:a61f4283-1299-4095-95d5-d5d4beeadcee}} and determined by {{formula:3fcdc7d1-0943-42c1-bc53-99b18c158836}} {{cite:b08066961f27fe7a80d4bcce9989f4efa2296133}}.
| r | ae413e96ff4199839a45a360f3d6ac28 |
Simple optimization algorithms are surprisingly able to find uneventful descent paths in the non-convex cost landscape of deep learning networks {{cite:c7abf72490301f9d33407e06774b26a5c269fff3}}, {{cite:f8be7d72c013eb993a4c9f74f21b17d47a8e78bb}}. However they also tend to construct features that capture spurious correlations {{cite:555e653e194304c68586718b29063d87b30765ff}}, {{cite:f0ec56713429b0ef6cc4e90ed1fbc071fe675553}}, {{cite:b622e5883eb8d905d1a4f8868613ecef567d6b7e}}, {{cite:94adc182de023fd5ee356539ee1b23c778880893}}. Several recent papers propose to work around this problem by leveraging multiple training sets that illustrate the possible changes in distribution. The training algorithm must then satisfy certain constraints across training sets, usually enforced with additional penalty terms {{cite:b622e5883eb8d905d1a4f8868613ecef567d6b7e}}, {{cite:91b79a4fc56a1e753159087c969d68b4f50399ad}}, {{cite:5bf3c02fcd014836397b89194bf8834ef7749503}}, {{cite:b20d1656b04a828bed948720099a633088780959}}, {{cite:3d3c209b4d044c3263993b0970108d47ba5c34dd}}, {{cite:2acd6ee58e301444690765e6466761ca4802f8f8}}, {{cite:1a17137aa6972716f9b2a02ca26309442f620932}}. The resulting optimization problem often turns substantially more challenging than the simple empirical risk minimization (ERM).
In practice it is often necessary to schedule the penalty hyper-parameters in a manner that weakens them so much that they no longer enforce the intended constraints. As a result, when one initializes such a method with the correct solution, the training process deviates from the constraint set and finds inferior solutions (Figure REF ).
| i | 767294a0ab3264da9056d6d5cb1d7fc2 |
A smooth metric measure space is a triple {{formula:85044ad5-f7c0-4dd5-b969-a977f92e53cf}} , where {{formula:9831cba6-4d8b-490f-be61-ad69d95ecbe4}} is a metric, {{formula:790d18fd-aa0c-44e3-bda5-e22474da544b}} is the weighted volume measure on {{formula:74d05430-cb0a-4215-aafb-d9ae044d6265}} related to function {{formula:bd9171c0-e61b-47b3-ae3f-2579f30723dc}} and {{formula:bdc5145b-01a7-4a9d-bf8b-26ace7d3690c}} is the Riemannian volume measure. Such spaces have been used more widely in the work of mathematicians, for instance, Perelman used it in {{cite:19e6cbec4179f77ca6a5db132fda3f866c2cb184}}. Let {{formula:9303cdc3-005a-4f4e-af8a-0deeeb06db36}} be an {{formula:0276ad40-dec0-4f0a-9c1b-65e19051b9b8}} -dimensional closed Riemannian manifold with metric {{formula:c9e89f84-9f27-4527-b1d5-5ea4b659da0c}} .
| i | e23662e74a3d9bede2af322c2554fea5 |
Example 1 Given the California housing dataset (that includes ZIP codes, square footage, number of bedroom and bathrooms, sales prices, etc.) {{cite:783c464e53c504f84557ef8b6285946eb834125c}},
and 20 ML models (e.g., regression
trees) trained on distinct subsets of this dataset,
(1) predict sales prices for a new test set (that includes the previous features but not sales prices), (2) provide confidence intervals for each prediction, (3) establish the robustness of the predictions to distributional shift between training and testing data.
