text stringlengths 54 548k | label stringclasses 4
values | id_ stringlengths 32 32 |
|---|---|---|
We describe the individual operations carried out on clients (data owner) and server (model owner), the two components of any FL system. Our proposed built on top of the FedAvg{{cite:1db59e109f4368fdf85405eeab2bebcbe469fe2d}} algorithm and some of the notations used in this description are {{formula:cb836aa1-e8cd-46a5-835f-7fd8e80254d3}} signify the set of {{formula:6761492d-0750-4553-82a4-cb12eb21a930}} clients, each of which has their own dataset {{formula:db9a86ed-4c0d-4264-868a-5bd078017f39}} Each of them trains a local model using their own dataset and only shares the model gradients to the FL server. Then, the global model formation takes place with all the local model gradient {{formula:3a6ab703-8201-4f7a-8aae-a7c4e2942caa}} updates which can denoted by {{formula:725d8483-903d-48bc-b442-afeb474bf721}} . The complete pseudocode of our method is shown in algorithm and described below:
[h]
Proposed Federated Learning algorithm
[1]
Clients number {{formula:a4e25a47-db43-4f9c-8e21-512007b1ac35}} per iteration, local epoches number {{formula:c861ea9b-151f-4439-8b80-d006a3638615}} , and learning rate {{formula:9f9a7a4f-aafa-4d08-9999-27aec75cc3c6}} , Local minibatch size {{formula:0701be11-8a2f-48d5-85e6-34a0745904a0}} , Total number of iteration {{formula:0b166a7b-acf7-4126-958c-84ada9e1e465}}
Global model {{formula:7033c085-9d89-46b6-ab73-c210c912fd02}} .
[Step 1](Server)
Initialize {{formula:171fc728-22cc-4fbc-8e46-f96664768842}}
[Step 2](Client)
LocalTraining{{formula:4bd3e055-7f3c-4075-87d4-9a69f6936e9a}}
Split local dataset {{formula:a93a4a9d-bb67-43b9-968b-19c52bad3538}} to minibatches of size {{formula:e1750348-f0c2-47c9-947c-20937e8c03d2}} which are included into the set {{formula:f29dd5ab-e86b-490a-bed8-11d6557031b9}} .
each local epoch {{formula:0a7bd75f-c446-4e77-b806-75afd6050a25}} from 1 to {{formula:587843e1-a6af-4326-86c2-d11a2ed5d12a}}
each {{formula:aab74585-d75a-46ed-93ad-f32d1f51f242}}
{{formula:156cf71d-009f-4a91-9716-f43e3e7df1b4}}
({{formula:9543679d-6b43-4272-9642-30e77395a7aa}} is the learning rate, {{formula:8aaf38cc-b34c-49c1-9110-31defa1f5bd8}} is the gradient of {{formula:1dd9f819-88ab-43f5-abfa-5afdd934af0d}} on {{formula:df1b2fbd-8901-4bf3-9b9a-3619086ea664}} .)
| m | 33c880bb1e2d79f7a36e6067e76e8289 |
and the server receives updates uniformly, i.e., {{formula:a86f27f5-46de-48f8-ac1a-e86320984dd0}} .
The above assumption is standard in the analysis of asynchronous methods, specifically in heterogeneous settings {{cite:ff3634bb2a8c11c49946ff35c4ec303a6c8f8d06}}, {{cite:23f6cb5bc58880e111fee24f88d9bcfbac618bde}}, {{cite:a0d8a54e1c50a09dddb2e906d034867c1220e313}}, {{cite:e1ceea13998e2f9edd7ad9615c9b8808bcd89b65}}, {{cite:eca940b8ab3ffff05508e65840603bc4cc7d5472}}, {{cite:29f81f304925415f433ae0c90ed756505e3ec8d8}}. Assumption guarantees that all clients remain active over the course of training. However, they have transient delays and perform updates with staleness.
| r | 8a4e41718a4bc97d24c9d137c3b24ca9 |
These questions are considered in the context of two standard datasets – MNIST and CIFAR-10. For MNIST we use a custom model as shown in table REF , while for CIFAR-10 we used a Wide-ResNet {{cite:ed057a6b7f95b4bcf7a06b5693598046d3057ee4}} with depth and width 10.
| r | e95333660b6c5ee271901380a3e79b6b |
To investigate this, we process the training, validation and test sets of the NSynth {{cite:4c67e000c5e3ac33e002da7d57369ae3e31e4f04}} dataset with audio effects. A state of the art deep learning architecture for instrument classification {{cite:d4f9f1464e73f2872f4e28ae0be44c933a09f6b3}} is then trained with the original training set, and appended with each of the augmented datasets for each effect. We use the model trained with the original training set as a baseline and compare how the models trained with augmented versions perform on the original test and on the augmented versions of it for each effect. The code for the experiments and evaluation is available in a public GitHub repositoryhttps://github.com/aframires/instrument-classifier/.
| m | 94e1ac55f67cfab180c1180bb4fdbb49 |
It worth noting, that few points should be addressed in any future implementations of the proposed SCW QKD.
