text stringlengths 54 548k | label stringclasses 4 values | id_ stringlengths 32 32 |
|---|---|---|
Multirotor vehicles, also referred to as multirotors, are becoming widely used in real-world robotics applications for their simple mechnical structure, agility, and low cost.
As multirotors are brought to new application domains, there is a rising demand to further extend their maneuverability {{cite:70596cfa2648f257d68cf7446f2734cc00086b18}}, {{cite:90228cc8e777cda85d37370b272e13340cb811ca}}, {{cite:62deca0348be7dacb1e97b47b5cfbde8a63e20ce}}, {{cite:3a52b03e8c874af2608dc0fc66a581508df6ffa4}}. To fulfill this need, research has been conducted to investigate fully-actuated multirotors using fixed-tilt {{cite:d3547c4a58d60923a6ea0cbd0699cd5a11e72bc9}}, {{cite:8c21e6402d829a97b93768032a1713b5406438c0}} or variable-tilt {{cite:e89523fc9f1b27d5b6f45fabe29a9219fcc5b079}}, {{cite:629b4072c1efe0d7a6e36047e8949ba8ad6e84b4}} rotor systems that enable the vehicle to carry out translational motions without altering the attitude.
| i | 75be5404b38e6d63adee68ade4ffc892 |
We show below that the brain neurons of crayfish act in critical states, generating neuronal avalanches on their dendritic processes and are self-organized. We compare the scaling relation with the BTW sandpile {{cite:c7f7c90f7c2b1ff07731fb15864d029b82af7c7a}}, {{cite:90e475d0e326c3c0cc021f840016b68e278ed930}}, the stochastic random walk model proposed for RP {{cite:5d2f4a91d1aee71813050df41d6708ecbfb9744a}}, and vertebrate species {{cite:39122979cc190f150f555a22924062d1a01c752d}}, {{cite:a32d425ca955ccfdf6e8da47011d2f46d8c14efe}}.
| i | 051e9e0b27fef5e14590051a896486de |
Convex generalized linear losses. Our first case of the study is non-smooth DP-SCO in the case of generalized linear losses (GLL). This model encompasses a broad class of problems, particularly those which arise in supervised learning, making it a very important particular case. Here, our contributions are two-fold. First, in the {{formula:8c6204f9-95fc-4024-8b5f-b596ac02a87d}} -setting, we provide the first nearly linear-time algorithm that attains the optimal excess risk. The fastest existing methods with similar risk work for general convex losses, but they run in superlinear time w.r.t. sample size {{cite:cd19ccaaa6b515b32f0d3a65e1d413097345ccf8}}, {{cite:27cba43539e0cd697535908b2dfb49d3f2658c70}}. Our second contribution here is a nearly-dimension independent excess risk bound in the {{formula:ce6c1b7f-451f-439d-894c-e87fa07bcc84}} -settingAs in all existing works on DP-SO, in the {{formula:fe58628a-3beb-486f-a91b-96448bd6273d}} -setting we also assume the feasible set to be polyhedral. for convex non-smooth GLL. This result circumvents a general DP-SCO excess risk lower bound in the non-smooth {{formula:633d10dc-992f-4ce8-b031-ba190ba53788}} -setting which shows polynomial dependence on the dimension {{cite:cd19ccaaa6b515b32f0d3a65e1d413097345ccf8}}, and it matches the minimax risk in the non-private case when {{formula:dcc8a936-d68f-42c3-9a8c-877e74472fcc}} {{cite:2c527199910fa771dcdd871ae2d2f32e3fbd0461}}.
| r | 8cd4ebc6b672408a41a19c5f551d197d |
which is often called system-matrix in tomography.
Here {{formula:9967adcb-5eef-455c-895b-001c6011128c}} has {{formula:de10ff0d-f2a1-4039-92d3-e89ef5a32fd5}} rows for the volume-cells, and {{formula:44b9e002-c0fc-4ae9-a342-e6617a93e405}} columns for the rays.
In our representation {{formula:af0f2250-5ee6-4289-82da-901b72b9a9ff}} contains the Cartesian-distance a ray travels through a given volume-cell, with the edge of the volume-cell having length 1.
From here on, we can follow different methods to reconstruct the volume-cell's intensities {{formula:5fc3d91b-55ee-4241-9422-0a87f524b165}} using the system-matrix {{formula:531d5699-d88e-41ff-9df4-2644bf383818}} , and the recoded light-field {{formula:c2cc1953-0e47-49e9-856c-c6735ce65c1c}} .
Here {{formula:86112eb6-9c5f-494a-91c1-526db1e093bd}} is a one-dimensional vector containing the reconstructed photon-emission-intensity in each of the {{formula:b4ab723f-c82e-4b8e-bbf1-3328f48219ab}} volume-cells, and {{formula:5fa1002e-efb8-4df7-bb62-828b180ed668}} is the a one-dimensional vector containing the measured intensities in each of the {{formula:64ae5817-37a8-4152-b257-24f934456499}} light-field-cells.
For the reconstruction of {{formula:4dd28d32-b53c-451c-8e91-43bbe8594a8f}} , we first explored filtered-back-projection, and then went on to iterative methods such as the (simultaneous)-algebraic-reconstruction-technique {{cite:d9de066ec42e495b8745803ff2459ba703ecdbff}}.
Figure REF shows reconstructions using our first implementation of filtered-back-projection.
| m | f3fdcd522e2609395fdbfbc542c238e5 |
Although the random shuffle operation performs well on the {{formula:c5551ff7-11df-4eda-91c1-081a451cc545}} and LiH molecules, the performance can be further improved by developing more advanced shuffling strategies. First, instead of random shuffle, we can design a problem-specific and hardware-oriented Hamiltonian allocation tactic, which can eliminate the deviation of the optimization path of local models and better adapt to the limited quantum resources of various local processors. Second, due to the existence of barren plateau {{cite:9c8a846d27c60272d4f540d8f9bd47650f094c9f}}, {{cite:5dd0025f3315b191ce97dcbf14d53978b9c36a5a}}, {{cite:efa6f0bb38f458e4925848b1c95a3b4d459e79d2}}, {{cite:d0dfc82bdef6fd031a4a4cc7691d31a41eb34829}}, {{cite:15f69e530e2d383d2ef4f90fbd9fb52a1912c602}} in the optimization of ansatz, the training of local quantum models may get stuck. Inspired by the study that local observables enjoy a polynomially vanishing gradient {{cite:d0dfc82bdef6fd031a4a4cc7691d31a41eb34829}}, a promising direction is to group Hamiltonian terms with similar locality in QUDIO to avoid the barren plateau of some processors. Finally, a more fine-grained partition of the quantum circuit structure besides observables can be employed to reduce the number of parameters to be optimized for each local processor, as implemented in {{cite:61b70598afafa03b95bd92f43c3c416a58d838ce}}.
| d | 27b287f9a5094163b8b91d1e151f1205 |
The internal correlation
of {{formula:1e0c936c-60a0-43c4-9e0e-708c598d76be}} has been studied in {{cite:272ee91a35e24a637833ac8b0362839f9e6fbf52}}, who
found that the proto-halo of a {{formula:d64bb11c-9e62-40ab-9951-aad30d712332}} halo with high {{formula:c5ecd576-5c83-436f-9e5f-ece7def7537b}}
is usually located in the vicinity of massive structures.
They thus suggested that the large-scale tidal field
associated with (neighbour halos or/and large scale structures) the high density can truncate the
accretion of the material in the outskirt of the proto-halo{{cite:94945dffb040ecdd3233c317b85f563ea6c80c7b}}.
The stronger the tidal field is, the larger the truncation effect
and the higher the assembly redshift. This can explain
the {{formula:66e6b02b-48c3-4f57-a5d2-56e47e55ff8d}} -{{formula:86f231bf-b998-4571-bce8-45a9606306e3}} correlation.
However, they suggested that this effect is only important
for small halos {{cite:09d6faf1baaa4b7b911479ce004b71327ac831ae}}, {{cite:6a9f7acc728e4d26c0b9013fbd38b60da8ef97a5}},
which is not consistent with our results that this effect is equally
important for cluster-sized halos. This indicates that the
tidal truncation may also be important for massive halos.
| d | 6bb05d77040d666cab8bf1ee09ba09e9 |
Our main idea is to include each edge of {{formula:490ce72b-8818-40b5-935a-aed24ab8c332}} in the sparsifier {{formula:d10e95cb-0c4e-4269-ad5a-3e4040d0e9e3}} with probability
proportional to its effective resistance. The effective resistance of an edge is known
to be equal to the probability that the edge
appears in a random spanning tree of {{formula:e21dfa90-d3ee-46ce-bbd0-805db6b11b9e}} (see, e.g., {{cite:9b534f10307bd84971e87af217210b027ceff202}} or
{{cite:d4c47e90a3f7de2a180b7fc2bae7ccd9e2c20645}}), and was proven in {{cite:351e4a2c2b52bedeb347cc6ca0f7c46e76d773a9}} to be
proportional to the commute time between the endpoints of the edge.
We show how to approximate the effective resistances of edges in {{formula:623bffa4-18ce-41a9-93d3-fdf8a03d375a}}
quickly and prove that sampling according to these approximate values yields a good sparsifier.
| r | 86c6882a0001b2aa62e4d34b96e525b2 |
The main idea of word embeddings is that their representation is obtained according to the context (the words around it).
The words are projected on a continuous space and those with similar context should be close in this multi-dimensional space.
A similarity between two word vectors can be measured by cosine similarity.
So using word-embeddings for plagiarism detection is appealing since they can be used to calculate similarity between sentences in the same or in two different languages (they capture intrinsically synonymy and morphological closeness).
We use the MultiVec {{cite:91261098fa0442771d9af89c9aeadc797b03390f}} toolkit for computing and managing the continuous representations of the texts.
It includes word2vec {{cite:c5b10bf4a700ba069fef0b78b9f95c558235733f}}, paragraph vector {{cite:60c8b43fa3290ef91efd27d9ff2dca05784344bf}} and bilingual distributed representations {{cite:8711299f2cc329c61727155fa09f29bf2525442a}} features.
The corpus used to build the vectors is the News Commentaryhttp://www.statmt.org/wmt14/translation-task.html parallel corpus.
For training our embeddings, we use CBOW model with a vector size of 100, a window size of 5, a negative sampling parameter of 5, and an alpha of 0.02.
| m | 0d26ce98fe5465e7ff164a627dd2b7f4 |
In the last few decades, the limits in accelerating gradients of conventional electron accelerators based on metallic cavities prompted considerable efforts in the development of alternative electron acceleration techniques. Hitherto, the acceleration of electrons in the wake of an intense laser pulse propagating in an underdense plasma (Laser Wakefield Acceleration, or LWFA {{cite:6dd43be5e9dd8c2ca3e21b605b16a1120dcbcd9d}}, {{cite:89638d3691c1460b280cf0902bb3d75f5dde4f59}}, {{cite:97aaf83b31b88443be7ad50309467a97680d02a9}}, {{cite:c26d65bda36bce90a2756647297e694721f52fb5}}) has been proven particularly promising, generating smaller electron accelerators with high accelerating gradients {{cite:2cf4e12cb42db72512048bb2b4bcc8692b91f00c}}, {{cite:92d369fc6cb91ba998b036ef53dfd2bb45e647b0}}, {{cite:d4a673a988b25bd629c0fdbdf8e23a2a9b0b34bd}}, GeV level final energies {{cite:ccf3998283a4660b08c142129227b4d6764a6b9f}}, {{cite:2bfc2b20a568e2a64bbc436823cce9c41873989c}} and femtoseconds duration accelerated beams {{cite:ccdfd8c3ea6b5919b204b539174dc920870e8e83}}. An important role in this acceleration scheme is played by the technique used to inject relativistic electrons in the accelerating phase of the involved plasma waves {{cite:97aaf83b31b88443be7ad50309467a97680d02a9}}, {{cite:c26d65bda36bce90a2756647297e694721f52fb5}}. Among the numerous demonstrated injection techniques, one of particular simplicity and often chosen is called ionization injection {{cite:6ed3edbd3002e88f120d1eeaea58648a61075ed1}}, {{cite:92733225acb0747104c527e679ea71a63d877a3a}}, {{cite:6ef9ff5c8a2bfba1b53b6558f5b0d394dd0647fa}}, {{cite:b305d830fb4ba8a5ebe7b947bbbf510f1c15301e}}, {{cite:370174b43331860fef5f2fba977d86bede378d13}}, {{cite:a3ea4206429e29236d617198dd7a6d0e225fcd9e}}, {{cite:571a714aeccd2eb595b72d2f7728159600de4fcb}}, {{cite:0e56380eddd80eab2f0a21b2d26ebcba29bc4f59}}. It consists in using a gas mixture to generate the required plasma. This mixture is mainly composed by a low atomic number Z gas, like hydrogen or helium, that is ionized very early in the laser-gas interaction. The other part of the mixture is composed by a higher Z dopant gas, like nitrogen, whose first levels of ionization are reached early in the laser-gas interaction. However, the higher ionization levels of this gas can be accessed only at higher values of the laser transverse electric field, normally near the peak of the envelope of the laser, i.e. later than the first levels of the high Z gas and the ones of the low Z gas. Tailoring properly the laser and plasma parameters, these ionization levels are reached only during the short period when the laser is near its maximum focusing. It provides a reserve of electrons that can be stripped off from the high Z ions near the peak of the laser pulse “just in time” to be trapped in the plasma wave in the wake of the laser (this case is referred to as self-truncated ionization injection {{cite:fdefb57a7dd6e83e8b01d6215d6ee302a922d2cd}}, {{cite:a3ea4206429e29236d617198dd7a6d0e225fcd9e}}, {{cite:571a714aeccd2eb595b72d2f7728159600de4fcb}}, {{cite:0e56380eddd80eab2f0a21b2d26ebcba29bc4f59}}).
| i | 2b43a89daaa418c15ae482eac68fac76 |
Word-level Auto Prompting: Prompt search {{cite:6c490366c72ece4863ca63a6a614ea0e7224349d}}, {{cite:14127e6ec80f9b58e407b7e97a1bb25291da15e8}} is one of the most effective means of prompting large language models. To use word-level search for inducing personality in language models, we seek the most functional three words for each big5 factor from candidates in {{cite:dd73b39f5c252b2c3ed151b3993dc85be5164daa}}. For faster search, we use GPT-Neo-2.7B and the shorter 15-item version of mpi for evaluation and apply the searched words to the final prompt for control.
| m | c7316413d7cf2df93b114c53d7aef5a1 |
Most recent SSL methods, , MoCo {{cite:2ede64ec01c5cca205209a73030374d56fdfc38f}} and BYOL {{cite:8c2c6571b9c932db0916e4bd52307185cd2f8141}}, encourage a query image to be closer to its own augmentation compared to some other random images. Follow-up works have focused on improving the positive pairs through generating better augmentations {{cite:d8106e01ee00e0a54ca5408a4489edfcce77ab42}}, {{cite:8e6ff5ff55c778a0f19d83556de67c055af5cf56}}, {{cite:d48ff77ae13226a8c9ed304bbb38e0c7ce97563b}} and the negative set by increasing the set size {{cite:2ede64ec01c5cca205209a73030374d56fdfc38f}} or mining effective samples {{cite:e249d40081a2abbf0576369b163c370b2630d58b}}, {{cite:f2499cae470befab2d82f7949dd6c027e7f8cb50}}, {{cite:79399af0b8b9c88600cf31e063e828ffdecb04c0}}, but have largely ignored possibility of utilizing additional positive images. More recently, {{cite:6d4a008c1be5e2dc8f18e5d941a013fd019258bc}}, {{cite:fe4e7750bf42200e3704d02848156d274b77e6e3}}, {{cite:2197ef8fcc932b2a9aa8c7b7ff6469e8a53a27c5}} expand the positive set using nearest neighbors.