Although selecting the model with the lowest error in the training set is a natural solution to (1), the deterministic nature of this solution is in conflict with the requirements of (2) providing confidence intervals and (3) ensuring robustness to distributional shifts.
Similarly, although using training errors to create a weighted ranking of the 20 models could be used to
provide confidence intervals, it does not ensure robustness to distribution shifts.
With the proposed approach, we
estimate the error of one of these 20 models (selected at random by Player I) against an
empirical distribution (selected at random by player II). The randomization of the underlying game provides confidence intervals, and its adversarial nature ensures robustness to distribution shifts.
Figure REF illustrates DTB. For that figure,
we trained {{formula:4291d66d-e030-4390-afe5-70ab7dac2402}} different regression tree models (each with a maximum depth of 15) on a train set with {{formula:c9b4d7a0-278a-423f-bdb5-32a692860fba}} samples, then exposed those models
to {{formula:5c0a52d6-6eec-4cf7-87ee-a1274b80737b}} unseen randomized realizations of a UQ set (with {{formula:fc7b944d-d3c8-49e0-836d-351c532da78d}} samples) to find each model's
probability, and finally applied them to a left-out test set with 104 samples. We chose a small test set to accentuate uncertainties for ease of presentation. Each model is trained on a random half of the training set ({{formula:4226d91d-6686-4cde-8bd8-71a6261bcfb3}} ). The sorted unseen randomized realizations of the UQ set are divided into {{formula:659b8370-8094-41e4-8583-8e4d2e794712}} subsets. The combined support of the {{formula:1cd9f5ec-64b0-4f13-b0d4-c102552216d0}} in each randomization encompasses roughly {{formula:ed065df3-76e5-4731-819d-d1b2db6cb628}} of the whole UQ set. The prediction distribution {{formula:c0dd3126-0bce-4f52-8810-6bc79456f8e1}} over the 20 models is calculated based on material presented in Section REF . For each 104 sample in the left-out test set, Figure REF shows the mean prediction (obtained by averaging the prediction of each model with respect to the distribution {{formula:205348b1-b720-48c2-8b86-061737a9e4d8}} ) and the standard deviation of that prediction (with respect to the distribution {{formula:a4a62a32-4880-4dbc-b8ef-5170fa6a3750}} ).
As can
be seen from this figure, almost all ground truth values over the
test set fall in between the confidence bounds of the predictions (note that we used one standard deviation to define confidence intervals).
| m | cb8b18dc81409104b351ccafe558fe8c |
Audio Generation Module:
It consists of an audio generator {{formula:0c09848a-6d0d-4949-9c27-6b86c7c9b90a}} , that generates audio from noise {{formula:0a16d763-ca21-4322-a7d8-53d869130f8f}} and a discriminator {{cite:3da686593b20216d9b70c6f06bdd9bdb3b5ec0da}}. This generator module is parameterized through a function {{formula:981b5db8-c0b6-49d6-959a-8c90c2872d0b}} where {{formula:4324a853-433f-4e52-8270-ee6c5c46da03}} are the weights of the model. We work with the raw audio waveform, and the generator produces raw waveforms instead of spectrograms. The generator module uses 1-D deconvolutional layers. Since there are frequency bands present in each audio waveform, unlike the image signal where the frequency patterns are not so common, an image discriminator architecture doesn't work well for audio generation. The discriminator easily learns the frequency pattern commonly known as the checker board effect and discards those waveforms.
These artifact frequencies occur for a particular phase. Therefore, our discriminator needs to learn this trivial policy to discard such generated examples. To resolve this hurdle, we employ the skill of phase shuffle operation {{cite:3da686593b20216d9b70c6f06bdd9bdb3b5ec0da}} to prevent our audio discriminator from learning such artifacts. It is parameterized through a function {{formula:3e9fb1c5-4399-4a2f-a872-2a020048f951}} where {{formula:ddb2cc43-9a5e-43c0-a7e6-f99607172d2e}} are the weights of the model. Generator {{formula:b9edaa7b-1564-41f0-b68a-b7227da43df5}} is trained to minimize the following value function, while {{formula:06956e8a-7557-48c9-9bf1-4fb2601b26b4}} is trained to maximize it.
| m | 787b67d6ae5557851ab14bd360aa4379 |
One interesting phenomenon associated with the existence of an ALP is
formation of the axion cloud surrounding a black hole {{cite:a22f77a6d8f00f4eb8efc825a7306b5e7a2921ec}}.