It is well known that Plug&Play QKD systems are particularly susceptible to Trojan horse attacks on Alice's modulator {{cite:607d3ca960564f51def27d10d24235d48ed31f88}}. To address this issue the signal, before getting to Faraday mirror, can be sampled by 50-50 non-polarizing beam splitter and measured by 2-channel single photon detection module (omitted in experimental scheme for implementation simplicity). The count and coincidence rates have to be monitored to ensure single photon level and exclude Trojan Horse attack.
For deep modulation regime an absence of the strong reference {{cite:39d2db571246dfcaadafd4c81733efa3d53e2204}}, {{cite:57e75b283feef3a0564ecbe5e5ef323889ea368b}}, {{cite:8b4279fd63632b0584c1e7ec7cbaf1ce0396ee8a}} would require a use of decoy state method against photon-number-splitting attack {{cite:125950ce8f9105bdd50870ccaf8e920eab575035}}, {{cite:acbadf6303902591d1f7a2f7334ffa22b33df0b9}}. It can be achieved by an extra amplitude modulation of the reference beam at Alice's site.
The minimization of the losses in Alice's and Bob's devices is an important future task, that could be solved in on-chip fashion {{cite:1b68618a26b3f7b1793c8413dddf1433a154a69c}}, {{cite:20c03c7d68e77a0a53f05118d7f93cfd4ecdaf92}}.
| d | fa71a4287b907fa3981cf2cf214d7757 |
[leftmargin=15pt]
HA (History Average): Mean of historical values of the data.
LR (Linear Regression): Ridge regression on historical values with a regularization of 0.01.
CNN (Convolutional Neural Network): We implement our CNN based on STResNet {{cite:4f0a5a5c4636e1c1a117663334c52598ccff5d0b}}. We use 6 residual blocks with 64 {{formula:c18dfed3-a364-4ae1-9b1e-cd02ae84f475}} filters. Batch normalization (BN) is used.
LSTM: We implement our LSTM based on ST-net {{cite:0aa75b958c844091c35e322f3813dcc8e0f030ba}}, which first uses convolutions to model spatial relations and uses LSTM to capture temporal dependencies. We use 3 residual blocks with BN for convolutions, and use a single-layer LSTM with 256 hidden units.
GCN (Graph Convolutional Network) {{cite:265ee40ad223e261467d232d4af02c3066cb8393}} and GAT (Graph Attention Network) {{cite:edbbc195f69d13e190bd3111df35e5b1faf927fe}}: We apply GCN and GAT with input features being historical values (single-task) or stacked historical values of all tasks (multi-task) of input data. For graph convolution, we take the adjacency matrix as {{formula:f7573e83-ee99-4589-9788-cc9e3243e208}} following {{cite:5ac896494d54c6a80f6470d3833330c6c86a1292}}, and use the normalized adjacency matrix {{formula:941c3374-54a7-44db-8b6f-ee1364a5abe1}} for GCN.
| m | 6b5af92457b7f7247893d6a9e88764dc |
As discussed in Section REF , label leakage happens when the target KC prediction {{formula:5927f391-08f6-42c3-a967-053a501d2107}} depends on the ground truth of the previous KC {{formula:71e51f1f-055a-4eed-8121-fc2566aa1613}} in the expanded KC sequence and at the same time, {{formula:8dd2ff6f-fa1a-473c-a2fd-6839c7c3614d}} and {{formula:e158d614-dd97-4191-82df-8b2c65bdc77b}} belong to the same question. During our journey of building pyKT, we found many publicly available DLKT implementations overlook this leakage issue {{cite:03a0ca24bfec517ed70c454efa715b0cf93a1179}}, {{cite:6eb5b1b04f8ccfe97afc280b9a578a63dabc9b0e}}, {{cite:796e6e5b8c5b47ee402961f3fcb560f45a4a21af}}, which artificially boosts prediction performance. We reproduce such “wrong” evaluation procedure and report the exaggerated results in the left part in Table REF . Furthermore, we explicitly show the exaggerated AUC gains ({{formula:9bdba267-3b80-434a-8b0f-b44038a51e22}} ) in the right part of Table REF by computing the AUC scores difference between results of one-by-one predictions at KC level (values in Table REF ) and results of all-in-one predictions at KC level (middle part in Table REF ). The related accuracy results are reported in Appendix REF . As we can see that, the leakage issue is much worse in AS2009 and AL2005, i.e., the mean values of {{formula:f90346b0-1d0c-4f9c-a7a8-7da50431b5f3}} is 8.54% and 13.23% respectively. This is because questions in AS2009 and AL2005 have more KCs associated and their average KC numbers per question is 1.1968 and 1.3521 compared to the value of 1.0089 and 1.0137 in BD2006 and NIPS34. Experimental results and conclusions impacted by the leakage problem need to be re-validate and we believe this is the reason why we cannot reproduce the results of many DKLT models on AS2009 and AL2005 {{cite:03a0ca24bfec517ed70c454efa715b0cf93a1179}}, {{cite:6eb5b1b04f8ccfe97afc280b9a578a63dabc9b0e}}, {{cite:796e6e5b8c5b47ee402961f3fcb560f45a4a21af}}. In summary, we do not recommend to conduct KT evaluations at KC level in the above one-by-one manner.