Inspired by classic mean-shift algorithm, MSF {{cite:6d4a008c1be5e2dc8f18e5d941a013fd019258bc}}, generalizes BYOL to group similar images together. MSF pulls a query image to be close to not only its augmentation, but also the top-{{formula:245bedf8-0903-42aa-8885-34409db12404}} nearest neighbors (NNs) of its augmentation.
| i | 3fd2f0e8ae5b8cb5c35120a82d157fe5 |
Using fluctuation theorems, we provided a quantum algorithm to prepare a purification of the thermal state of a quantum system {{formula:c989e3a7-2817-404c-bd9f-89e07233aa95}} at inverse temperature {{formula:1f328c90-b719-419c-9040-6973e83b3b66}} , starting from a purification of the thermal state of a quantum system {{formula:0e93580c-1d31-40de-9b93-a4052f19a32b}} . The dominant factor in the complexity of our quantum algorithm is {{formula:05268ce4-4065-4056-8ce9-fe7d238f0b55}} , where {{formula:ebc7da50-acdb-4c82-ac67-c83ac8cc408c}} is the free-energy difference and {{formula:94d5268b-cf3f-41fa-9a1f-53f58d7b07bd}} is a work cutoff.
This result is a significant improvement of prior quantum algorithms {{cite:3fbe29deb72ea0955050d27e9120e461797c88dd}}, {{cite:97cc48c739d27a2c7233ccdb380a9f0101b0f5ad}}. The work cutoff also depends on a non-equilibrium unitary {{formula:f1becc98-d122-4b5f-b7e9-b09ee59183c8}} that can be arbitrary.
We showed that certain {{formula:cd3c7cd8-1b52-48bf-a676-8234e0338175}} 's allow for an improvement in {{formula:e6c115e7-f501-4c24-a48e-f536d9717aba}} and the runtime of our algorithm.
We also obtained suitable choices of {{formula:4ce1cdf4-43df-407b-9b07-524b86207231}} for large classes of Hamiltonians, including general Hamiltonians, commuting Hamiltonians, and the local spin Hamiltonians. In all these cases we found a lower bound of {{formula:326f0d57-6017-4334-a4b7-01026b01721c}} that, for {{formula:b620d9fc-0528-4814-95de-1c372876eafa}} , depends on the strength of the perturbation {{formula:ab3b5284-c1cf-48e0-8045-26aeaf697278}} .
These results are especially useful in the regime where {{formula:46a4a6ac-04f3-4c26-9f27-46235721cb66}} .
| d | d8be622e10b3d163cdbd3240259da1c5 |
We propose UniMix and Bayias to improve model calibration by tackling the bias issues caused by imbalanced distributed train and test set. We improve the model calibration and accuracy simultaneously in an end-to-end manner. However, there are some methods to ameliorate calibration in post-hoc (e.g., temperature scaling {{cite:8a7e1e75fc7074df1dc1a8a3d922284d06fb6ce1}}) or two-stage (e.g., label aware smoothing {{cite:7bf9788569a752e63b391593c25b54eae284f1c0}}) way. As we discussed above, such pipelines will be effective and we integrate our proposed methods with one of them for further comparison. The compared results are illustrated in Tab.REF .
{{table:30d0e64b-d7e9-44a2-8e5f-7db61204c38b}} | m | 2b15cc2e5017433e08eb261f45b90a0f |
Research on HPO of GNNs for molecular property prediction is still in its infancy. For example, the pioneering work of GNN presented in {{cite:284e2d86bd45935b56c4a900c1b950c17eeab538}}, {{cite:5b7b60c3d2948a664b74381b82855dddb8aeb2a8}}, {{cite:e176543f63464aef020e0c31157faf77deaeb091}} did not discuss the problem of HPO in detail. Meanwhile, most HPO methods have not been explored in terms of their efficiency on GNNs when facing this type of problems. Their performance may need to be further investigated because the sizes of molecular datasets vary from hundreds to thousands, which are far less than those of the datasets used for typical deep learning applications (e.g., image recognition, natural language processing). At the same time, predicting molecular properties requires more sophisticated neural architectures to process irregular molecular structures, which is different from image processing problems that have regular spatial patterns within image patches. Therefore, it has become necessary to explore the performance of existing HPO methods on GNNs in molecular domains. This motivates our research, and we conducted methodology comparison and experimental analysis for RS, TPE, and CMA-ES to assess their effects on GNNs as HPO methods. We expect that our research can inform researchers in both molecular machine learning and chemistry and material sciences.
| i | 80f530abad7a2fee4bd608e613e6ed15 |
We presented a comprehensive empirical evaluation of our proposal on the well-established MNIST benchmark dataset {{cite:483408cf9f847bb834540cf8c747a54dd5dbed1c}}. Results across ten experiment iterations found that our conformal loss function is competitive with ACP for approximate validity and predictive efficiency. The crucial difference is that our direct approach needs only one model and therefore has significantly improved computational efficiency, solving a long-standing challenge with ensemble CP variants. The computational benefit increases proportionally as the number of ensemble models {{formula:ed10c9f7-e50e-4b4d-a916-8cf417d46ab5}} in ACP grows.
| d | 363ef7238112d76902df81c2498543a2 |
LIME can be described as follows {{cite:837f00ab10c4d465bcc8a8f96f46b5503507ae5a}}.
Let {{formula:e28f9abf-43d3-4469-9d1d-0347c81a8ba6}} be the function learned by a classification or regression model over training samples. No further information about this function {{formula:c8b1fc98-525f-4f3c-af9d-6498dacf17ca}} is assumed. Now, let {{formula:7eb1f0ef-a2ab-4760-a267-1a9dcdcf4f3c}} be an instance, and consider its prediction {{formula:633fd46f-dd68-445b-b87e-6ce13856afb6}} . LIME aims to explain the prediction {{formula:95ccf923-7957-4038-843a-bcfbe8dbdeb9}} locally.
Note that the feature space of LIME need not be the same as the input space of {{formula:fa28f9dd-43c6-490d-a395-b8c6d5baa6a1}} . For example, in case of text data interpretable space is used as vectors representing presence/absence of words, whereas the original space might be the word embeddings or word2vec representations. Indeed, LIME uses discretized features of smaller dimension {{formula:37b81e91-f7a7-48cb-9367-4a3b31d06beb}} to build the local model, and aims to learn an explanatory model {{formula:c97e6629-3533-45fd-ae67-17c276659f13}} , which approximates {{formula:76641e8b-7fe0-41a3-9068-28a90997f7c6}} in the neighborhood of {{formula:aec2fa90-3742-4a22-a67d-c8b98d74022a}} .
To get a local explanation, LIME generates neighbourhood points around an instance {{formula:8c15b6ad-435b-4036-892a-56994d4e422f}} to be explained and assigns a weight vector to these points. The weight is assigned using {{formula:a9edd2e9-e1c5-41ce-89f7-c87ec303f77d}} , which denotes the proximity measure of {{formula:95cff812-3931-4843-82a6-7c0519ef7715}} w.r.t. {{formula:077e78f3-369e-406d-9157-f36a34922d93}} . It then learns the weighted linear surrogate model {{formula:1619243c-f487-4463-b730-67519dce4199}} by solving the following optimisation problem:
{{formula:ffaf00a0-5784-4dde-8d85-db2881a48c02}}
| m | caea2602ee17a39dad964c7370a250cb |
The environments tested are simple for the purpose of establishing proof of concept in using a global DNC model in the continual RL setting. The environments represent multiple tasks requiring an RL agent and predictive model to learn continually. In the directional path navigation environment, the numbered tiles change for the same navigation task as the agent solves each level. In the obstacle-based grid navigation environment, new obstacles are presented, requiring the agent to adapt its navigation strategy. Other work on continual learning in the RL context examines environments where the tasks involve locomotion of agents represented as half-cheetahs, ants, or spiders {{cite:033f503b49d7f83e73b356214c5d02e50622a3c9}}, {{cite:685c26a5f812bd33d49c7750434b68c0e7b0b674}}, {{cite:436b6eb86ba002ea719c245b75f526cce92acaf7}}. The tasks are adapted in a number of ways such as varying the nonstationary locomotion environment the agent is in (the background, the slope of the terrain, and so on), or crippling the agent by disabling a joint or a leg. The agent's continual learning capabilities are tested as the tasks are varied. An extension of our work would be to test the NCA on such environments.
| d | b737b3bde89a4b70c38442b37109e36f |
By following the mathematical correspondence between Bayesian inference and statistical mechanics {{cite:250118797c19277bd13a8382df75f6830939dde7}}, {{cite:e9013dffaf47f36cea53676ecddd3a790b5af4b5}}, {{cite:4262a7e1ee626a3d78b746e9e5fb3c4f07f25ab1}}, the form of {{formula:0d0671e1-9c66-44fa-8a41-686a39e283d6}} is a "Boltzmann distribution" whose "inverse temperature" and "energy" for a "state" {{formula:30700d3b-8adb-430c-9799-0167830300b2}} are respectively {{formula:e5deb80e-2a32-4468-8018-9457b80f610a}} and {{formula:019683f4-6cd4-4590-812f-7d673229a6d7}} as an analogy. Whereas WLS is to obtain a "ground state" as a best-fit parameter set, Bayesian inference is to obtain a whole "statistical ensemble" subject to {{formula:244a726a-e75b-4bb6-acf2-9f3916e6b001}} (or to obtain {{formula:91a425c3-a747-4830-b1bb-556613c1db85}} itself).
A "statistical ensemble", namely numerous realizations of {{formula:af947744-7da9-4514-89db-92c992c32c56}} , subject to {{formula:4ee558c4-c96b-46a7-a4ea-62944cf95dfd}} is shown in Fig. REF (b).
One can see the posterior probability density {{formula:2e70715a-d63f-4094-87f4-0cd89daff113}} , represented by the colour gradation in the inset, tends to be higher as {{formula:6cf77e55-d710-4855-9add-11e52848f1f6}} is smaller, where WLS solution is located around the highest density.
This tendency is supported by Eq. (REF ); WLS solution corresponds to {{formula:5989036a-538d-4095-80cb-689596a5f1ef}} that maximize {{formula:2f31e3ef-e9e9-4600-974f-26008845ba3a}} .
Note that this explanation is not strict but intuitive since the two values of {{formula:57010995-9508-4c09-bb74-c425e2fb4508}} that respectively maximize {{formula:deaffdb3-7606-4598-9f8e-4f7f3c0dc5cd}} and {{formula:d47f1cfc-4855-4035-9467-4f3b0eb74294}} are almost same but strictly different {{cite:6122fac756e4a58846b78ce8f66cbaf593ff283b}}.
In the Bayesian approach, the "ensemble" average and standard deviation are respectively taken as one of the estimators and its error bar, referred to as the posterior mean and standard deviation, since the {{formula:d7aa911d-1ac3-4abd-b9f8-9194c4594dee}} 's "fluctuation" means the uncertainty in parameter estimation: e.g. {{formula:1666ffac-0053-45d7-b324-90d55967f402}} (a.u.) and {{formula:ef9d9654-ae20-4c17-9a6a-e6c5d83a738b}} (eV) for Fig. REF (b).
| r | 50c8efb29d35b96da0b507ce44757976 |
Moreover, deviation from the standard cosmological evolution is possible, as long as any non-standard contribution to the energy density is absent before Big Bang Nucleosynthesis {{cite:ecf1f6c123cd4345e912188f5cc01629d7d1522b}}, {{cite:ab49f9338e5da89aa74d04d5e36b32144a7b90f3}} becomes active; for temperatures {{cite:ff9cef13640405e21e2e0bdf2249b2e8e21b911b}}, {{cite:fec84d4badb2b061b0b8167a58b920e3954bac5f}}, {{cite:5bd666ba14c138253afa040ddaa3d2668f2e9c68}}, {{cite:8539f9da504af05911f6fcd763b4e8a38ed7970f}} {{formula:f9965e1d-1696-4606-a795-aae326188d0b}} . For the QCD axion, such studies have been performed {{cite:e03227007a1377a238652d936cca45fff96661d1}}, {{cite:2aba95c70ff340baa2157664a77e5f111162f898}}. However, updated experimental data or new ALP models may require more similar systematic studies. To the best of our knowledge, a library for the calculation of the axion (or ALP) relic abundance does not exist.
| i | fc610b4604773dfed8960d5ea6138c4b |
In this work we have focused on modeling multivariate extremes and followed a proof of concept approach. We can use the same idea presented here for inferring parameters in other statistical models. In particular, non-Gaussian models for dependencies tend to be hard to estimate by using classical approaches. Extensions for future work include nonhomogeneous Poisson processes {{cite:dcbe3cc35d9fb7b970b3c6de1f1d29ac075ec778}}, epidemiology models {{cite:617994d44392fb85c0590d83de2e9fe4bcb55a07}}, discrete stochastic population dynamics models {{cite:bcbeae4c1ad62424ccbb1ad0241fb80a2cc8b6b9}}, and random set models.
| d | c0f06384b398d2e649aed048bd0f4b45 |
Stereo image super-resolution (StereoSR), aiming to increase the spatial resolution of stereo image pairs, becomes a growing research direction and benefits a lot from the deep-learning. Though remarkable advances have been achieved, convolutional neural network (CNN) based StereoSR methods {{cite:8e3493699f5cf6f6e693f9eb6cc0b61a63893097}}, {{cite:b33cbb327231e30c1773a171d28b73b480230ac9}}, {{cite:c3e9b5f89601d7b140d6005792e627b74eb342ab}}, {{cite:0f162dce0cadef831ccb3cc5b39e6deef36e2933}} are mainly designed to improve the common objective evaluation metrics (peak signal to noise ratio PSNR, structural similarity index SSIM {{cite:5ab1726c0f8da7559cceac16d8f2bb6ceb4832f9}}). These StereoSR models are trained with the pixel-level mean square error (MSE) or mean absolute error (MAE) loss. It is well known that MSE/MAE loss forces the restored pixels to be the average of all possible solutions, which leads to high PSNR/SSIM accompanied by poor perceptual performance.