For a massive bosonic field around a rotating black hole,
superradiant instability takes place{{cite:645b2d8df3883b9117077ccc3f93079f07a8d4fb}}, {{cite:245fc3a3aa64ebff5b68e4e382489153c4736f3f}}, {{cite:e832e6ea228eb88ce35794bf0489e5b8965ad8a7}},
and the specific modes of the massive field exponentially grow with time.
The growing time scale for the massive scalar field is analytically estimated in Refs. {{cite:645b2d8df3883b9117077ccc3f93079f07a8d4fb}}, {{cite:245fc3a3aa64ebff5b68e4e382489153c4736f3f}}
under certain approximations and numerically calculated in Ref. {{cite:573f3a30483d148845c373403a75208d974e2379}}.
Possible observable phenomena and those relevance have been discussed in a number of papers {{cite:41a876a2714b49828c245e520f0841ba23cfbcea}}, {{cite:493330b097f6bcb89903ec2983a3b0ec6639e50d}}, {{cite:31baaee9c53c66c9fb80eec2c919ed46bbe41894}}, {{cite:05fa7f441d98969c847a48d1b66983fd01799b8f}}, {{cite:157113d68ef892182710491d5c338b71aed450d8}}, {{cite:9d0be78a30fc29746cd6ef041b551f700a53a55f}}, {{cite:478699ca66e998727023e700a5c744528e689b1d}}, {{cite:c8a83f956856225c9e386fbf4f572a0586554f08}}, {{cite:f77ec05687d8351b4a056269987e8d76813f0905}}, {{cite:289769627dd0e85b689efd22bd83b9458098aba3}}, {{cite:4650f7b5e92f13439e7794379d2ebcfa09973f8c}}, {{cite:ebe1bae3699505209e56937e75b3d89432bdb958}}, {{cite:c36e9a45bcb4c8c7ba9b2e5b0e8681a18ca936ca}}.
If the Compton wavelength of an ALP is similar to the gravitational radius of a rotating black hole,
the angular momentum of the black hole may be efficiently extracted and exhausted by the superradiant instability
within the cosmic history.
Therefore the existence of a rotating black hole may give a constraint on the existence of an ALP
in the corresponding mass range {{cite:640e57431ade9cf362f42f03d692ba8902db6871}}.
However, once we consider more general situations,
there may be an efficient decaying process which spoils the persistent growth of the cloud.
For instance, the existence of a binary compact object
may dissolve the cloud due to the level transitions {{cite:222e929f863534d463ab66af60d3ea45fc50650d}}.
In this paper,
we focus on the dissipative effect through
the coupling with the electro-magnetic field.