| r | 43f5c8576d58295120b10d38e1a19323 |
Step (1): Wikipedia sub-network of trends.
To extract sub-networks that comprise trending Wikipedia articles, we use a graph-based anomaly detection algorithm presented in {{cite:7bdf820a51b6e3188b320c5ff7beee00cf1e9c6f}}. The algorithm extracts a sub-network from the Wikipedia network of pages and hyperlinks. From the initial graph, it keeps the pages that encounter surges of user interest (viewership bursts) over time, as well as their connections. A viewership burst corresponds to a spike in the activity of a page. The desired magnitude of a spike is controlled by a hyperparameter in the algorithm. Besides, the algorithm sets a weight on each hyperlink connection. This weight is increased for pages having a correlated bursty behavior, reflecting their correlation strength. This automatically creates clusters of trending, correlated, pages.
These clusters of densely connected pages within the sub-network are extracted using a community detection algorithm {{cite:005ecb90a999ddd5cc803a34232fa56bbce30860}}. We may assume that every cluster of pages is related to an event or a topic that attracted the interest of Wikipedia visitors at a certain moment in time.
{{figure:ba8720a9-34c7-4d95-bb8c-ef1430919996}} | m | 345f29fb566416408b81cec580415474 |
Reproducing results for benchmarking.
There have been many attempts to establish a benchmark for the offline RL approaches by building datasets {{cite:8e25c9405145d6326f138aee920c6e8974a9adb5}}, {{cite:e836b400a2cd0116aefb7db212d726d0d9f07eff}}, sharing their source code, as well as producing a library focusing on offline RL {{cite:4cc400e5bc1a3bc356cc368e4b73cdf81bc757a0}}. However, we still found some conflicting results between papers.
| d | e7138dacf570ef1c34e63e642e2abc6f |
EOVSA data processing tools are included in the open-source SunCASA package (https://github.com/suncasa/suncasa-src), which is based on the CASA (https://casa.nrao.edu). Codes for processing GOES and SDO/AIA data are included in the SolarSoftWare (SSW) repository (https://www.lmsal.com/solarsoft/). E-CALLISTO processing codes are available at https://www.e-callisto.org. The core wavelet analysis software is provided by Torrence and Compo{{cite:59b9c6a6a7c259278e6ea38921cd28ab76090613}} and available at http://atoc.colorado.edu/research/wavelets ; the open-source code for fitting the background power spectra and determining the significance levels is available at https://idoc.ias.u-psud.fr/MEDOC/wavelets_tc98. The MHD code is accessible at https://princetonuniversity.github.io/Athena-Cversion/.
| m | b6019b23f5dafb51dcfb873c29cc344e |
In this method, the attackers can access the information of the model to generate adversarial examples using the fast gradient method (FGSM) {{cite:ac753e68c02bea67e0f40d51ea41a72c2337a891}}.
Specifically, FGSM uses the gradient of the targeted model to compute adversarial examples. Then, the speech input is manipulated by adding small noises to become the adversarial examples.
The advantage of FGSM is that it can be faster to generate adversarial examples because it is based on one round of iteration to adjust speech features.
| m | 74dd607f11d4d1ea1723bb7c5d8e02d5 |
We demonstrate the effectiveness of ALBEF on various downstream V+L tasks including image-text retrieval, visual question answering, visual reasoning, visual entailment, and weakly-supervised visual grounding.
ALBEF achieves substantial improvements over existing state-of-the-art methods.