{{figure:1154f066-6d7c-4906-a0ae-01b6cbcea1ab}} | i | 30e8279bd911d018a10fe19ac82ce0a1 |
We report the IoU scores of all the methods on the CARLA and KITTI {{cite:79eaf549d2e831aabe5ccde14f93dee645ede62d}} dataset in Table REF and Table REF respectively.
As we can see from the tables, SBEVNet achieves superior performance on both the datasets.
We also observe the increase in performance if we use both stereo information and inverse perspective mapping.
IPM yields a greater increase in performance in the CARLA {{cite:24263fbf45f5f82dd9475f8ba11f9f87a352f7d8}} dataset because the ground is perfectly flat.
If we use only RGB IPM along with stereo, the results are slightly worse on the KITTI dataset because the ground is not perfectly planar.
We see that degradation does not persist if we also use IPM on the image features.
For the KITTI dataset, we see a sharp improvement over pseudo-LiDAR approaches because of inaccurate depth estimation. On the other hand, our model does not depend on explicit depth data/model.
The results of MonoLayout {{cite:26a1e9b4c4588df19acfb7f580de54d66d22424b}} and MonoOccupancy {{cite:734392a7c9f76507f6e03a3f0e2d2e1c0aad1978}} are inferior due to lack of any camera geometry priors in the network.
We also show the qualitative results on the test set of CARLA {{cite:24263fbf45f5f82dd9475f8ba11f9f87a352f7d8}} and KITTI {{cite:79eaf549d2e831aabe5ccde14f93dee645ede62d}} dataset in Figure REF .
We see that in certain regions SBEVNet gives outputs closer to the ground truth.
For example, Psuedo-lidar fails to segment cars in the KITTI dataset.
We also observe a drop in quality in the estimated layout as we move further from the camera.
| r | cd335d8fcb06f0a4cc756f1d939cf7d0 |
At the multi-modal fusion stage, we utilize the early fusion strategy. Early fusion incorporates multi-modal information before an encoder stage and has the advantage of retaining finer local structures and neighborhood relationships. Opposed to early fusion, late fusion is usually adopted for multi-modal learning with modalities from different domains to capture higher-level semantics, such as fusing information of images and depth {{cite:7baad4c0fadde67c814a3e4c50f3dc27b29adeb1}}, {{cite:abac947ac77d5e303ab2f336d3bb5c3b32fb0071}}. Our SCADC operates information fusion only on the depth domain, and thus early fusion of retaining local features and structures is more desirable.
| m | 00438a0d9e27b8e47558202c5693370c |
Some simulation results are presented in this section with the intent of highlighting the strengths of the proposed system state estimation GRDS approach with respect to the state-of-art solutions described in {{cite:7f37938f303e719c272cd7436dced5c4f9f6b156}}, {{cite:b73493abfef722275acf7655313a5669502b5711}}, {{cite:07a05ba5543775dc89479a100287306269e34e04}}, especially in case of MASs modeled by graphs characterized by {{formula:124e7ed0-bacb-42a2-a9d5-950844460000}} .
| r | e99a4bb999503c3a106f9bc7b38c14c1 |
Before concluding this chapter, we would like to mention that
the connection between CM Lang–Trotter conjecture and Hardy–Littlewood conjecture was realized
long ago, for instance already by S. Lang and H. Trotter {{cite:44ae4e25b287ced110f60aa333c75f145b158ddb}} and B. Mazur
{{cite:c78aa9f5376e4dd6c77a5ece32adc45d955ae1f5}}, at least in some special cases.
The precise relationship, however, was not so clear due to several subtleties.
The main difficulty lies in the understanding of the Lang–Trotter constant defined
in term of Galois image.
This constant is necessarily
quite delicate because its conjectural answer (by the Comparison Conjecture)
is already quite delicate as will be seen later.
| r | e0d1062904502cb63c928a97954a3c83 |
To enable reconstruction-based OOD detection that is not dependent on a fixed information bottleneck, we propose making use of a trained DDPM {{cite:039631ecdd43c7d2a1fd87a87db2bc635f3788a5}} to reconstruct images. During training, samples {{formula:48ce89d1-ed30-41e2-ac6a-85eba968b650}} are degraded according to a fixed process with Gaussian noise according to a timestep {{formula:79e2628b-b401-447f-b97d-5c4052d7628c}} and a noise variance schedule {{formula:9ec0d935-d55e-425f-a568-1a09057746f6}} to produce noised samples {{formula:debed03c-6974-42a4-8ed5-9e9f40c9bea9}} , such that
{{formula:e6bac2e6-a71f-4b7a-a63d-77aa2c2dddbd}}
| m | 1dc9bdff0bce95a7ad2d872394fbdbba |
Visual understanding is an important research domain with a long history that attracts extensive models such as Mask RCNN {{cite:2cf5e5a2466ee392a8893f9028d382f31a59eb85}}, ResNet {{cite:a467da4ba20205426e83bffa127e48f575a550bf}} and UNet {{cite:7581d405fa963e4cb28f28865b62d8bdc1107143}}. They have been successfully employed in a variety of visual understanding tasks such as action recognition, image classification, pose estimation and visual search {{cite:7ec67b7e424e6cc0e38a4c1fe2227793caaa1f60}}. Most of them gain high-level understanding by identifying the objects in view based on visual input. However, reliable visual scene understanding requires not only recognition-level but also cognition-level visual understanding, and seamless integration of them. More specifically, it is desirable to identify the objects of interest to infer their actions, intents and mental states with an aim of having a comprehensive and reliable understanding of the visual input. While this is a natural task for humans, existing visual understanding systems suffer from a lack of ability for higher-order cognition inference {{cite:401ec67a9490d089df8b0554b5cc0b9042337631}}.
| i | ec009eb2813b56cba8a531b1b24a36de |
where {{formula:5385c32f-afec-4e83-85e9-623a195f42dc}} is a temperature hyper-parameter for scaling the feature vector distance {{cite:b69ef286553368b1d256f7585187c29ab6f9e7cf}}, and is empirically set to {{formula:bf5d6f9a-9812-4605-a88b-b16a056c9758}} here. The patches are randomly cropped during the training.
| m | 784373d2490da147bb139d368b9d4ba5 |
Furnished by the above results, we can proceed now to calculate the central charge of the dual CFT. In fact, the Poisson bracket of Hamiltonians (as the generators of symmetry for the diffeomorphisms) becomes a Virasoro algebra with a central charge {{cite:10fb44041a588495d29e3a04178e0247457bc646}}, where the central charge is given by
{{formula:1cc8b944-ce04-462c-83b8-47f1b0fd768b}}
| m | 7879b68ae4e611fc3eb25fabef43d634 |
Model Generalization: Recently, domain adaptation and generalization are very hot topics in the natural/medical image segmentation fields {{cite:2a4e7d3c85f180e34facf4211315f4c898ec2517}}, {{cite:2c0db663a4545f01f066c5bbe02f1b765405beac}}. But for the abdominal multi-organ segmentation task, there are very few studies {{cite:8dca824b598a601ec7809986e61a19f56c0ed2ce}} focused on investigating the generalization problem. This is mainly due to lacking open available multi-sources and large-scale datasets/benchmarks. In this work, we investigate the domain gaps between our build WORD dataset and open-source datasets BTCV {{cite:db294a03e82aa1359d2a3aa0aaa9e948bba651f0}} and TCIA {{cite:b5d863db94008a363d61ca6a1f578a4b0bf30452}} and find that there are significant domain gaps across different source datasets. To boost the deep learning-based clinical applications, it is very desirable to train models with good generalization and high performance. So, we build a benchmark for robust and generalizable abdominal multi-organ segmentation research.
| d | 2888993fd740f66dc34945ed39a1364d |
Unlike prompt tuning in NLP, the meaning of the uLM's vocabulary is not obvious.
In NLP, it is usually simple to identify how to define the verbalizer {{cite:adf3b2ff60a0f56664e56e595e6d7bce9ccf31e2}}, and often the verbalizer is even an identity function when the prediction target is the vocabulary itself {{cite:81cabc4da79afbcb48385e2b586e74d2cbb63026}}.
This paper leverages a heuristic, frequency-based approach to define the verbalizer. How to better identify the mapping between discovered units and task labels is critical for performance and remains unsolved.
| d | 63613820157945097d8ee17908d7b004 |
Deep learning become extremely successfully in the last decade. The image recognition performance on ImageNet {{cite:47b0660d83a539822c5de176621c27c268b49d43}} has surpassed the performance of humans []. Google and Apple built remarkable intelligent speech assistant using speech recognition and speech synthesis technology. Self-driving cars has been invested a lot and given high expectations. We can see them in the near future if we are lucky. However, most successful deep learning systems are supervised learning systems which means that they demand a large a mount of labelled data. The consequence is that it will become extremely difficult if you want to train your model on mobile devices and you can't expect your model to work well if labelling is too expensive.
| i | f485f3e66027619b253f32a999442663 |
which improves the numerical performance. This is inspired by the approach suggested in Ref. {{cite:7458e532a420da94fc20ce9ca9e64b76304d1496}} for learning rational functions. Some care needs to be taken when drawing the initial conditions of this model, so that the atoms do not fly off to infinity. Besides working in the zero momentum frame, we have to limit the amount of energy in the system, which is given by
{{formula:fed72144-a765-469b-bc87-1ea885a6d645}}
| d | 186e0fc50864eb227d4a6213a7201927 |
where {{formula:dd3e679a-bb5a-4f64-a143-59e1d4074a01}} is the Gamma function {{cite:36f164ce00199fd15a4ae5aa6d6ee8fe4d317da7}}.
Figure 3 shows the shape functions for numerous different values of {{formula:f0daf51b-e1c3-47dc-a0c1-8a0647df904b}} . Note, that all positive {{formula:10794f70-299d-4618-8608-cc0ba1884191}} s mean solution with asymptotic
decay which means that for {{formula:f6eb8af0-2bf3-401f-8c03-d5e578babc20}} the {{formula:2d5cfda9-7f92-4f37-b4c8-53b691ad7d6f}} .
This can interpreted as certain kind of boundary conditions {{formula:15c84e9d-e192-4e26-be7a-beba6d105fc7}} and {{formula:963ffe21-6505-4536-9b2f-bc03f093a36d}} .
The {{formula:1bb18589-60b6-470b-bdc9-2c89679802e8}} means additional oscillations.
Zero alpha value means a solution which converges to a finite value,
and negative values means divergent solutions.
| r | 5b0e859b699fa439a6ac2aa3bfaab512 |
Inpainting networks are typically benchmarked on Places2 {{cite:7c25357f523f25724cae71e37127a8e516e94f6f}}.
However, this dataset does not have high resolution images for evaluation purposes.
Instead, we will use images from the Unsplash-Lite Dataset, which contains 25k high resolution nature-themed photos {{cite:dd6bc6f4fd6031a90b901bfcd19de0fecbb9e13d}}.
We randomly sampled 1000 images to evaluate on (linked blue[here]).
| r | e42ccbcda6392f51754d69e6767e9282 |
For the decoder, we evaluate the following approaches (1) Graphite {{cite:ad2e847da1862d016608b22cf86b12edf19b5cdd}}, which uses a simple dot product decoder after
GCN-style iterated message passing to transform the encoder output node representations {{formula:14e7c6e2-53e3-44cc-a270-30bdbce42369}} to final node representations {{formula:69b926ae-8eb7-4e8c-b16d-2d3672f1f5d7}} . (2) DGLFRM {{cite:158c8cfa23d512a5e0e3d77a2c9f290a5aa51cf9}}, which applies a stochastic block model decoder. We combine the SBM decoder with the Graphite technique of transforming the node representations {{formula:9b6fc015-2aa2-4db6-b8e5-341683dc5dc0}} to another set {{formula:ee5963aa-1eec-4e92-8e66-59d87dfeb501}} . (3) FC, a fully connected neural network {{formula:f980b1b0-3775-4c40-8991-6987595e93e9}} that directly maps the node representation to a probabilistic adjacency matrix. We include the vanilla FC decoder as a baseline because it avoids confounding the effect of kernel regularization with the influence of a specific decoder architecture.
(4) MolGAN {{cite:99baf9e85beecb481b5a97a53a34d78a0b0dd87e}} is a SOTA method for that generates edges independently based on a GAN architecture.
| m | 69796c3723bee4a0cdcb742de962e478 |
In Fig.REF , we explore in more detail the known
anticorrelation between the impact parameters versus {{formula:911f202b-6b37-43ff-9042-eaa2620550c9}}(H i). Considering only the high-confidence DLA associations from
the literature and the IFU-based searches probing large scale
environments, a clear trend seems to emerge, although with large
scatter. Indeed, high {{formula:5f8f8abe-d4b1-4d74-94e2-b11c804d67d6}}(H i) systems are observed at preferentially
small impact parameters, with a Pearson's correlation coefficient of {{formula:93ce61f2-29c6-4f7b-af36-585a2f53bf50}} and a {{formula:b0c94ebf-0bb5-41ec-b6ef-555c98c86ad3}} value of 0.029 {{cite:5667dc8eae1586274a83b6435deaf5b589ad0eab}}, {{cite:86be6f6099a025c7b9e80b740a9c5b6aed32e5d4}}, {{cite:c1d9787eff24cca70854798508798c437a0b5951}}, {{cite:85c8c0409162f5629fd591d7feca986a8a02a703}}. It should however be noted that some of these detections rely on long-slit spectroscopic measurements, for which only small impact parameters are accessible. Larger samples studied with large format IFUs are needed to confirm the significance of this relation.
| d | 8126e71375a3d32a15084e12cb2a1c2d |
Intriguingly, the score of {{formula:7b6af89f-8d51-430f-bde4-7d8ac25192a7}} for our baseline mode parallels previous results observed in overhead imagery. {{cite:30c0916394a8cc8cc7a6c531e76769a3824b4fe4}} studied object detection performance in xView {{cite:049c7136dd7d464e7fbd370ff2343dc7e41417ab}} satellite imagery for various resolutions and five different object classes. These authors used the YOLT {{cite:f3556e0ca6146a27d74dd6e2e6123873041db4d0}} object detection framework, which uses a custom network based on the Googlenet {{cite:6eb737f349bce7f95e4783051c3b1f17180461be}} architecture. The mean extent of the objects in this paper was 5.3 meters; at a resolution of 1.2 meters objects have an average extent of 4.4 pixels.
| d | 57698685742034d75e0d05d6cd2d8761 |
We have presented a method to assess the effect of weak and strong lensing on the estimation of the luminosity distance for a population of astrophysical GW sources, taking into account selection effects due to the finite sensitivity of a GW detector.