| i | f8fdfb802811435d17b787b613ecbe16 |
Under the assumption of incremental network growth and bounded human effort, infrastructural and the socio-economic features are written in Ref. {{cite:9450e3ac3cdb3189b4ccc0280b2c2639e5277999}} as power laws of the population size respectively with exponents:
{{formula:19c428d8-9cc7-4e73-96ea-a7a7b0996d68}}
| d | dd89f88f2b418be516deac4dd892902b |
Answering these questions requires first establishing that a power law is indeed the best description of the data, as opposed to other descriptions, namely other functional relationships between {{formula:21163a09-8fab-4de2-a186-06d0854b60a2}} and {{formula:bcfedbf0-8d63-4107-b3b6-689a7da956d2}} . Assuming the power law to be indeed present, the second, related, problem is to estimate its parameters: the pre-factor, {{formula:7ffc5720-cde3-4f67-a5f8-2cf30eddce55}} , and often most importantly, the exponent {{formula:93a35f23-ce0a-4d51-9a9a-ac2e24403d1c}} . Solving both problems in a statistically sound way may be quite cumbersome. The presence of noise in the data and not having data spanning a sufficient range can lead to misleading results: identifying something that is not a power law as a power law, or finding incorrect values for the exponent or the pre-factor. These issues are not commonly addressed when discussing power-law relationships in neural data even though, as we hope becomes clear in this paper and as has already been noted in other areas of science, the implications could be far-reaching {{cite:7a5a2cf94210f1059bddc63a8e3d7a6cc0c5db26}}, {{cite:bc47eabec055e7333dd8a68944382c8429669dc1}}, {{cite:a41325ffc053d1781b26f8fb4dda7eccd82de140}}.
| i | 5430d67b7457099394da25877ce75356 |
In this section, a comparison of the proposed monolithic solver with alternative minimization solver {{cite:f29ca55b9532666cdbe007135bfccc4ddd7a8cdf}}, {{cite:6f47328ba17c42006f893a22b1f57fbe623892e6}}, {{cite:601a264bf9684b3bff3889a512a1ea12b9ac97e4}} and the quasi Newton-Raphson method proposed by {{cite:3d1080e26f53f16c4b92e8b4bee4df1a0fba2cfe}} is carried out. The comparison is based only on the load-displacement curves obtained from the respective solvers. The alternate minimization solver solves the displacement and the phase-field sub-problems alternatively until a certain tolerance criterion is met. The quasi Newton Raphson approach proposed in {{cite:3d1080e26f53f16c4b92e8b4bee4df1a0fba2cfe}} adopts a linear extrapolation of the phase-field variable for the momentum balance equation. Although, the extrapolation strategy is questionable, it yields a convex energy functional, resulting in a better convergence behaviour of the Newton Raphson method.
| m | ec62fceda92c2fc985a046f70be9da39 |
paragraph4
.5em plus1ex minus.2ex-.5emSmall normalization batch size is known to have negative effect,
as observed in previous works {{cite:86b41205e1432bd56ebbfecbf27d9acb94931be1}}, {{cite:3ff430ad04a76ec2bf374a80ef77d562a54be80f}}.
We analyze the underlying reasons,
and show that on ImageNet, the large performance gap between small and regular normalization batch size
can be reduced by 3.5% with PreciseBN (Table. REF ),
then by another 6.5% by using mini-batch statistics (Fig. REF ).
These improvements do not affect the SGD training at all.
| d | 870fa177962f6d62ab5bada7052082ea |
Attracted by these different properties several recent works, such as
Novelty Search with Local Competition {{cite:e94d4ac7ccc3fa02ae3eb59b82cd75cd68c1fceb}} and
the MAP-Elites algorithm {{cite:b90be5e7db05d52dc21bc508c3d6bcad0d294811}}, started to
investigate the question of generating large collections of both
diverse and high-performing solutions. Pugh et
al. {{cite:6c34a54f1168444947f5f3f2630ffd61c0470ef6}}, {{cite:155fa9c6782382bd869ff4fc865ae93ffb93d01b}} nicely named this
question as the Quality-Diversity (QD) challenge.
| i | a6806d5e05ea869f6a2d043b5a72b50a |
Fig. REF provides a more comprehensive comparison in the bit accuracy against various intensities of distortions.
In general, our combined model achieves better performance than HiDDeN {{cite:8b2978aa70b7569df2d3e1798655a931e5b3c9f5}} for resisting distortions.
Although our method fails in Dropout (p=0.3) according to Fig. REF , the curve of Dropout in Fig. REF shows our method surpasses HiDDeN {{cite:8b2978aa70b7569df2d3e1798655a931e5b3c9f5}} when the remaining ratio {{formula:1a70624f-9df5-491b-8a9a-29b17daf1bdd}} .