On image-text retrieval, it outperforms methods that are pre-trained on orders of magnitude larger datasets (CLIP {{cite:c8358ebfcb8061f188b2914155beff2a86fd4512}} and ALIGN {{cite:39ff5128ce8fc970941638b4167e2a23a43654ab}}).
On VQA and NLVR{{formula:b6cf8f55-8c04-416c-b07d-c6382f84b3c5}} , it achieves absolute improvements of {{formula:05cec9d0-14fc-4c47-baf3-00c5cf0f15b7}} and {{formula:de87d710-f3cc-4d67-b365-e291aa9d4666}} compared to the state-of-the-art method VILLA {{cite:af1652cf953952e3890da2b63c95497688cc857f}}, while enjoying much faster inference speed.
We also provide quantitative and qualitative analysis on ALBEF using Grad-CAM {{cite:3d89ea1a05e35658c28af708ed6e8c7911877bfb}}, which reveals its ability to perform accurate object, attribute and relationship grounding implicitly.
| i | 19a7a141ab1b8fe249b07d2857e00520 |
In Fig. REF , the sum rates achieved by different algorithms are compared versus the UAV's altitude (also known as hovering height in the literature). From the figure, it can be seen that the sum rate first increases when the UAV's altitude gets larger, but the sum rate starts to decrease when the UAV's altitude is large enough. This is due to the fact that increasing the UAV's altitude results in an increase in the channel gains. To illustrate this point, we consider an example, where the UAV and a GU {{formula:cc1bf0c0-6eb9-4587-a718-1dd0379e6d9b}} have the coordinates of {{formula:e7be4f4a-d22f-4a2f-bb81-0a2101d5380a}} and {{formula:b68f4a0b-5a15-4c2f-a467-1c9a1ad8ffd0}} , i.e., the distance is given as {{formula:2f8ccd84-ff2d-4387-b51c-5b2a8e9fb0dc}} . When {{formula:264c0395-0a64-4e6c-aad1-2e69746658e2}} , {{formula:e49853df-efa0-4f2c-ad1a-9250b3958081}} , {{formula:16c29ab4-5a1f-45b4-a059-d2a64e3f64e2}} , and the channel gain is {{formula:eac5d05d-fe1c-43e6-ab33-2a7c5c28142d}} (here {{formula:633b190b-4d2e-4ff2-8bcf-9665ed44511c}} is a constant and the value is computed via Eq. (REF )), and similarly when {{formula:972ce88d-0663-4502-b53b-1bc9adad501a}} , the channel gain is {{formula:af6b9556-6b78-400a-b444-9abb21a4ae3c}} . Obviously, the altitude {{formula:4503bb09-3a79-4bdb-8e11-f1563c4ef23d}} offers a higher channel gain value than {{formula:0bc900ae-deac-4d83-8c55-913ffa11bf7d}} , and the GU's achievable rate increases as a consequence.
However, the channel gain and sum-rate performance become worse if the altitude keeps increasing. Using the same example as above, the channel gains for {{formula:cfcd7334-1a32-4b16-b041-9f2aefd03ce9}} and {{formula:8c3ac067-5b18-4574-a2f4-9c9161f36881}} are {{formula:9ffb8e5b-8f8e-4d3f-85fe-2643aaed7034}} and {{formula:06232bab-0622-4cfa-ac1d-53a6dc2f5806}} , respectively.
Moreover, the OFDMA system is better than two NOMA schemes (GRPA and RandP) at lower values of the UAV altitude. The main reason is that the NOMA schemes typically achieve a better sum rate than OFDMA when the channel conditions of GUs are sufficiently distinctive, but this is not the case when the UAV altitude is small {{cite:837c8e2e7d949cc213332ee54a13c0614b87836f}}. Besides, the figure shows that the proposed HHOPAP algorithm provides the best sum-rate performance in all cases of the altitude.
{{figure:125a6c2c-2d79-48d2-b59d-e9e532d79b16}} | r | f7f6b13409ce23462485f98a5b12c2d8 |
Large-scale quantum computers promise to solve classically intractable problems in areas such as quantum simulation, prime factorization, and others {{cite:f7fe4396744348ae3d47bb2b7c615452126db50f}}, {{cite:d53bd938164ec739dab7785fc3369c107acb4147}}, {{cite:747c61127239e9b532b52584ba3abf29f36ca777}}, {{cite:fc55044d73bdc08f0f5b776ce9e5a87335437ec7}}, {{cite:cfb0d51c239ed4d22d527246caf10de6d05bfd3d}}, {{cite:298447c0207e94f91aecaf42c439e2ddb5cd921b}}, {{cite:302d796a569e4b4f4aa1e22d0ebc6f1d98255468}}. However, these complex quantum computations demand levels of precision that are currently unachievable due to imperfect control and noise in gate operations between physical qubits. In fact, it
is unlikely that analog physical qubit control will ever be able to reach the precision demanded by large scale computations. Quantum error correction (QEC) {{cite:c23b971d8f386923e138d22a939941e6f332456a}}, {{cite:b7f776093cfd47af721b48c22ad1ccbb5b5864e9}}, {{cite:54a5fbeaca69b4b90fdba9e4487620a63b39be95}}, was the key ingredient to digitize quantum operations, making extremely low error rates possible in principle. QEC works by redundantly encoding quantum information into a protected subspace within a larger Hilbert space of many physical qubits. Using a polynomially scaling number of physical qubits, the probability of a computation being corrupted can be suppressed exponentially, making arbitrarily precise quantum computation feasible.