Since a realistic GW detector has a finite horizon, the probability that a source detected at redshift {{formula:9da2c10a-9f83-4133-8c2c-c44b543448af}} is magnified is higher than the probability that it is de-magnified. As a consequence,
the mean of the distribution of magnification for sources at sufficiently high redshift, is shifted from 1 to higher values. These effects, which we dubbed lensing selection effects, disappear in the limit of a perfect instrument (or analogously, in the limit of sources at redshift much lower than the instrument horizon).We stress that our derivation of lensing selection effects relies on the geometric optics approximation. We expect wave effects to become non-negligible in the LISA waveband only when dealing with diffusion off sub-galactic structures, see e.g. {{cite:aa373dadeea1ec9004581efee170e763c8e6048b}}, {{cite:05c0456f7a45ad933811844a221ebc23ba379dfb}}, {{cite:28e688e897c3da42a0e66e47bf5b6103e54f605e}}, {{cite:144ed7cf489c0f04be69aa02181a8ea357bf2093}}, {{cite:b13bb144306799284ea49200b86d9b96c02c2db9}}, {{cite:47c155f69f3a904ba58347041fe4cfb8ac83eb6b}}. Diffraction on sub-galactic scales makes lenses on those scales effectively transparent to GW in the LISA band {{cite:aa373dadeea1ec9004581efee170e763c8e6048b}}, in contrast with what happens for lensing of EM sources (see e.g. {{cite:00dc2ab74377050c33b778d5d72e2b6c28c3f61c}}) as the EM spectrum is at much
lower wavelengths than any relevant astrophysical structure at cosmological scales.
| d | f91d87667939a7e830828fbc09b0928d |
Despite considerable progress, there is still a long way to go before fault tolerant topological quantum computation is experimentally feasible {{cite:2764030af7b08c080eecb77ee4a8d98bb7989ed2}}, {{cite:171ac399220547bd38ba71f8d5f088b0afeefaa7}}. Thus, there is a motivation to explore less traditional schemes for realizing and manipulating anyons. For example, it was proposed that anyons could be synthesized by coupling weakly interacting (or noninteracting) electrons to a topologically nontrivial background (or topologically nontrivial external perturbations) {{cite:3270efac964d0fdd9fb820e6937477157bbfc815}}, {{cite:1c5ae9a737104d75180af9ab16ef1cb391778443}}, {{cite:8c2bbedac8a8d1028070c1042968ce34604170cd}}. In refs. {{cite:3270efac964d0fdd9fb820e6937477157bbfc815}}, {{cite:1c5ae9a737104d75180af9ab16ef1cb391778443}}, anyons are proposed in a system of an artificially structured type-II superconducting film, adjacent to a two-dimensional electron gas in the integer QHE with unit filling fraction {{cite:3270efac964d0fdd9fb820e6937477157bbfc815}}, {{cite:1c5ae9a737104d75180af9ab16ef1cb391778443}}. A periodic array of pinning sites imprinted on the superconductor will structure an Abrikosov lattice of vortices {{cite:3270efac964d0fdd9fb820e6937477157bbfc815}}. Anyons are bound by
vacancies (interstitials) in the vortex lattice, which carry a deficit (surplus) of one-half of a magnetic flux quantum {{cite:3270efac964d0fdd9fb820e6937477157bbfc815}}. In ref. {{cite:8c2bbedac8a8d1028070c1042968ce34604170cd}} anyons were proposed in integer QHE magnets. Magnetic vortices in this system are topologically stable and have fractional electronic quantum numbers yielding anyonic statistics. Anyons were also proposed by using topological defects in graphene {{cite:6536c2f2af29a60325438bc31eed6d71a190c67e}}.
| i | 1ad2fd210f62875347116d5b3c50e053 |
(a) By 1.6.11 of {{cite:8b286f4b41521767049a466c20d925829501456f}}, in a finitely generated group, every subgroup of finite index is finitely generated.
Thus by the restricted-finite condition on {{formula:548b3457-c1d7-4b9f-aa85-54d9504ce0bb}} ,
all subgroups of {{formula:3fe4e7bf-7d7f-4023-a5d8-4d61b7bb1f7d}} are finitely generated. The last statement is now clear.
(b) Suppose that {{formula:c50ccc37-ec6f-488c-ac3c-77e1c63e8ef4}} is not finitely generated. If {{formula:eda6f728-4ea5-487c-b98f-8e48daaa7604}} is a finitely generated subgroup of {{formula:53202c98-b3bb-41f8-b1a6-9763597efedc}} which is not finite. By the restricted-finite condition on {{formula:9e20253d-25e8-48cc-9cbb-dc55125fb952}} , {{formula:03342716-b9e6-4cf5-8165-507d8671e4db}} is finite.
Hence, {{formula:67ead9e9-aaf0-42b2-8227-c0e14c44ed51}} must be finitely generated, contradiction. Therefore, {{formula:dcec7d16-1dda-4f40-b7b5-61a8f6519a3f}} is locally finite. Clearly, locally finite groups are torsion.
| r | 400ace333db7c11ae1891085cf222091 |
In our study here, we examined a doubly holographic model, which from the brane perspective, described two two-dimensional bath CFTs coupled to either side of a two-dimensional eternal black hole, as illustrated in figure REF . While the bath CFTs are in thermal equilibrium with the black hole, the entanglement between the two subsystems grows as they exchange quanta at the microscopic level creating the potential for an information loss paradox {{cite:2b42e73b7de4e7291f03256343bca9b15086535c}}.
Similar doubly holographic models of two-dimensional gravity were introduced in {{cite:17fcd802faeba1a8915ae0dee1af70e80ad204b3}}, {{cite:649357d83b25d5278075bad5bb21a8c27b5d1a79}} and extensively studied in, e.g., {{cite:f3f20b7fc333824b32c56cb083c2e3b1c2a73c5c}}, {{cite:681371282ed5029f17ad5abec6faa83a18207489}}, {{cite:0fb94c9b499561d6e1103d0253d041b678c4a519}}, {{cite:58fc4e290a7b4b1e61f58730079c220d9713ce8d}}, {{cite:ccdb32b826ed6b315698c691897bd9f74508be19}}, {{cite:59e44bc26e7973453acf998374afab7e2d71a972}}, {{cite:faef412e5ebe32e9b5df66d95ce462b095bb1ea7}}, {{cite:49e0f73d6b3068213b76eaac2f7eb46f594fb448}}, {{cite:3459ef344ea43c7847618142eee9d9f3f28fe979}}, {{cite:566c6cfe91b5f5b9ea75dbfb04ff44736eb3eef1}}, {{cite:6191b7f5882764fa8519d02e9675228d79252597}}, {{cite:dc6c53c6af094a6f7552e390737624b4466f2be0}}, {{cite:b391cb349a9d6788babbf926511ceeaf2519107f}}, {{cite:d08c4cdb2eadbd71f7fc58bbb8715348d0ce3ab3}}, {{cite:bf0932e43c79c343d9c42238fc79574665235bd1}}, {{cite:4ec7c978df0cef977f07df55a22c104eae0465b5}}, {{cite:dad07c93e0ee14aa6d5c43e8c3f405f1d3cc061d}}, {{cite:c89a46dbe33bd98c2a46f3a455bb7bce92a58811}}, {{cite:b64a310f6e1f478f401e92c8d8cd3774f1f3e1dd}}. A minor difference between the present work and those earlier studies is that, for the most part, they involved {{formula:16564884-8b2a-4d0a-bc48-4702ae91f193}} -orbifold or end-of-the-world branes analogous to those appearing in the study of holographic boundary CFTs {{cite:c938f0c9c8f4534f6768cc12bc98587ecca5ce56}}, {{cite:101cefee5ef3125099fcc366b0acef3e8344642b}}, {{cite:c183d4310a7ebdc79599131ba35a23f36a95aa94}}. In contrast, following the constructions of {{cite:d77cd9334a844ee00bc5527c42331ef80796af72}}, {{cite:78098cb492e5284d4e22db2209282878a65d2997}}, our model involved conformal defects which are dual to two-dimensional branes immersed in the three-dimensional bulk geometry.
| d | 7851d4c12aef44a5a02970a2f9c4a864 |
When dealing with complex biological phenomena, there are necessarily limitations in the deduced models and acquired data.
To assess criticality via finite-size scaling, ideally cell density is varied by orders of magnitude. However, this is not
feasible in this biological system. If cell density is much lower than about 1/3 ML, cells do not aggregate; if
higher, experiments would need to be conducted in 3D with major technical difficulties.
Despite the approximations, our model allows the identification of the key ingredients for certain observed behavior.
For instance, an earlier version of the model showed some level of aggregation but no finite-size scaling.
By investigating this short coming, we noticed that streams were too narrow due to nearly negligible volume exclusion.
However, quasi-one dimensional streams restrict cell movement and suppress criticality, reminiscent of the missing disorder-order phase transition in the 1D
Ising model according to the Mermin-Wagner theorem {{cite:3d0878b29337113a48a38afa3fda9a30b14da4d8}}.
(Note that the 2D Ising model is a borderline case, but it is still possible to formally define a phase transition according to Kosterlitz and Thouless {{cite:df00b9f413f7e62ad50fa993e2add852f3a15167}}.)
In our simulations, only when volume exclusion is increased and streams become broader, critical-like behavior emerges (see also discussion in {{cite:9d1b624af84e6e5dee5f3bf293fefe66900b723f}}).
| d | 9ebcf12d581f458cf1baba6de5425a2f |
where {{formula:f1e0001f-f5fe-47dd-a761-32d78bd62369}} denotes the number of quantum bits (qubits), spin variables {{formula:efc8f64e-4d1a-4b6d-8678-52ea30a9417c}} , and {{formula:19dd946a-d8d3-4e3e-96f4-065d785e55b0}} and {{formula:2d376fa8-6803-402f-9a53-989c5e9b95fc}} represent local fields and couplers, respectively {{cite:1a85615af7a4c43618d667da26c888b1375561f4}}, {{cite:cce8923d4a7454ea06feff8c9ff218df8332e47b}}, {{cite:02655119d1b14eb6ab58f70ad802c4bbcfefb6f8}}.
Quantum annealers can efficiently recognize the region of the ground state(s) of the given Hamiltonians; however, they generally fail to get to the global minimum, regardless of how close they are to the ground state.
In other words, unlike classical annealing that always converges to a local (or sometimes the global) optimum, quantum annealers generally yield an excited state(s) that are not necessarily a local optimum {{cite:6b600484b591b658d0e19fcdb58cad60c0f059f2}}, {{cite:11c09680696d78a6cd9939a463c7d43ea05720da}}, {{cite:0f9e286a03d8020d85082fd6d16d1e5a8c023745}}.
Hence, we can expect that applying optimization heuristics and meta-heuristics on samples attained by a quantum annealer to result in new (or synthetic) sample(s) with lower energy value, specifically on systems with glassy landscapes {{cite:c70a329b06e34a694fed1fde2fbb93575dfc8cb8}}.
In this section, we start with a local optimization heuristic, called single-qubit correction (SQC), and then extend it to introduce multi-qubit correction (MQC) scheme for mitigating errors in quantum annealers {{cite:6b600484b591b658d0e19fcdb58cad60c0f059f2}}.
| r | 044254980f5077c96d87673a22f72de3 |
From the results in Table REF we see that our more structured method is the most optimal for extracting rules. MAP provides a coarse representation of the labels and enumerated rules with no parameters. It can express these rules better than LSTMs but lags behind our method. As the number of free parameters increases, the LSTM models are more likely to pick up spurious signals in the data that are useful from a cross-entropy optimization perspective but deviate from the underlying generative rule representation (Hits). This can be seen with highly parameterized LSTM models and video models: I3D {{cite:cb5dfe62854c368c3c568b6480c065515b66d0b2}}, TSN {{cite:e65c9b1c5599e6446d6f205375e765868e0cadd9}}, and TSM {{cite:e156b105bd870f488ad81d26addaa732866b86d4}}, leading to larger gains in mAP in Table REF .
| r | 5a703ee4bfc4444f53dd723a8521889d |
Our goal is to learn common-sense physical reasoning in a purely unsupervised fashion directly from visual observations.
We have argued that in order to solve this problem we need to exploit the compositional structure of a visual scene.
Conventional unsupervised representation learning approaches (eg. VAEs {{cite:a07410e99f871c671ee5215754edc580056af7b1}}; GANs {{cite:041c3c282c5efcb4601ab1625a0a2728fce4cb29}}) learn a single distributed representation that superimposes information about the input, without imposing any structure regarding objects or other low-level primitives.
These monolithic representations can not factorize physical interactions between pairs of objects and therefore lack an essential inductive bias to learn these efficiently.