When comparing the identity model to combined model, we can see that the noise layer {{formula:7b4584e5-bba6-497c-b0b5-fbd59bebee19}} has obvious benefits for Gaussian and JPEG compression distortions, as the identity model generally fails under such two attacks.
| r | 9203e7d604eee3d581e32552e909312c |
Graph cuts for {{formula:5d3e8182-a75a-4ada-99bf-121051ee6cf3}} {{cite:d89333224dc0d7565f40f81ac385a52d2d234ee1}}, {{cite:400f6315032a8a12de88b8d1a7b8c7f3e8cf6637}};
Random walker for {{formula:74c19266-d9af-4673-af66-88d67ce74e89}} {{cite:f6d55672cbfd9409e1fe9dd2f3a53fe033a35048}};
Shortest paths for {{formula:b4bcba54-e13e-40bc-b2f6-dea19da60410}} {{cite:d1a693f8bd0e43762d6ed2b5d06f6ebcae187eb1}}, {{cite:4b6fce3bc6ea7234cc359ef2a4c8eebf466909b0}}.
| m | aa38c03bb5301dc3ad05f2ad8fb38ac2 |
All experiment settings are kept consistent with {{cite:b5ae2981ed15142a2a46ab5387e6f090b23db3ab}} ({{cite:11cb2df4f7ce3f52dd71adc8fd6ee4244567f31b}} on ClueWeb12).
All the candidate documents to rerank are from their base retrieval methods. Conv-KNRM and BERT Ranker use open source implementations {{cite:ed006a971c184519a4dbfd1eaeaed3ffbdcec5a1}}, {{cite:f9b6048fc5b85bbf3bd1898155053359d17d0ffe}}.
The evaluation scores with the corresponding paper are thus directly comparable. This is crucial for reproducible IR and to compare ReInfoSelect with the supervision from private Bing search log {{cite:b5ae2981ed15142a2a46ab5387e6f090b23db3ab}}.
| m | e82769475da9dd49940ad28d63ee260c |
Note that feature selection is a research topic highly active in the Machine Learning community, especially on non-physics-related topics and classification tasks, with plenty of articles dedicated to the question (to name a few, {{cite:8e565d0597b88b67aac69e5812aa0a8dfff70864}}, {{cite:1ef2b509a1c96ec3b9a766ff988f1da3ba8b6022}}, {{cite:5808c9318f8430f7798a859524ca415181f51dad}}, {{cite:794bc10e3f32e89feddf78d5a7e46ca5b05a503a}}, {{cite:aaf36de13cca6303945ff43fc0e3a9dbc369cefb}}, {{cite:fbace08df3f4f7a24eea36ffbf5b35a99c2d950c}}).
Researchers have gathered existing methods into five main categories: filter methods, wrapper methods, embedded methods, ensemble methods, and integrative methods (see, for instance, {{cite:f6fecf7064b3adeb87060d56b55d21a49349eae1}}, {{cite:fbace08df3f4f7a24eea36ffbf5b35a99c2d950c}}, {{cite:52be3efed8652c389467dba95ddc9f58cffe02a2}} for details). For this work, we are particularly interested in finding methods to identify a priori (before any training) some relevant features just by analyzing existing data. The techniques developed in this article can be viewed as assistive tools for classical or data-driven modeling. It falls into the filter methods category, where numerous approaches have been considered based on linear correlation, Fisher score, etc. (see, for instance, {{cite:f6fecf7064b3adeb87060d56b55d21a49349eae1}}).
| i | bccc2bea8535d2dfb7733e457e22deb7 |
Detailed models of pedestrian motion inevitably require to take unobservable internal states of humans into account, e.g., for decision
processes and interactions between pedestrians.