| i | 089e19696092979db5f4fc44079f8371 |
Another frequently used measure to quantify the strength of a gravitational wave signal is the
power spectral density {{formula:b1fda22b-cd12-476b-9084-ef287789861b}} {{cite:230cd3ae2b50e8e06aba68266c776cba252b9e28}}, which has the dimension of {{formula:19a7b47c-a56d-44a0-892b-1e35651e303e}} and is related to
{{formula:65df1ee5-6234-4423-9f7c-46bb25e0b003}} as
{{formula:af1faee6-c963-4814-ab4b-9c7dda3fbba5}}
| d | d74586be55bafda1096462c9f61b7297 |
Completeness The sum of the relevances of features of an input is equal to the prediction {{cite:6164bad06e82abf9cd358d04708743a85a24522e}}.
This axiom stems from the intuition that a prediction of an input is a prediction of a composition of features.
| m | 6bcc188c578d15613a94f28eba06405b |
Implementation Invariance Two neural networks that are functionally equivalent, i.e. give the same output for all possible inputs, should receive the same relevances for every input {{cite:7608ea6c27dc03482315dc4457fe03996ac20a71}}.
This seems trivial but is important to state as methods that work with internal weights of neural models do not necessarily comply.
| m | ba2aadf0c864880262afcaaf709b2d1e |
To test this, we selected the two model parametrizations considered in {{cite:931a627ee5272072c356e59c160605f19caf9e10}}, namely fully connected and full-width convolutional models. In addition, we used shallow (one hidden layer) and deep (three hidden layers) versions of these models, along with linear and nonlinear activation functions. These models were trained on CIFAR-10 dataset {{cite:fba8ea3bbb94cc1eef99f55baedc032006e38294}} (MIT) using PyTorch {{cite:fb2416e6e54e564fcbf577d879cbf225b24cfbaa}} (performance in Supp. Table REF ).
| r | 633c21a4273769adc03930612da388ea |
We have checked that the above result agrees numerically with our previous calculation performed in ref. {{cite:06230fd785c9a4560f9686d16e1b63a6f07580f1}}. This comparison does not include the terms {{formula:b3e8defd-efa1-41e3-813f-9182d87a8ff1}} since those were not computed in ref. {{cite:06230fd785c9a4560f9686d16e1b63a6f07580f1}}. Given the two calculations were performed with almost completely independent methods this represents a strong check on the correctness of eqs. (REF ,REF ). Since one of the main objectives of the present work is to document the calculation of the two-loop amplitude used in ref. {{cite:06230fd785c9a4560f9686d16e1b63a6f07580f1}}, we will next explain in some detail how the two calculational approaches differ from each other.
| r | afa8dbdb3f982af328a163efe7f10b12 |
A huge amount of biologically relevant information is available from cell tracking experiments in embryos. Future research on anomalous transport could investigate other cells beyond hemocytes, later stages of morphogenesis and other types of embryos e.g. zebra fish or murine. Heterogeneous fractional Brownian motion informed by neural networks is a powerful new tool to automatically segment the dynamics of cell motility in time and space. Future studies could link the generalized diffusion coefficients and anomalous exponents with specific molecular processes in the cells {{cite:48022c6389a2c7c41a4a918d281e0a4df6890cd0}}. This would provide an accurate quantitative model to connect the molecular biology with the cell motility over the complete duration of the embryonic stage of development. For example, a current challenge is to explain the anti-correlation of the generalized diffusion coefficients and the anomalous exponents in oscillatory hemocyte motility, which has not previously been observed in the field of anomalous transport. The feed forward neural network used provides state of the art dynamic segmentation of the tracks, but it is possible that other architectures could improve on this performance e.g. the Wavenet convolutional neural network with causality constraints and a wider range of dynamic sensitivity {{cite:b98f11ae9afe8fe916f933589cc25914bb75d9ea}}. Lattice light sheet microscopy could allow the tracking of smaller features inside cells with higher resolution {{cite:8d6d427f180428db046678108117a77642de5e21}}, although it is expected SPIM will still be competitive for the analysis of whole cell motions.