Hence, we require an alternative approach that can discover objects representations as primitives of a visual scene in an unsupervised fashion.
| m | e12fa76c20a81df7d2d726f48082e651 |
A more detailed insight can be achieved from linear perturbation theory {{cite:2e0453728b052c8e25c1a8e4df534c24c094227c}}. It can describe the properties of structure formation, with respect to their sizes, masses and time-scales. Understanding these processes in detail is important in order to get a better view of the previous history and subsequent evolution of a disc. However, in these considerations, a razor-thin disc was assumed, characterized by its surface density {{formula:a068cc8c-dc43-4abb-8354-f96062ad6cf1}} . Already {{cite:bb035ba97e53dca162bcc030537432dd18301e60}} pointed out that a disc with finite thickness leads to a reduction of the gravitational force in the mid-plane, where structure formation takes place. Therefore, the disc remains stable even below a {{formula:ac406946-8464-4604-8fb4-d81e0dbab244}} -value of one and a revised critical value has to be determined with {{formula:eaae978a-0809-463d-8f75-f04dc3af86be}} , describing the unstable regime.
| i | 00e05c85963f5c70505af4782c80bf9c |
In the study of compact stars in modified gravity, {{formula:a088799b-13fb-4ac8-98fb-150619085e84}} gravity has been intensively investigated so far, where the action is written by an appropriate function of the scalar curvature {{formula:69c8c6d4-9917-49fe-b65c-3e0be55001e1}} .
The existing studies formulated the modified Tolman-Oppenheimer-Volkov (TOV) equations for the general {{formula:99f25e9d-36c2-4b55-a243-17b2fa0ac12e}} function, and they have revealed the mass-radius relation which can explain the massive neutron stars with concrete {{formula:3fc5028c-d585-4ea6-8e9b-e749828591ed}} models and EOS.
Those studies indeed show synergistic constraints on the {{formula:97de0b6a-b257-4ffe-bd22-b7bd8f65cd7f}} gravity models from both cosmology and astrophysics. {{cite:2b91bcf1ea6947d37de285fc547dcf265eccba70}}, {{cite:654ae0fefcc6710f11d3bcaa18a0314dc38e5d21}}, {{cite:c9216bebe4699271c6c4c783204ec7599901ca3b}}.
On the other hand, the degeneracy problem still exists, and one naively expects that the functional degree of freedom for the {{formula:20d270aa-a795-487c-9cc1-0375022cd9cb}} function degenerates to that of the EOS.
It is analogous to the well-known reconstruction technique for cosmology {{cite:cfb5d7c2b55bbd226a1ed82c14361b956e91d19e}}, {{cite:13fa021d378607ebc8faa3af82005a984872a800}}, {{cite:f687a1ba0e0a4358e4d72749c1369aaa4426155c}}, {{cite:6c95fa39b23f12571bf9105390549e7de89548d9}},
where an arbitrary time-evolution of scale factor can be reproduced by choosing {{formula:82862f3b-bdba-4906-8ee6-abd498fa16e0}} function.
Furthermore, the field equation in {{formula:20a3d56b-c098-43ab-adbf-b0bae4f3c965}} gravity includes the fourth-order of derivative,
although the Einstein equation does the second order.
Thus, one needs to impose additional boundary constraints to solve the modified TOV equations numerically.
There are lots of arguments about the boundary condition;
for example,
one can solve ordinary junction conditions for the Ricci scalar {{cite:b7296b27f8c80e6b0896f94abb84f2755f599515}} or impose a condition that the Ricci scalar takes the same value as that in general relativity {{cite:b6c00bd9c6485f8686a682c205be2da667b843c4}}, and the scalar degree of freedom in {{formula:d476aca0-cb6b-4767-ad7c-4291a162bb50}} gravity may allow the nontrivial exterior solution {{cite:ad5814e3f06fa32f8af46cc46dd9414422f7308a}}.
However, one also expects that
the boundary condition at the surface of compact star potentially constrain the {{formula:00519f25-98bf-462d-acb3-be19dff3a23a}} function,
assuming the Schwarzschild space-time as the exterior geometry.
| i | 2761f8a63876fb0cec5c5bb7c98b4543 |
We compare the searched transformer with two CNNs (ResNet and ResNeXt) and the state-of-the-art vision transformer Deit {{cite:d4e0bc5427c6243bf1a10c74a7149eef83182bdc}}. Table REF shows the results under different computational budgets. The results for the existing models, such as R18 (Resnet-18) and R50 (Resnet-50), in Table REF are copied from the results reported in {{cite:d4e0bc5427c6243bf1a10c74a7149eef83182bdc}}. We also use the training setting in {{cite:d4e0bc5427c6243bf1a10c74a7149eef83182bdc}} and report the results for R18, R50, X50-32x4d (Resnext-50), and X101-64x4d (Resnext-101), with {{formula:c2c78d1d-d6a6-43fc-bf82-71e2172153b4}} followed by these models in Table REF for denoting the same training configurations.
Our models achieve better accuracy than all compared networks under similar FLOPS restrictions. For example, our searched model with 1.3G FLOPS restriction achieves {{formula:9759d5ee-7967-4fad-8f88-6d36f852bea3}} accuracy score, which is higher than both Deit-Tiny and ResNet18 (R18) by more than 4 points and 6 points respectively.
Our searched models achieves obvious improvement in accuracy from the symphony our two designs: local information and architecture search. The local information brought by Conv1d without proper distribution has limited improvement according to Table REF . However, the searched global and local information distribution shows much better performance according to our ablation study and the detailed architecture search will further improve the performance of our GLiT model.
{{table:d8eb7183-857c-403c-aebb-e27fa2538fe1}} | r | 465cbc92cf34b66686a0c5b6d541ebd3 |
Table REF shows the quantitative results of our proposed method on the KITTI Eigen split. Note that {{cite:814085d224003a7f5aca7ced7a2e6486632c4225}}{{cite:862409ef291c1a2316f8c9b6833b5168ff2d6177}}{{cite:f45ce30e6bca186f78040fd05024cf07ac955793}}{{cite:ff9bcad90ab6d72636b45828414193b077189996}}{{cite:ff32a7ea8f5806c1d44ad12f815c423e7a12e7ac}} employed the same split strategy as our method.
{{table:740f42f4-d939-4375-b76b-d50ad5a534f5}} | r | 658eb075b39e7c81e0de1b639773f355 |
Learning approximate equivariance has been recently approached through novel layer operations {{cite:a280f146edb36a77c1fbaade6925df65e7427153}}, {{cite:aa07877d9f8c6776c7da0812325dc78d1953820e}}, {{cite:4116d97dd46514fddc4d3b93fa00ce328ae9bfff}}, {{cite:9f4b75788333c940618cf3b85d2f9ee4412e7a4c}}, {{cite:5b8b6669a954932cdc050c3a2a00e75e28dd1055}}. Separately, the field of neural architecture search (NAS) aims to optimize full neural network architectures {{cite:4f545e9a9d0d1bc8a003a25ccc63b844ea9f7b42}}, {{cite:dc5e02888f812f47711616211483e07256094cdf}}, {{cite:d3e872bbcade9f9de50a44e5eb736f3b65b4ce98}}, {{cite:49972066249f67e93e0e16c4edf6c502322b4933}}, {{cite:09563100cce75add6378c569dc33e5eb729b11a5}}. Existing NAS methods have not yet been developed for explicitly optimizing equivariance, although partial or soft equivariant approaches like {{cite:d59753e009d030b151a5b04470d4d601024eb758}} and {{cite:da68b470b4b410927618073c8964e9ba7da30592}} do allow for custom equivariant architectures. An important aspect of NAS is network morphisms: function-preserving architectural changes {{cite:3362c71f0342e7b7bc3a1b924a0f8c691aeaa904}} which can be used during training to change the loss landscape and gradient descent trajectory while immediately maintaining the current functionality and loss value {{cite:fcda01a2a179e0973458ea372d23a3bf3b4fd2db}}. Developing tools for searching over a space of architectural representations of equivariance would allow for existing NAS algorithms to be applied towards architectural optimization of equivariance.
| i | 698bcdac938c3c146fc755691235cc26 |
Where {{formula:f128e360-fd9d-4a51-8083-f508d020dd86}} denotes the element-wise multiplication and {{formula:ef505929-4009-4fbc-8a9d-f2026b3c80db}} is a Gumbel-Softmax {{cite:c294709fe004a4aa0ce6c6197803265088b592e0}} function to obtain the spatial-wise soft attention of the basic adaptor activation {{formula:f09d1d8e-b0fe-4495-bdcc-97ab65a68e69}} .
Furthermore, instead of fully utilizing the pre-trained model as prior works {{cite:5b5e4ec08e32d2c17fec8cecf3dde7de040a5d6c}}, {{cite:10e1b3f8325e55cca9743d6e393ab2501121e79f}}, , we select the important weights for current domain by
turning the soft attention {{formula:03215e50-36d5-4a2c-8016-62f54b0fd279}} into binary hard gating {{formula:41793309-cb06-42ec-8fc1-9a6c4b51e880}} .
Then, the Eq.REF can be further modified as:
{{formula:69c26983-c1c1-4692-9886-ef10b6c7c3dd}}
| m | e9819e5d1eed5cfa32c74a22f16d3c9d |
In the numerical calculations, for the concerned models we also consider the constraints from available experimental data, such as the most stringent upper limit on the sum of neutrino masses by PLANK{{cite:c4dcb149b4dcb8c249f71c08db30262566496cf9}} {{formula:e3d872b2-c357-47ae-9a29-0a842fda7d67}} ; the neutrino mass-squared differences at {{formula:9abdc939-f368-45f7-8427-4337753b4651}} level errors {{cite:c4dcb149b4dcb8c249f71c08db30262566496cf9}} that are obtained via analyzing the solar and atmospheric neutrino oscillation data:
{{formula:f51ed3cd-886e-4168-b436-2ce9462e2602}}
| r | 56238227da3a8d1f58778f2eba6de2b9 |
The Liquid Argon Time Projection Chamber (LArTPC) detector is a proven technology that has been adopted by many accelerator-based neutrino experiments, including the Short-Baseline Neutrino program at Fermilab {{cite:05a7da8c512d4d90e5e529ba8ab0908a2b4ad0e9}}, {{cite:ff6feece19a0614271d86dbf66218637df58d64e}} and DUNE {{cite:2611c879ddedbdee4fdf6342529c54906300c426}}. It offers millimeter-scale spatial resolution and excellent calorimetric capabilities in the detection of particles traversing the liquid argon and the measurement of their properties.
| i | e0fad9b321f8c210a5667a9ea68a37ce |
Recent blind MFSR approaches, on the other hand, utilized a fixed kernel to upscale every frame {{cite:14b7f9696e2fdba6db02550a1c218359fb8dec81}}, {{cite:f56d0357520e4f5f7bbfb6d810806509735cbb2f}} in the same video – we hypothesize that this fixed kernel assumption can also lead to kernel mismatch.
Therefore, in this work, we attempt to investigate and answer the following questions: how does the kernel change temporally in real-world videos, and how can we leverage this change in the video restoration process?
| i | ad353887c89ee69c005444f1d3d77bdd |
In this section we will survey some of the complex-analytic ideas that play a decisive role in the theory of (multi)critical circle maps. Since these ideas are quite deep, the narrative to follow is by necessity very sketchy. For the general theory of complex dynamics we refer the reader to the books {{cite:cc06b8f75c9c39f4e34c8d5e600d34da62855328}}, {{cite:a13d195b15212530d1c00f35e0032a5a261fb8fc}}, {{cite:8adefba4594e6f1c2f171acf1a4ce6c2ddae9ce7}}, {{cite:61de0508d48cd11849b2d90da1fb06fe9e2048cb}}.
| m | 62a8ed159b3d48a282ecbb4b40164acf |
Reconfigurable intelligent surfaces (RISs) can enhance the system spectral efficiency (SE) with low energy consumption and deployment costs, which is believed to be a promising technology in next-generation networks {{cite:3ae1478399a7f71fbfb31d4e7383cae8a8e63c14}}. Specifically, by properly programming its adopted phase shifts, an RIS can control the wireless physical layer channel intelligently and then give it a performance boost {{cite:6b1496e5b9773c247ddc80a50b7cd81e95118e0d}}. One of the key scenarios where RIS shines, is for SE improvements in multi-antenna systems {{cite:328748f669f61f2ca80a5886bae2d11f55bf2170}}.
| i | e100c807569491e24b4010fbc8f242d5 |
The communication modules for multi-agent can be divided into two types: integration module and recurrent model-based module. The former involves centralized communication, using a centralized network to combine the states for all agents {{cite:89141d9aebcdb0b815b1a409afccd3d7421e0d0e}}. For example, Targeted Multi-Agent Communication architecture (TarMAC) {{cite:e5f163971717a5a9b8b1a99d871d0f0f16f64943}} leveraged a soft attention module to conduct a targeted communication behavior, by which every two agents are assigned a communication channel. The latter regards the agent group as a sequence, using recurrent models to process the latent state for the agent sequence. Deep distributed recurrent Q-networks (DDRQN) {{cite:c2fccd47bf207d03108168df7a501d08ae55d42a}} is an earlier recurrent model-based method aiming to solve communication-based coordination tasks without any pre-designed communication protocol. Then several improved models are developed, such as CommNet (Communication Net) {{cite:76869c6ae666fac01e31cc51f36251a18a36510e}} and BicNet (Bidirectionally-Coordinated Net) {{cite:7b4c778306e446becf638092f4c4ed2f41461453}}. Some works also utilized attention mechanisms to enhance recurrent model-based methods. For example, Attentional and Recurrent Message Integration (ARMI) calculated the correlation between the message and the observation by attention mechanism. At the same time, Liu et al. {{cite:f97d68bf9397bd4d8751d644c5c34956c4000cf3}} leveraged a self-attention mechanism to build a communication framework, which can learn both to construct communication groups and decide when to communicate for agents.
| i | ae4db4978ce7824d8628942415cfbf2c |
Our NoisyTune method also has the potential to empower other PLM finetuning techniques.
We compare the performance of the original RecAdam {{cite:239e0aa25d262bd3e01264397325eb2bfebada2a}} and Mixout {{cite:533c974de077a97bc398f71b4f5013bd50cc369a}} method and their variants combined with NoisyTune.
The results are shown in Fig. REF .
We find that combining NoisyTune with existing PLM finetuning techniques can further improve the performance.
This is because NoisyTune aims to address the overfitting of pretraining signals while these methods aim to prevent overfitting in downstream tasks, thereby they can be empowered by NoisyTune to improve model performance.
| m | 4465ff049bed4a49d65c41d756b00d0a |
Let us mention two key properties of the Tikhonov regularization, which we will use later in the analysis (see, for instance, {{cite:4da0c6ce95aa69fa516139b66e5431814bbe732e}} or {{cite:555aa08283048c09bf45055448a1d55f5dbac6b8}} Theorem 23.44 for its classic analogue). First let us introduce the strongly convex function {{formula:e7d2ff5e-b1a4-408e-85c5-dd5107fcd654}} as {{formula:e867d673-beff-4179-9c96-635a69c98e33}} and denote the unique minimiser of {{formula:a53f81e2-4ad8-4760-a6cb-9ca0a237be4d}} as {{formula:2efda6e4-3475-4c09-a6d3-05f39df544aa}} . Thus, the first order optimality condition reads as
{{formula:e619f8a3-759d-4c02-b086-35eed6b30601}}
| r | 8872803f79289df1ce6c20aba19ecd58 |
It is observed in Table REF that the classification accuracies of all deep feature-based methods are generally poorer than the perceptual hashing method, OSF-DNS, for both datasets.