Our findings suggest some principles in evaluating the current and future states around each individual.
Pedestrians try to avoid interference in their
personal space, considering the cost that depends on physical
constraints (distance, angle, etc.). Such future anticipation cannot
be captured by one of the current paradigms of
pedestrian motion that is described by a driving force to a predefined
target and local avoidance {{cite:bef65fb3a612beb0e877057d7f3b94df8bee14c3}}, {{cite:a85dc34b66303f9a5be6cb2324ad343d6eb6f322}}, {{cite:14d2b2c14856dace94fe2e36f8271bdb78528447}}, but requires models that include
decision processes, like most cellular automata {{cite:43df244e70e92fdd75a3de5e90a2e076c93d1022}}. For this reason, the inflow process has not received much attention in
pedestrian dynamics research until now.
| d | b763c0549d07eb9500905316444f8070 |
In order to preserve meaningful values at the boundaries, the convolutional filters used in the proposed architecture incorporate a symmetric boundary condition into the padding operation. Classically, padding is used to preserve the spatial dimensions of the field being convoluted, but the standard zero-padding approach doesn’t usually represent the expected physical behavior. Indeed, padding with zeros everywhere would violate the representation of existing boundary conditions, for example, the notion of wall boundaries would have lesser significance if a region is padded with zeros on all the sides in a channel flow. To preserve the boundary conditions after multiple successive convolutions, a boundary condition formulation was implemented such that the walls could be padded with zeros if required, while the other sides could be padded with adequate values from the symmetric cells. The resulting model is trained with the Adam {{cite:0192e74f47639e68e23182e1c8eb0a683a73aafe}} optimizer, to iteratively minimize the total equi-weighted mean squared error (MSE) loss defined by:
{{formula:201057c2-9bf6-4d84-a8e7-f7a456d16d5c}}
| m | 1795223b19a4497984ea5707b4e71e70 |
Empirical Risk Minimization (ERM {{cite:2948d51ac4d6744f4353c9286e57f197fdfc4a4b}}, i.e., CNN baseline): minimizes the sum of errors over data. We regard it as the naive baseline to learn a single model on all source domains.
Meta-Learning Domain Generalization (MLDG {{cite:771040b76be02752c42e8df8fa53f0d4ecf62685}}):
learns how to generalize cross domains by leveraging Model-Agnostic Meta Learning (MAML {{cite:003506944d0c91bf01621c75b5222ae244755ea0}}).
Domain Adversarial Neural Network (DANN {{cite:7d7e212402947fd09801b3b088417feb5a57664c}}): employs an adversarial network that consists of a generator and a discriminator to adapt feature distribution. Under the domain generalization setting, we perform DANN across multiple source domains as no target can be accessed during the training process.
Group Distributionally Robust Optimization (GroupDRO {{cite:48f81b75d58601b853597f6eb306b18605db6cd3}}): couples group DRO models with increased regularization to increase the importance of the worst-group loss.
Representation Self-Challenging (RSC {{cite:6d2d781d51e9cd18b4340cc0902f875ac5aae1c2}}): discards the dominant features i.e. representations associated with the higher gradients at each epoch, and forces the model to predict with the remaining information.
AND-mask {{cite:ff2cdacf70a1a2d384adb334e5114f1f1d523c19}}: learning explanations that are hard to vary, which uses AND-mask to improve the consistency in gradients for better generalization.
| m | a4ea29ed05e08bc905029050a5acfd9f |
Although, to stabilize this approximation, approaches were introduced in {{cite:ffe1643c82f76d9fc76076fb88c55e196a64164a}}, {{cite:b1200049fa699207babdb4fed2855777cb9938f9}}, they rely on the learning process of a single network that can still be biased by its own initialization. These constraints can be removed with the use of a second NN. If we run the same approximation procedure with two independent initialization, the solution is more robust as the optimal value is obtained only when the two networks converge to each other which is in contrast with {{cite:ffe1643c82f76d9fc76076fb88c55e196a64164a}}, {{cite:b1200049fa699207babdb4fed2855777cb9938f9}}. This type of approach allows for the steady improvement in cumulative rewards we observed from Table. REF results. Another crucial component of our methodology is the use of buffer and the use of the second network paired with the exploration strategy thus reducing the need for a large replay buffer which we demonstrated in Fig. REF .