| d | 8cf3a7d14c26ed8ffba1da3fe50c99f8 |
Robust04 contains long documents that would require large amounts of memory due to the quadratic cost of the Transformer models. Thus, during finetuning and inference, it is not possible to directly feed the entire text at once to our models. To address this issue, we use a slightly modified version of the MaxP technique of {{cite:365e287b492fc603ef6a12397c76a550e6f96290}}. We first segment each document into passages by applying a sliding window of 10 sentences with a stride of 5. We then obtain a relevance probability for each passage by classifying it independently. We select the highest probability among these passages as the relevance probability of the document; that is, we do not use the original (BM25) retrieval scores.
| m | 13e6fdb0311ccd46e72f0b6ca64e1a1a |
Deep neural networks (DNNs) bring wireless communications into a new era of deep learning and artificial intelligence.
One of the insightful ideas is end-to-end learning of communications systems {{cite:02f5a1f58b3f8297a99cf19ccdf198e77bcefe04}}, which re-designs the physical layer by employing a neural network instead of multiple independent blocks at the transmitter and the receiver.
Particularly, an autoencoder architecture {{cite:2b31d91901cb35d217f2d3bf62ea9ea69a96a8dc}} is utilized for end-to-end communications, where an encoder neural network (NN) and a decoder NN are respectively utilized in the transmitter and receiver to replace signal processing tasks.
Through jointly training the transmitter NN and the receiver NN, the end-to-end communications system can achieve global optimization and considerable performance improvements {{cite:02f5a1f58b3f8297a99cf19ccdf198e77bcefe04}}.
| i | 7fb94773e38a9c79818f8a3dc0adad36 |
Symmetries exist throughout nature. All the fundamental laws of
physics are built upon the framework of symmetries, from the gauge
groups describing the Standard Model of particle physics, to
Einstein's theories of general and special relativity. Once one
understands the symmetry of a certain system, powerful predictions can
be made. A notable example is that of Gell-Mann's
eightfold-way {{cite:97fa4f5a5bda619b1c573cfa24c8b7e0c086f583}}, built upon the symmetries observed
in hadrons, that led to his prediction of the {{formula:8adf69c6-45f7-4caa-8d1e-39623eb14649}} baryon, which
was subsequently observed 3 years later {{cite:51bcf05ddd45428de527e3846547ba81fa450ac8}}.
The study of symmetries and invariance in deep learning has recently
become a field of interest to the community (see,
e.g., {{cite:e0066ba7bf0aeed8456841d13e75c29337121747}} for a comprehensive overview), and
rapid progress has been made in constructing architectures with group
theoretic structures embedded within. Two fundamental architectures in
machine learning, the convolutional and graph neural networks, are
invariant to the translation and symmetry groups respectively.
| i | 8c654cceb78eb3814bfa055782764f6c |
One limitation of our privacy preserving approach is that we still consume a small privacy budget for each query. Thus, given a fixed privacy budget, subsequently querying our ensemble will eventually use up all the privacy budget available. This problem is addressed by the Private Aggregation of Teacher Ensembles (PATE) approach {{cite:c46d4641f3bbb6b7e1dca1d360e5b3ac155481e8}}, where an ensemble of teacher models is used to annotate incomplete public data and train a student model on the annotated data. In doing so, the consumed privacy budget does not increase once the student model is trained.
In principle, our ensemble strategy could be easily extended to a PATE-like scenario.
| d | c02b68e7a316f301733c012c60209d22 |
In each application, we use the KL-divergence to evaluate the discrepancy between the posteriors from the two encoders in the SOS-DVAE caused by the supervision. There are other commonly-used methods for comparing distribution similarity {{cite:8373b07422e0bd940daf085ab879085f27250561}}, {{cite:a0ddcabd22f5c5eb028c952adf8c2939a9c2b87e}}, {{cite:548429ef1c7eaa3d470b21ad6d546ddce0b68c58}}. However, KL-divergence is chosen for two important reasons. First, the likelihood is nearly ubiquitous in statistics to evaluate model fit, and as this work is motivated by generative modeling the KL-divergence is a natural choice. More importantly, these models are learned to minimize this KL-divergence, and evaluating similarity using the KL-divergence makes it clear that the bias stems from the supervised objective rather than the similarity metric used for training.