The reason may be that occlusion leads to a distortion of the face embeddings obtained by the convolutional base of the networks, which makes discrimination more difficult.
Mainly, ArcFace attains much worse performance than the rest of the CNN-based approaches.
In ArcFace, a more reliable method to increase the feature distances is applied {{cite:a4b318cdaaf6c6592f4a877de40bbbba3f76c677}}: an arc-cosine function is applied in the angular domain so that the decision boundaries between features corresponding to different classes are more distant from each other. During the experiment, we mainly extracted the embedding features of a face and its eye occluded version through the pre-trained ArcFace model. Though both faces are the same except the occluded region (i.e. they belong to the same class), the method provides very different embeddings for them. The reason may be that in ArcFace, the embeddings of the faces are distributed around each feature center toward the hyper-sphere and it uses an additive angular margin penalty between feature and ground truth weight to concurrently enhance the intra-class compactness and inter-class discrepancy.
| d | 8876147577333eb839253449bfde48bb |
PPO {{cite:ed727cbf82f603ea41f8ad2f5d4fcb4d99161b6a}} is a model-free deep reinforcement learning algorithm, which has seen a broad adaptation in the literature as a strong baseline.
For our experiments, we rely on the PPO implementation as provided by {{cite:1b14644ecee1d372c57f0969f639db959f77905d}} {{cite:1b14644ecee1d372c57f0969f639db959f77905d}}.
| m | c1bdba0ee92b0bf859d01bd98ce6883a |
The proof is similar to that of Lemma 6.1 in {{cite:c11225fb129644243cd375e224de5515dfe3337d}}.
| r | 65273143362dbeb8d97bb08a2951dd30 |
Table REF reports the accuracy scores for several pairs of languages. Although SWOT-Flow has not been specifically designed to perform word translation, these results show that its overall performance is on par with the adversarial network (adv-net) proposed specifically for this task in {{cite:ca335ae411fdb2fd67fdcd37fa7e264c84c5f552}}. In particular, SWOT-Flow seems to perform well for translation between languages with close origins.
{{figure:23800a01-fd95-4ed0-9ea1-4ed21d4c1a9d}} | r | 8a7878437fe4d7eb00930a498ce3b803 |
The time series data available for the cryptos is subject to numerous limitations. The most important one of them is that different coins were introduced at different time points, therefore, the data available for each coin has different lengths. For the clustering problems {{cite:532921af62bdaf29f7732ae65b5600d55f5fc6f8}}, defining the distance metric between points in time series with various lengths is not conventional. For many analytical problems, this issue is easily tackled by truncating the time series to the shared sample period. We refrain from doing so because, in the analysis of cryptocurrency prices, the evolution of the data in time is highly crucial for an investigation in the short term and long term dynamics and therefore, truncating the time series would lead to loss of important information. Hence, we deal with the time series data of cryptos with different lengths and do not directly impose a distance metric on the input data points.
| m | 7d2157e511d9b0e730770bf7c907cce5 |
blueQ-learning is a fundamental approach to RL but we also experimented with other state-of-the-art model-free RL algorithms, such as PPO {{cite:b7b50940cc0aee4b26f80031883e02f11201a546}} and its predecessor, TRPO {{cite:5c0abce6f1ae7199ab6e52d5c1106fd391a09e15}}. Table REF summarises our experiments that show there is little improvement to be had. It is worth noting that PPO and TRPO demonstrate much lower standard deviation in final performance across multiple training runs compared to DDQN+PER. This is to be expected as those policy gradient methods are revered for their stability, but the best seed for DDQN+PER consistently outperformed them, albeit marginally, for the networks we tested.
| d | e15d0d7ddec627f36d4a6e444e47a2cb |
{{cite:27429d92a7134642d2da77fe583e6ab2dcff6ab0}} suggest that bursts igniting in a relatively hot neutron star envelope leave a substantial fraction of unburnt fuel at shallow depths. This fuel is brought down to the ignition depth by opacity-driven convective mixing on the observed timescale of minutes and produce a subsequent burst.
Although recurrence timescales observed during these SWT X-ray bursts could be explained by the model proposed by {{cite:27429d92a7134642d2da77fe583e6ab2dcff6ab0}} we do not find any bumps before a follow-up burst in T1 as predicted by these authors. Such differences were also noted by {{cite:3f1b104645e0034aa440985d4fa6ffcd48ccbb54}}. {{cite:c2db1c32648471218acc3560afb8f7a7db368108}} noted that all sources that show SWT bursts have a fast spinning neutron star ({{formula:22783276-d40a-4587-8c30-093b708bf3ce}} Hz) suggesting the requirement of rapid rotation for multiple-burst events and the importance of rotation-induced mixing processes. They also proposed a waiting point in the chain of nuclear reactions such that a decay reaction with half life similar to the short recurrence times temporarily halts the nuclear burning. However, {{cite:27429d92a7134642d2da77fe583e6ab2dcff6ab0}} observed that a considerable spread in the recurrence times (3-45 minutes), even for SWT bursts from a particular source, disfavoured this scenario. One plausible explanation to the existence of the short recurrence bursts could be based on the spreading layer model as suggested in {{cite:c6419c4ab9ed4034d20a257d42491a0eb8725aaf}}. However, these authors also note that further refinement in the model is needed for the complete understanding of the existence of SWT X-ray bursts.
| d | d5ac5734355f149e1ef2b045c05a9a8a |
Table REF shows the results on ImageNet, including top-1 and top-5 classification errors on the test set, number of weight parameters (millions), and search costs (GPU days). Following {{cite:d4562029a3c8e3967e294eb36bda583c9c5fd891}}, we take the architectures searched by SGL-DARTS-1st on CIFAR-10, by SGL-P-DARTS on CIFAR-10, by SGL-P-DARTS on CIFAR-100, and evaluate them on ImageNet. As can be seen, applying our SGL method to DARTS, P-DARTS, and PC-DARTS, the top-1 and top-5 errors are both reduced significantly. This further demonstrates the effectiveness of our method.
| r | 6d2e2b3be56bde2865787bbb7fd621f5 |
The spin exchange interactions, including Heisenberg and DM interactions
{{cite:138ce8b54790d3e0574fa31989c88094d348676c}}, {{cite:df5c51f198734254ea74d2e50d8fe67f40ed5026}}, are calculated using first principles based on combining
magnetic force theorem and linear-response approach {{cite:5b48af401ee03cf7e6f65adc595cdb53a59da78f}}, {{cite:f89bb8ac829e64ee541aaf1ea0f14618ce4962ea}}, {{cite:db3b956e8beb42f8f68120cd9f4279c497a8d391}}, which have successfully applied to various
magnetic materials {{cite:f89bb8ac829e64ee541aaf1ea0f14618ce4962ea}}, {{cite:db3b956e8beb42f8f68120cd9f4279c497a8d391}}, {{cite:c059072e99482c055ef78547cd5b745cc9e1eaea}}, {{cite:18748dbbec695376bf2920cb73c9e896c2765770}}, {{cite:65dd6ff39ee13c219fad4b7fbed584421e0d41f1}}.
| m | bd569c48a72b7ecd878a6414d5feffa3 |
Unfortunately, the converse of this theorem does not hold. A counterexample is given in Chapt. 10.6 in Jonsson's book {{cite:4b5731d619d5d69b0821b5f69a1bbad62bb863e4}}. Nevertheless it turns out that some tools of the topological approach to evasiveness suit as well for a topological approach to {{formula:1282d63e-a683-4d24-a5ef-08e6b2359d6b}} -hardness of {{formula:8c673c1e-8d40-4445-af1d-3d53e7fa59bb}} . The most important ingredient in our proofs is a theorem that goes back to Smith {{cite:3df83e0eaa7232ea632a17b843efb491a29d1d2e}} (see also {{cite:fdf117f17f8f5b680f2debf8a1a80387e7a86a65}} and Chapt. 3 in {{cite:509ee362aa538fbca90c752b67257316f7ab8217}}), intuitively stating that, given a simplicial complex {{formula:2d4b228c-d8fb-4f8a-bad1-b885a353dcd2}} and a {{formula:ae219a21-5c55-4ba8-8bd2-d55398ee31b0}} -power group {{formula:15298dcb-a8d0-4415-a728-ac8118f1da7f}} for some prime {{formula:f61a5d60-855b-4c30-9fde-b312be4f194b}} that operates on the ground set of {{formula:4dfdaaba-c891-4179-8e05-8cbe08bf4863}} in a way that leaves the complex stable, it holds that
{{formula:5360e695-33df-446f-bde5-4abe03b500f1}}
| r | b8b350b9e300ed34bcf54b923d04707d |
The main goal of this paper is to establish the subinvariance of uniform domains in suitable metric spaces with respect to quaisymmetric mappings, freely quasiconformal mappings and quasihyperbloic mappings. We start by recalling some basic definitions.
Through this paper, we always assume that {{formula:10625a8a-0dc9-450d-b8ac-c3e806706c1d}} and {{formula:9509b2d0-b385-44bc-a383-a1b76bd3eabd}} are metric spaces. We follow the notations and terminology of {{cite:d42b50b1d3b6c384f33bdddf2b08f90a926d9ce9}}, {{cite:709567003947d84a02d2bc165f4d8db1f9c7f3ea}}, {{cite:0beabcc20b4ce024e7ae5359c1ef2a284552edd7}}, {{cite:4c741b2f35fb3a291d976330932bc4ed28ba8ac7}}, {{cite:078ae4e17d783ae2e066821338df514d9f2e6965}}.
| i | 82f2caa0b928aecf5426a193e9d8b01d |
The availability of a relatively big parallel
dataset for formality transfer has made it a go-to task. Extensive
research was triggered by {{cite:775d06e95233e21cd7ec7c8e13d0033e2ad091ad}}. They
benchmarked the performance of phrase-based and neural machine
translation with respect to this style. Following their work,
{{cite:7e82bcd9b87ade351538d6d244a89ca198332ebe}} performed style transfer on the
Gyafc corpus as a problem of grammatical error correction.
| m | 57dab05d737363eac4e4740f8dd14bb0 |
We state that bulk {{formula:10800a21-6631-44fc-987b-d732e01e7a2d}} NiO{{formula:ccb65f83-0c42-4bc1-97af-751bf38fdb3e}} samples show a universal spin glass behavior, without any sign of long-range magnetic orders{{cite:dd7a01159755fcbdd7ab349005fd249fa2470c81}}, {{cite:2e3f3564f82a5466a8049d2057f12e79f4154362}}. This is distinct from the parent cuprates which host AFM long-range order.
The fitting parameters from the spin-dynamics models are similar for all {{formula:53649304-ac16-40d9-9642-c0611ec72ed8}} NiO{{formula:f2a1c259-126c-497a-9eec-c4ef9a703492}} as well as Pr{{formula:434cf8db-f8d2-4ec6-a104-342c1a0379e6}} Ni{{formula:eb2f77ea-08e9-4194-a238-845371116066}} O{{formula:867f5c83-6ce7-454e-a7c8-c3128f51b5f7}}{{cite:6720a1bb98b68198224f6f7a5b62638159ac9d29}} (see Table REF ), implying the similar spin-glass nature in these nickelates. According these results, the magnetic entities in {{formula:b9fe6a0e-8298-49dd-9474-dd8e647e9cb1}} NiO{{formula:c6750510-6db5-432b-b528-23199b805ba3}} are expected to be cluster-like, rather than atomic single spins, and the interaction between the entities is not weak.
Within the {{formula:e446d41c-e443-4d02-9257-072344f21291}} NiO{{formula:af0e1d76-e1bc-43ff-83e2-2d529c25e5da}} system itself, we find that the relaxation times {{formula:349bb307-5ebe-45f1-835e-719db1610ffd}} and {{formula:adb57575-1272-4902-aae8-6e6f8de122bd}} become shorter and the energy barrier {{formula:26e43745-f742-43a7-a3e3-cf1bb466ff0c}} smaller from {{formula:091f5686-809f-4adf-a36a-377d1d4820d9}} =La, Pr to Nd. It is interesting to note that while the spin-freezing temperatures {{formula:12330ec6-2a7f-4f8b-a637-e4b7e27e1022}} decrease from La, Pr to Nd, the critical temperatures in the Sr-doped thin-film versions of {{formula:01c56b70-6e3e-4173-9073-faa9abb9db1e}} NiO{{formula:593aff2b-1693-44f5-83f6-3e437476b3d9}} increase (no superconductivity for {{formula:5b4a6f0c-4b93-474b-a499-46fe716fbacd}} =La{{cite:a9b0a1cd3d1fcf0d0aef161725dad892df6edbd2}}, and {{formula:4e5ef75b-490c-4075-8206-bef314bd81c9}} = 7{{formula:1df33c80-c1f1-456f-bc3f-875d60fbbe6d}} 12 K for {{formula:f3126a03-c24c-478f-bcb0-3fe1509a286b}} =Pr{{cite:1100d17159d672014a957e8bc1acab6cebf931dc}}, and {{formula:481f6522-b7cd-43a7-bc18-977a339e5613}} = 9{{formula:42cfacba-255e-4aa5-9b5a-f2a921129154}} 15 K for {{formula:c85b834f-2888-43a6-a29e-64b8c5acd40f}} =Nd{{cite:a9b0a1cd3d1fcf0d0aef161725dad892df6edbd2}}. This has been attributed to a possible essential role of the 4{{formula:fa5a5589-1060-4826-b93e-a60ec763d425}} -electrons of the {{formula:413b800b-f6b2-4fdd-a44e-b530d3b60cc6}} ions, and/or to stronger interactions between the Ni 3{{formula:c62c7c38-ee69-496b-b743-3b5b74c195ab}} orbitals owing to a smaller {{formula:63aec39e-e337-4c62-904b-1b6748c9e6b7}} -axis lattice parameter{{cite:9d8339c624311214f0862cc0402221123dadb7a4}}.