| d | 7721ec858473c3ff63d2aa32a84aef43 |
Based upon the current experiments in atom-cavity QEDs, we theoretically investigate the strong single-PB to two-photon bundles emission in a trapped single spin-1 atom coupled to a single-mode optical cavity. We show that the energy-spectrum anharmonicity associated with {{formula:f6e29d53-3eaf-4602-b691-ceb398a600a9}} th dressed-state splittings can be significantly enhanced by tuning the
quadratic Zeeman shift in spin-1 JCM. The high-quality single-photon emission with {{formula:7352f5d3-f2e9-4d23-a2da-824387c16f16}} and two-photon bundles emission with {{formula:4f6408eb-0b0d-4bf3-8ff4-66532f36846f}} are achieved by driving the cavity and atom at a moderate atom-cavity coupling, respectively. Compared with the seminal spin-{{formula:995602be-8c3d-4795-be36-0cf281257722}} JCM, {{formula:9eba15e5-abfa-483b-a720-50d06b35a7c9}} in our scheme can be reduced by three orders of magnitude with a large cavity photon number. The strong optical nonlinearities are generated by the advantage of versatile spin degrees of freedom in high-spin systems, which helps to weaken the strong single atom-cavity coupling regime required by the seminal spin-{{formula:8e7664db-b130-48c3-b761-bcda9e112235}} JCM. In particular, the photon emissions can be highly tuned from the strong single-PB to two-photon bundles with interplay of cavity and atom driven fields, corresponding to the transition of strong antibunching for isolated photons to separated two-photon bundles. Furthermore, the scheme for realizing {{formula:c89643a0-7977-49b3-b35d-8332d44ebf06}} -photon emissions with large steady-state photon numbers beyond the strong-coupling limit could be used as high-quality multiphoton sources. Our study offers an exciting opportunity for studying novel nonclassical quantum states by utilizing spin degrees of freedom in high-spin ({{formula:a74cbadd-0994-400a-8972-520b460f32f2}} ) single atom-cavity QEDs, e.g., {{formula:dd61e6b1-bd99-4127-acf1-912c48aabbce}} Yb with {{formula:8a6595fa-cf87-4671-9292-ad15cf8fc216}} {{cite:d7e2c4d8df3ff741001bd062c29d4758ec11de7e}}, which could provide a new building-block for exploring fundamental physics and applications in quantum computation {{cite:1ab3346f01509c27a1b8f13bf7a3561e0f90c014}}, {{cite:b68056fa159e0d8be8964c1e7c8caa66925c38a6}}. In the same spirit of our approach, we point out that our scheme can be further extended to explore new nonequilibrium phases and strongly correlated many-body physics in {{formula:1e9fb1ea-3242-4326-b4dc-4a02feb3285d}} -level Dicke model between atomic ensembles and optical cavity {{cite:95c600d3f8bcb56eeaba9d4dc1985aaee3151eee}}, {{cite:f43773237a5379d050c793bde97d1c1dea23e6fb}}, {{cite:9f7104d52657558e2696bb717bead96ec1db8303}}, {{cite:64419dcff3b2c52d194e5679340d705d8eb45c95}}, {{cite:96e49c0d944e8e99033dcdff3230923b93dac134}}, {{cite:c190f4da7c118c4fca4176fc102118d3ab4d5091}}, {{cite:13204f9ed686cb92c00ecaa43a68360fc4d104dd}}.
| d | 0d348a66bb9c1cb892f2967a4d19c7ed |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.