| r | b827352e6b9aef2caac854244c1832fc |
The typical situation is when the Hamiltonian of the system is invariant with respect to
the cotangent lift of the action of {{formula:5c8324ba-68f0-4b5d-9e7a-740a5e947d7e}} on {{formula:06627987-90b3-403c-8439-f62d3be86e6f}} . Then the quotient manifold {{formula:2517a3cf-af5c-4405-9350-2b93f1aabfb2}} becomes
the reduced phase space of the system. For example the above happens in rigid body
mechanics {{cite:746fc09ae1ab1671c9314903d40bcb1641ec84e5}} or when one formulates the equation of motion of a classical particle
in a Yang-Mills field: the electromagnetic field is a particular case {{cite:4a2ff80083f0d509b8db2d4b641b5c75547322d3}}, {{cite:6bc94ece71a76c2b29ae197ae16fc2811958255a}}.
| i | 5690a468468fc73dbccf9636e2539df9 |
Bilby {{cite:f7ef3d0905b913ad7a8b890613aa6d968d057fe3}},
PyMultiNest {{cite:1268688a8f513e9e05e1e675903ad4a105271229}},
PyCBC {{cite:a2fe3675191d87216e4e0dd290aa636710f0f6d3}}, {{cite:b25de63f9aeaa7ecd7e3c7d0ad4eec2b4cc0ec58}}
| d | 3636fa0e91e9a11f638533c0be568c7e |
We aim to reduce the domain shift in adverse weather conditions while not acquiring additional real data. Hence, we propose a novel training approach that leverages synthetic data, while making the architecture aware of the weather condition and nighttime. Our architecture is trained on both synthetic and real data simultaneously (see Figure REF ). Our methodology is based on three components: i) adding two simple networks WAS and TAS that work as supervisors to teach the model to learn weather and nighttime specific features; ii) the full-model is trained using multi-task learning where the baseline learn semantic segmentation and WAS and TAS learns to predict weather condition and day-night, respectively; iii) the model is trained on images from synthetic domain {{formula:471d1ba7-fd52-486a-8fa2-24de3d1e711c}} and real domain {{formula:69229410-15a4-4fb3-89a8-5fb6e414ff10}} in alternating fashion to ensure that the model learn to extract adverse weather features only from synthetic data which presents a proxy of the adverse real domain {{formula:73012936-be97-484d-8e65-a2e7b234d711}} . At the same time, it does not overfit to synthetic data and still ensure that the architecture other components leverage real data. Throughout the paper, {{formula:138b9445-c72f-4ce2-ac9d-fcdbcfa86e3e}} , {{formula:b592bfc7-acda-47ae-84aa-8f11c0bb6949}} , and {{formula:c99a60e3-f12c-4703-82c9-3d4c60b09bfa}} are represented by Cityscapes {{cite:90df78495a8d8532d33995c81579eab5dbe79984}}, ACDC {{cite:8a0a3d361c8e4e664b9f614d48e9190d93bccca6}}, and AWSS datasets, respectively.
| m | c642a47fdfced64e5dafaf5fb181125b |
We recognise that the second term in the above expression is
the Hilbert transform {{formula:497361c3-bfaa-4d2c-9567-d42d34fdec7b}} of {{formula:fb3e0d43-f2bd-4ef0-9b3a-a45cd1822acd}} (see e.g.{{cite:7353999232b03531ba78669fdb8e2e24b2d0635f}} or {{cite:7196a89c3a7e02812262d06f54ac7c6d2ba4564e}}).
We recall that the Hilbert transform {{formula:1b61b47d-23ec-4be5-9166-e2358ca774d4}} is a tempered distribution
which maps {{formula:06263bb6-28a0-4053-b558-8156458b3f41}} into itself for {{formula:51735136-aac1-4148-ae03-7f960b26d97a}} ;
moreover it is a convolution-type distribution and consequently
it commutes with derivatives.
This implies that, for a constant {{formula:f668b089-7a98-4687-aab2-e52b74250798}} depending only on {{formula:b71aa4f8-64da-4021-a312-aea6d3a171b0}} ,
there holds
{{formula:2dc5bfd5-3709-4188-ab83-d6d27c242ba4}}
| r | fef7b40cd70fa35eed942082b59bc820 |
In this section, because many images in ImageNet only have classification labels, we use the hidden layer saliency map as the mask of the target object. And we apply metrics of pointing game (pointing) {{cite:3745923336e3f454f6177f3769324efd804cc4d0}}, Spearman's correlation (spearman cor), and structure similarity index (SSMI) {{cite:3398c28f50f25779428666dc43db5a43c429f644}} to evaluate concept classifiers' performances on ImageNet. VGG19 is used for testing.