| d | aceb771b6944e0c627c2d570d8eb5320 |
Broken trajectories and occlusions:
In settings where classical keypoint detectors are unreliable, one can either use state-of-the-art pretrained keypoint networks, like SuperPoint {{cite:e9dfc9440d91d8fb810d48de8600040a02da774b}} and LF-Net {{cite:b1b554e44928ff0ef3ba8da51d1e88e8ad7ba2a3}}, or pretrain an unsupervised keypoint discovery network {{cite:891e0132a38ddab0a0379db5b0afe81d71cafd10}}, {{cite:9dc9c0982a2e2a0473dc9227624976fa86f1480d}}, {{cite:3ca83dd28fb2d8dc023cbed0a3c8e5d23ae213eb}}. However, we found show that a simple grid keypoint detector yielded more reliable tracks than using classic (SIFT, ORB, FAST), or modern (SuperPoint, LF-Net) keypoint detectors.
| d | 92cfbecff8d9e18dc5ae6df40da775d3 |
Next-Active-Object Prediction:
Our method is designed to accommodate videos captured from a third-person viewpoint as we need to have a view of the human joints and the surrounding objects. The most related work to ours is the work of Dessalene et al. {{cite:fe3f44a0a95d4e2aee36a6e0498a2ae16199e347}} which is currently limited only to egocentric videos. This does not allow for a comparison with that approach. We compare our method to the recent work of Furnari et al. {{cite:ba9e9693d3116d9a604287bc251f83b7a5a76329}}. This work performs on both egocentric and third-view datasets and is the method that {{cite:fe3f44a0a95d4e2aee36a6e0498a2ae16199e347}} compares with. Their performance is comparable for the task of next-active-object prediction. However, instead of following their experimental scheme and evaluating only the accuracy of the prediction of the next-active-object, we also evaluate the accuracy of the prediction in relation to the time prior to the start of the action where the next-active-object will be used. Predictions are made in the range [2s, 1.75s, 1.5s, 1.25s, 1s, 0.75s, 0.5s, 0.25s] before the beginning of the action (see Fig. REF ). As seen in Table REF our method can correctly predict more objects as we move closer in time while {{cite:ba9e9693d3116d9a604287bc251f83b7a5a76329}} can predict less accurately the objects and is not affected by the time horizon.
By comparing the graph of the partially observed video with those of the reference videos, the pair of graphs that have the smaller graph edit distance and object correspondences between the graphs are estimated (test and reference videos may have different number of objects). The work of Furnari et al. {{cite:ba9e9693d3116d9a604287bc251f83b7a5a76329}} is tested using the CAD120 dataset and the publicly available implementation. We extracted the 1024-dimensional features by using TSN {{cite:c5993b6f9c4e25b75b957198439dabd19c9771b9}} and calculated object features using the ground truth annotations. Their code accommodates the extraction of predictions at different seconds before the beginning of the action as described above.
| r | 5d80a15f0d9fe975e81234191c18c62e |
Some exciting applications that fall under the scope of intelligent code analysis are vulnerability prediction {{cite:511495c82a8e0e5efdce912834faae3a2064b437}}, {{cite:e803882413ac37263c996b2b088f0031a6a5cda0}}, {{cite:733fd3291180aa6418b5aa13f6354842f39cb0e9}}, semantic-based code search {{cite:28f143fb9989dd273c88238b32dc3e45ef80abd2}}, code summarization {{cite:759344251e359833bc3a33c2b68675effe338d9b}}, {{cite:28f143fb9989dd273c88238b32dc3e45ef80abd2}}, code captioning {{cite:759344251e359833bc3a33c2b68675effe338d9b}}, code classification and clone detection {{cite:e345e13cd15f112b73d20e9df632f47cc926557e}}, semantic error identification, code review and completion, synthesis, repairs, documentation and more {{cite:f5efda997dff3f1c67af3c2d7f5e45ec0296f69f}}, {{cite:759344251e359833bc3a33c2b68675effe338d9b}}, {{cite:958486e0a65dc10dc83aa8e4fa97d317754c8fd2}}. These are mostly developed based on Artificial Intellifence (AI) in general and Machine Learning (ML) or Deep Learning (DL) in particular.
| i | f21537e332a49141265e71bfdd7ebf59 |
The first characteristic is the invisible agent. The struggle against
corruption targets `invisible enemies' that deviate budgets from education,
healthcare, etc. – “they are often difficult to discover,
to prove, and to punish. Such crimes are usually committed in secret, by
powerful people, and with some degree of sophistication” {{cite:34aaedc181b6ca1d12e8022169f6142b7ee6b6ec}}. Like other coronaviruses, SARS-CoV-2 has four structural
proteins, known as the S (spike), E (envelope), M (membrane), and N
(nucleocapsid) proteins; the N protein holds the RNA genome, and the S, E,
and M proteins together create the viral envelope. The diameter of
SARS-CoV-2 virion is 50-200 nanometers (from {{formula:ce0bb83a-b2fc-4ec2-bad7-18c0c90551f1}} to {{formula:29653f10-adb4-4c41-9d2a-52b2ca58a378}} {{formula:f4495deb-2139-408c-bb17-bc61f6942c55}} ) {{cite:fef5eefa49ceb4fc14eeb65d1883c6f9965b3ba6}}, and the image at the atomic level of
the spike was obtained by cryogenic electron microscopy {{cite:821343d53c42194622e414f875790150b13ad277}}.
| d | 77d82e68147b11968d163ea960a25df2 |
In these papers {{formula:6c0d1287-4f8c-4750-9856-aa0c86d73530}} was defined by the method of “combinatorial quantization", which yields presentations by generators and relations given in matrix form. Assuming {{formula:815b7db4-1432-401e-bce2-9d6bec6ec0ac}} has one vertex and no edge contractible in {{formula:a8b6f056-41b8-4137-9183-a280b6967584}} , these presentations make {{formula:4605ea3f-afe3-43ef-b91c-fd57e122d8ff}} a natural deformation quantization of the ring of regular functions {{formula:f9bb2828-d281-4298-82a1-e9fbafd43e6c}} endowed with the Fock-Rosly Poisson structure {{cite:62c6fdd96fb704814d323bcdb9971eb90107116c}}, {{cite:d78aeaf09b04824a845d165f111f3708ad68d967}}, where {{formula:11a70a9c-aba5-4138-b4f9-ac6c4072c6b2}} is the number of edges (loops) of {{formula:34835ec7-c871-465f-90f1-4511a722f92d}} . Hence {{formula:1b6378fc-e79f-4005-ad8a-51b6ad038779}} is a deformation quantization of {{formula:a13962c8-e373-4a3a-8ae6-1ad893657d8a}} , the ring of regular functions invariant under the coadjoint action of {{formula:1a64b0b0-b832-44e3-89cb-9ccb798f3e6f}} , ie. the ring of regular functions on the variety {{formula:560980a9-01c3-4f35-9ead-f4ce2fb98fdc}} of characters of representations {{formula:cf2a7335-6a41-49a8-8b4a-26b21d638a8f}} , endowed with the Atiyah-Bott-Goldman Poisson structure.
| i | 6512fd1eb2e9684c346fcab97b93be2f |
The irreversible deformations that we observe happen under persistent stress at slow rate.
They occur at very low stress, significantly below the material yield stress: the typical values obtained experimentally for the local stresses are at most {{formula:a1ffed4e-cd21-4a22-9d12-ecb5b7780957}} .
These small local stress values are obtained despite significant global displacement, a consequence of the slender nature of the filament.
The process is linear and unaffected by time or repeated treatments.
It is therefore likely that material creep is the physical phenomenon at play, hence the mechanical properties will be similar to those for other polymer glasses in terms of frequency, time, and temperature responses {{cite:c56c94af7953a8d048006724e28bc885c56e0280}}.
| d | b6c0cfe5e662049ab2ca802325ffc8d3 |
The PASCAL dataset {{cite:4b72fcc6aaf78e61e9a835ba4290dfd1a77d821d}} is a popular benchmark for dense prediction tasks. As in {{cite:f14335fcf3a23259880567310c42b691b54c3ed1}}, we use the PASCAL-Context split {{cite:09da0c880c63c24ae7ccb0587b2daf210b92680a}} that has annotations for semantic segmentation, human part segmentation and edge detection. In addition, we consider surface normal estimation and saliency detection. For all methods, the initial set consists of fixed 1000 labelled images with five modalities each (normals, segmentation, depth, human segmentation, and saliency). At each AL iteration, Random, RBAL, Coreset and Learning Loss query 300 fully-labelled images and PartAL queries 1500 image / target pairs. Again, in the end, all methods use the same labelling budget.
| r | 88005bb429f6d134f5a6581af6640ef6 |
This should be compared with analogous results for random walks.
Under finite first moment condition, a random walk on a hyperbolic group almost surely sub-linearly tracks geodesics from the basepoint {{formula:66147687-6083-4dcc-9d7e-c27cbd0fa50d}} to the limit point of the random walk in the Gromov boundary.
We refer to Kaimanovich {{cite:1ce91516263d548baa5624a5044a4749da4e21a2}} for a proof, see also {{cite:c003e1f88ae608a9ac92e2dcebaff12ee5c55dd7}} for the case of trees.
Sub-linear tracking of geodesics is one of the most important result in the study of random walks on groups with hyperbolic properties and is related to a celebrated multiplicative ergodic theorem of Furstenberg and Kesten {{cite:93c6baa92c9671aebb406f0c6dd95093e8e71795}}.
It was first coined by Kaimanovich in the context of symmetric spaces {{cite:542fce89f664eea0fe00e933fb0cbd6b6a6d3d7d}} and was then extended to groups with non-positive curvature by Karlsson and Margulis {{cite:825951c85908feb3d4aa2dabd1f96f42e8390768}}
and to more general classes of group by Tiozzo {{cite:d95c2188e71d85c0447b09085f711f7a153f2a5e}}, including mapping class groups.
If the random walk has finite support, then the tracking is in fact logarithmic and this is true for all weakly hyperbolic groups, i.e. groups admitting a non-elementary action by isometries on a Gromov hyperbolic space, see {{cite:a9c849c1793c817c187f79dd29b600802f27fde5}}.
If the group is relatively hyperbolic, then the random walk actually sub-linearly tracks transition points on the word geodesic in the Cayley graphs, see {{cite:969b48fee67f1c5642aa0a61808e3bd2ab9bcc0c}}.
Theorem REF can thus be thought as a generalization of these results to branching random walks on relatively hyperbolic groups and is new, even for hyperbolic groups.
| r | 83c5a5e7866cb2e2ecf6bd4e3434811e |
In recent decades, hadron physicists have expended great effort hunting for evidence of the multiquark states. Since the arguable state of {{formula:ff804d5c-0783-4fb0-9660-e0c512c788d7}} (1540) was proposed as the first observed pentaquark, tremendous progress on experimental and theoretical explorations of the multiquark states has been achieved. Disregarding the large number of XYZ tetraquark candidates, in recent years the LHCb Collaboration has reported and confirmed the observation of three narrow pentaquark-like states: {{formula:61c66c36-d052-4612-a7ee-03d4bdb443e3}} , {{formula:68f5f356-f1ce-4572-aa1a-1bd39a352490}} , and {{formula:349180b6-f2e9-40f7-8e5b-e7c97204032f}} . All three pentaquark-like states may have the quark content of {{formula:0d29b31c-a25c-4057-b376-66355995187c}} {{cite:ca610484367c4f60a2c05e5042c73ed90acc34cc}}, {{cite:6dfffd3528c2f209a6b7acc993ea71dd0162449b}}, {{cite:ff6bf398a42ca4c024518ba28aebd59e4f206b7e}}, but their internal structures as well as quantum numbers are still unclear.
| i | 3dc345ca98eb0b78d79fb901f6a7d415 |
The (scaling) dimension of the operator {{formula:6ebc208e-8402-4a26-a1dd-11d898cc0512}} is thus {{formula:f876e1e3-e03e-49bf-a691-97993d63366d}} , coinciding with the engineering dimension {{formula:e13a90ff-8cd3-4003-a7c3-6252d3043701}} of its top term. The lower powers appear due to the tadpole corrections. At the non-linear level, expanding over these {{formula:5f3643c2-3dd2-4444-b44f-24bfef1a771d}} operators reproduces perturbation theory {{cite:fb58616a4476ab3c2f4de5862ac5a8cadb511a07}}, {{cite:32e79feb8b8e6c117a36d23dc2c9d9a2bdb7787b}}, {{cite:08d9d0029ddd2858284c50dc24a4ed3f9c13e23f}}, {{cite:ff5511553f7a78eb8ba997ceccde4cb4fded3ea8}}, {{cite:1e9522e18a939264c5e0b8d4dc76287a2c997358}}, {{cite:74f50530190cb4685939533cab8fc86039008d8a}}, {{cite:89aa691f1c589326469e6a01e431d78a1f913c47}}, {{cite:7cb95d00d4e47f99d014925b2af37511244be58e}}, {{cite:dd0956657148252acb9bfd690fd76a7860ed81d7}}, {{cite:307309526d5da63c67e0e212efabf9c814de66af}}, {{cite:c83bde3b98a4aabbb4d93378ef1559226bd02c3a}}, {{cite:c2bb68aef4e59a94eec3bd242ee5b22a45510470}}, {{cite:5c16296ebc4d5ee3a14fee51af02d6561932ed15}}.
| i | 68a1f84001daf0182fc8f97a3515026e |
We generally discuss the whole distribution of the FPT and its
asymptotic behaviors. As said earlier, the short-time asymptotic
behavior is determined by “direct trajectories” that go straight
from the starting point to the closest point on the pillar
{{cite:7553a2d7d8a156249aaf7b0ad240b4cbb82a092d}}, {{cite:998cb11a325645ecf2b389a7cc34579a4a631698}}. As a consequence, the left tail of the PDF
is very sensitive to the starting point and to the closest part of the
pillar. In turn, the geometric configuration of the system does not
almost affect this behavior. As earlier discussed for other settings
{{cite:0a2b9c76edb5cdd30f6c22ffc5a7f43efeeae819}}, {{cite:3787e4034880d84f98ef1a2be6d22a21f685bb41}}, {{cite:3b4363da6e4ed176726db46b185d659e5153521f}}, {{cite:7f308e0dd323a0706edce8b7af317b14f7b75310}}, one
generally gets the Lévy-Smirnov type behavior,
{{formula:4cbc405a-9609-4dc4-9123-7aec66684c6a}}
| d | 9e0249d36dc7c2c6f7620ba37606bc25 |
Synthetic adversarial patches: First, we create synthetic robust feature level adversarial patches as in {{cite:f7bbb96854c9bf8c5741947fc70eebf3c61bf489}} by perturbing the latent activations of a BigGAN {{cite:b4855261b54685d37daeba1849d114ef14622160}} generator.
The synthetic adversarial patches were trained to cause any source class image to be misclassified as the target regardless of the insertion location in the source image.
| m | 994fd873870c5a6c747f0b8b033d9b9b |
3D Proposal Recall:
3D proposal generation is evaluated using 3D bounding box recall at a {{formula:55d9e121-1096-4aad-b15a-513e19bd408e}} 3D IoU threshold. We compare three variants of our RPN against the proposal generation algorithms 3DOP {{cite:412144720da765534af64177f8fddb959e1ab855}} and Mono3D {{cite:1a8fcbc71dd777ec7457ed44db59058a71f37504}}. Fig. REF shows the recall vs number of proposals curves for our RPN variants, 3DOP and Mono3D. It can be seen that our RPN variants outperform both 3DOP and Mono3D by a wide margin on all three classes. As an example, our Feature Pyramid based fusion RPN achieves an {{formula:b98d4b1f-d201-402a-a2c6-32647e7d6169}} 3D recall on the car class with just 10 proposals per frame. The maximum recall achieved by 3DOP and Mono3D on the car class is {{formula:553b5b99-8dce-4e5f-81cc-5ba486032621}} and {{formula:2296b018-c863-4a3b-a2f3-23b155581f5e}} respectively. This gap is also present for the pedestrian and cyclist classes, where our RPN achieves more than {{formula:5084fa9d-8dc8-4af9-803b-f3b28ed71b74}} increase in recall at 1024 proposals. This large gap in performance suggests the superiority of learning based approaches over methods based on hand crafted features. For the car class, our RPN variants achieve a {{formula:b12234b7-4522-416a-bb55-4cb3cf0ab0ae}} recall at just 50 proposals, whereas MV3D {{cite:ed477753967ddeb63a508b67a23e58257a022e35}} reported requiring 300 proposals to achieve the same recall. It should be noted that MV3D does not publicly provide proposal results for cars, and was not tested on pedestrians or cyclists.
{{table:7d03cabb-919a-4b10-8cf4-ab7e4cdb03f9}}{{table:f76a0290-56d3-4c95-8a2e-8a639c8f5168}} | r | b51ee96162a2f0e1f6bcbd3e6509f909 |
Based on an analysis of an {{formula:efa594ef-ab59-4965-8291-75d46f59d3d2}} data sample with an integrated luminosity of 2.93 fb{{formula:6bfd67b1-8eef-450d-a4f8-f7bedf219104}} collected at {{formula:d7b563c2-b107-4a00-9105-dfbb0759969c}} GeV
with the BESIII detector, we measure the branching fractions of hadronic {{formula:08193354-14b5-4cb8-ab1f-aeab0100a58f}} meson decays to be:
{{formula:77d1c5fb-1dc6-4aa3-a281-b74ba38bf820}} ,
{{formula:90a3d6f2-7ab9-4677-80f6-f91bfe23cd56}} , and
{{formula:980e9fe5-1980-416e-9ee4-c03999a41d9c}} .
The measured branching fraction of {{formula:448e4b63-eedc-48aa-924b-fd6dbf76c74f}} is consistent with the previous result
measured by CLEO {{cite:82f046431fae888051c4348747b3fa0e84b24f92}}, {{cite:9e491e9e3fa460c4d712519f5f36232c87c8052d}}, but improved with a factor of 4 in precision.
The branching fractions of {{formula:afd03ca5-5e43-4de5-a405-403cd8a715f9}} and
{{formula:9a55c55f-4e4d-4a71-99ad-1370198a7e1b}} are determined for the first time.
| d | 4774a48a10332b42a392d3e15a5af3ae |
Furthermore, we show that self-distillation on this baseline provides an additional boost. Self-distillation is a form of knowledge distillation {{cite:1533dde8803a12aa7997cd7fb310b56f3ff43790}}, where the student and teacher models are identical in architecture and task. We apply self-distillation to the pre-trained network.
| i | 865e137ef4861c5a8a618c7604b3e368 |
4th order quadrature with preconditioner {{cite:ad11ac80f33363733b2f767677fa0fc6279d8a24}}, {{cite:f461dcc5f8801303fa746423c76a597ee5dec38b}}; see the Matlab functions quadrature.m and weights.m in the supplementary material.
Wiener-Hopf method using the sinc-based fast Hilbert transform with no zero padding. In order to counteract the oscillations on the recovered function, we used an exponential filter of order 8 on the final stage of the fixed point algorithm. The maximum number of iterations of the fixed point algorithm is set to 5. In fact, in most cases the final error level is achieved within 3 iterations. We discuss the use of the sinc-based fast Hilbert transform and spectral filtering in Section REF below.
Wiener-Hopf method using the symmetric sign function in the fast Hilbert transform, i.e. with zeros placed at both {{formula:705b5ffb-0e93-41fb-8d60-7e509eaaaa80}} and {{formula:60a4b25a-4dca-48d6-84d5-9adf197a297a}} , similar to the method introduced by {{cite:605f70ef4fb50e8647dd4078b958e0b48cd044cb}} and {{cite:ca0e6de71b4e627bd1ab725f644a64d2769dba21}} and tested by {{cite:2e1d1eb900f7aa8439a09ecec19a910c4c74528f}}.
Wiener-Hopf method with Voronin's variant using the symmetrical sign function for the Hilbert transform.
| r | 0f7f4c55e666a515ba1cd819306f76bd |
A very interesting observation made in {{cite:087fc6232adaf817c929284a75ebf5c3b30ebaa8}} is that the coupling constants of Chiral Theory determine a certain (kinematic) algebra in the light-cone gauge and the product in this algebra is a remnant of the star-product. This statement covers all vertices. For the FDA at hand, it is the {{formula:ad545ac3-120d-4d85-a7d8-46a5385ab268}} -vertex where the star-product structure is manifest. The other vertices correspond to the Chevalley-Eilenberg cocycles of the higher spin algebra. Nevertheless, according to {{cite:087fc6232adaf817c929284a75ebf5c3b30ebaa8}}, what survives of these vertices in the light-cone gauge is the same star-product. It would be interesting to clarify this statement.
| d | 5c92aec1e5ea2cc86a9c441c4c6f653d |
In fact, the likelihood principle is still applicable under a rather mild independence assumption that, at least, approximately endorses some distributional characters upon observed or computed recurrent events occurring along with the time series. For instance, ({{cite:ecf80ea3119a292408fb95746d03aac5ad72045a}}; {{cite:352dd3e29574ae90cb77b545359312390408cf75}}) attempted to estimate the likelihood ratio using KL divergence; ({{cite:dd599ae9cf6e19c6772956ad39ac9b18051e7ecc}}) proposed a graph-based approach and applied it in multivariate non-Euclidean data. ({{cite:1b65761b8134c7e1bc38fdac468a4263a73e1f7c}}) developed an empirical likelihood approach to discover an unknown number of change-points via BIC. ({{cite:6275760f5ef3cd81016a539988f7d071434b21f8}}) present a U-statistic to quantify the difference between the characteristic functions of two segments. ({{cite:d9f728434e2efb375011f5a7e1f958f8f22efa2f}}) generalized Mann-Whitney rank-based statistic to multivariate settings. ({{cite:c4c68c9e2da346b4f2c66531c138b09885af98dc}}) improved the kernel-based method by ({{cite:04013fb5bc8ca65d7214c7229ab899f2aa75b403}}) with a generalized model-selection penalty.
| i | 36a68cee989103b0d6bf98b6e823bec1 |
In this paper we restrict our empirical investigations to IRM and ERM under the settings described above. In this linear setting, we present the ERM results through both the analytical formSimilar to {{cite:8d0281a255d91f3787ba2f36869cee8aa1f0494d}}, we use an off the shelf scikit learn solver for this. and SGD optimization. Our empirical observations confirm that under the mild scale change, SGD optimization are still largely capable of convergence under ERM settings, achieving the same performance as the analytical solution.
| r | 5f5c83fbcd0a015c890978f96e60e27d |
The output scores of MultiXrank can be used in a wide variety of downstream analyses, such as clustering and embedding. For instance, shallow embedding methods need similarity measure for the optimization of the loss function {{cite:64d2ab745164d8af42e497f6124dcc3dab0e38d8}}, {{cite:034f2d7e902c6d01e31f511919544a341ebdaa52}}. MultiXrank can produce such a similarity measure respecting the global topology of the network. Using MultiXrank output scores with subsequent embedding would allow taking advantage of the embedding space on the robustness to increase the predictive power of the gene-disease association prediction task{{cite:a5336c402081ee06fb39e8a70abeb75c7a5ced01}}.
| d | f5aa1bfcd83d0884927ed01af6c747c8 |
Using this formula for {{formula:1c348739-f5f6-40f0-b5a5-07205134c3d6}} together with {{cite:23fee3e2df7e58196255ed91ab04274a07a0dc62}} gives the polynomial
{{formula:b410e99e-5561-4899-b78d-51580357ae4a}}
| m | 5b370b14e15a27c85cd2707cfc153ae5 |
We found that image regions contain information about their spatial position relative to the lens, refining established assumptions about translational invariance {{cite:5602882bd2631bca79bc98ea20498bddd9624df3}}.
Our network has automatically discovered various relevant clues, ranging from subtle lens flaws to photographic priors.
These features are likely to be acquired to some extent by many self-supervised representation learning methods, such as contrastive learning, where cropping is an important form of data augmentation {{cite:ebc9490a661ae288cd2ddbd3862bffef0ff8d5d1}}, {{cite:99d4664cf50b5a33f9b90e2418914e628b42cfe7}}.
Although they are often treated as a bug, there are also compelling cases where the clues could prove to be useful.
For example, we believe that our crop detection and analysis framework has implications for revealing misleading photojournalism.
We also hope that our work inspires further research into how the traces left behind by image cropping, and the altered visual distributions that it gives rise to, can be leveraged in other interesting ways.
| d | a87de69cfe74a155a93bb5ed80c1e0b1 |
KL divergence-based loss. This loss utilises the different ranges of the built time labels in the dataset. Intuitively, an arbitrary pair of predicted built year must have a similar relationship (i.e., the closeness) to the ground-truth built time, which is not as straightforward as computing the distance between two samples, since the built time labels come with different ambiguities. Inspired by t-SNE {{cite:8ef36cf4fd131eb05c5cbfa6113dea6971aeaf48}}, we encode such pairwise relationships into conditional probability {{formula:d0b2caa1-56fe-45e3-9451-84f56be206b3}} of prediction {{formula:dab1f52a-470c-4b27-8cf7-b776b05615ce}} given {{formula:e5272203-7827-4781-86e1-4a767afdd4cb}} (or conditional probability {{formula:27c07245-b956-413e-9430-dad4f3839bc5}} of ground-truth {{formula:3f3b357c-96ce-4785-bd8a-e4f367c57e3c}} given {{formula:98e29ea8-0abe-41aa-b974-9774d311a3d0}} ), where {{formula:9464aa02-7a1d-460c-b852-204cc5dff7e0}} and {{formula:b5046b3e-e9ad-4cd9-911a-2482057cf119}} are in {{formula:b85a0f46-532e-476d-97e5-d6a69c947d75}} . We enforce {{formula:b8e03a4f-2970-4325-a203-f6e7f7650ef1}} being similar to {{formula:82a83c38-f06a-4a42-bb48-566e43158c58}} , so that the pairwise relationships in the ground-truth built year labels are maintained in the predictions.
| m | 8a5a74d2ffa1d502270774e3a2ec80a9 |
As above, we show that the family {{formula:a5047a7b-105c-4a94-baf6-1bc2c4978c49}} for {{formula:e64843ae-e3fb-4053-9280-2ee40ff2e25f}} given in Definition REF is generic in the variety of {{formula:76bed4d3-91c8-440c-b927-88455f8b10ff}} -dimensional commutative nilpotent algebras and inductively give an algorithmic procedure to obtain any {{formula:beec7e5a-1096-4b73-be59-48f1b677481b}} -dimensional nilpotent commutative algebra as a degeneration from {{formula:8d1485b1-6dda-4b5f-bc65-b8ca9096d9ad}} .
The cases of {{formula:a9b47840-4645-4cfa-9217-4011a7488a32}} and {{formula:98eeec41-19d8-4591-a683-61d40e5dc621}} follow from {{cite:cb69507b5609ef1be20e09f87accbcb2f1bf1f5b}}, {{cite:f89527819da6041b6ce826a8e7f8c4ce8d7487ba}}.
| i | ddeb075bc8af85ed1fe14fd4dd095d65 |
The paper is organized as follows. In Section 2, we will first discuss the salient features of the bottom-up magnetised EMD model of {{cite:c4ede254e675679f70ca5ea2105d1f4c04cb28b5}}, {{cite:6e0b2dad11dce517e34804bbfb95496723399f23}}. We will then discuss the free energy, entropy of a quark-antiquark pair, and entropic force in section 3, and compare them with lattice QCD results. Here, we discuss the results for both parallel and perpendicular orientation of the quark-antiquark pair with respect to the magnetic field. The discussion on the dissociation time is presented in Section 4. Finally, we summarize our main results with some discussions and an outlook to future research in Section 5.
| i | 7c0eef2c212fcaaf47fd891bc8ccf721 |
The same applies to AB, with the exception of the inclusion of two dice losses {{cite:3ff566237bcce36e193076dd269db57023ea1c82}} on both transferred outputs for the segmentation task to impose data fidelity:
{{formula:7688d798-9a2c-4193-845c-61406ce70f3e}}
| m | 85cc0523bcd3278025a2669717a835e8 |
In its current form a CAT makes specific assumptions about the conditional dependencies between actions (Section REF ). Following {{cite:e590aafa4e824730465603ae1cd3f0f58b722265}}, a potential future research avenue is to explore the possibility of modelling more complex dependencies. Namely, by contextualizing further the selection of a {{formula:48817653-545e-4d3c-9b89-fa3b788048df}} with with an encoding learned from previous components {{formula:00ba647b-00b2-4d91-8fc5-e82c821d0a82}} .
| d | 5fbdcb4c9ace73bd5c91a7bedea199ce |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.