{{table:63fe47ab-3976-42f8-93f3-619cafaa0cc7}} | r | 604ae5449195ebf32655debd590e9ff6 |
Blockchain-enabled IoT networks utilise authentication and authorisation schemes to limit access only to only authorised nodes. However, these schemes can not ascertain in situ reliability and trustworthiness of authorised nodes, as these schemes do not monitor the behaviour of nodes over the operational period. The case of Mirai Botnet exemplifies how IoT nodes can be compromised post-authentication and become malicious {{cite:827075eb8c47378e93a537d1d59bd4e92a0dbc8a}}, which could severely impede the security and resiliency of the network. As critical infrastructures are being increasingly connected to IoT networks {{cite:488e3d629bcf210900aabe94a8fc842f0b88ead1}}, the presence of malicious adversaries could eventually cause severe detrimental effects, as exemplified in the recent Colonial Pipeline attack, where a ransomware attack halted a major gas pipeline in the US {{cite:63ba83649a3b799139b1ea6b9df19666792c9e44}}.
| i | 5ba0ecc858098306e0b8f5fe90433be2 |
As it is well known, when negative muons, {{formula:1dba8d7d-2331-412f-b1c1-d17ab9a2ebcb}} , produced in a meson factory, slow down in matter, it is possible for them to be captured in atomic orbits. Afterwards, fast electromagnetic cascades bring these muons down to the innermost (1s or 2p) quantum orbits (in this way muonic atoms are produced) {{cite:024d26528eae3c9c11d316fccf957c5d74fbd557}}, {{cite:e22f5bcd71e03a6e5ba836a3447940d018796df9}}, {{cite:0e13baab6d972f780d52f7c4f4b8972ffa69b06c}}, {{cite:7c1e96035006a40a9b48da1452d98cabe4ba6ee4}}, {{cite:a421c38462b74a2bc832af5d1dc9f43c24d21698}}. A bound muon in the muonic atom may disappear either by decay known as muon decay in orbit or by capture by the nucleus the main channel of which is the ordinary muon capture represented by the reaction {{cite:032d0839fb47cd1bb6c94abaac71c2001def4f89}}, {{cite:038f8dd73043f10428cf0941b2f2df8be7fa732c}}, {{cite:aade8e64619c82ae31daae370e9dd58acacfb34d}}
{{formula:9062b96e-ab84-4f7e-91ed-2349423c53af}}
| i | b3752d693f201734ad35b221444a5f18 |
Uncertainties in the percentages of IDPs over classes of disorder (Figures 1 and 2) were estimated through bootstrap resampling and reported as standard errors {{cite:19f2658e9ddf79caf30d8f2ef4aa0dd871ebafea}}.
| m | e1cac15ca5562f7a8c00d1919d87b689 |
The prototypical implementation of the use-case-centric FAIR metrics presented here accounts for 10 out of the 15 FAIR principles {{cite:31a9d98b91fea790935ab473118c8aee1fe5674d}}, as can bee seen in tab:fair.
A checkmark depicts principles that are covered by the overall execution and not by a specific quality criteria.
The main claim of the presented approach consists in the hypothesis that the use case chosen provides a specific meaning to the vague criteria F2 and R1 and therefore makes adherence to these principles measurable.
The implementation is a proof-of-concept to justify this claim.
| d | 6786669d3c3fdc4a41ce683b64a737bd |
Data preparation. We use publicly available high-speed (240 FPS) video dataset, the Need for Speed dataset {{cite:cfc3df9485cecce8e2cfb29dc9b6a9fea6facfc3}}. The reason we choose this dataset is because it has rich motion categories and content (100 videos with 380K frames) which involves both camera and scene/object motion. As introduced in Section REF , our RD is trained on the output of DMR process. As a proof of concept, we simulate solving a single-frame prediction problem, i.e. given two consecutive video frames, we first simulate the latent event frame. Next, a DMR is performed to predict the end frame.
| r | 6d3e2213aeaa81c20f9b3e06bd2253f2 |
To better understand and exhibit effects of various types of noise on WB neuron, three examples of voltage time courses of both deterministic and stochastic WB neurons have well been replotted in [fig:Schematicfigure]Fig.REF B. Deterministic WB model, [WB:isolateneurons]Eq.(), has been successfully applied to describe the dynamics of periodic regular firing for mammalian neuronal excitability in [fig:Schematicfigure]Fig.REF B,Top. With further consideration of input noise to WB neuron, such as sensory noise or external input noise, isolated WB neuron can generate higher frequecy but irregular spikes (in [fig:Schematicfigure]Fig.REF B,mid) when {{formula:572a01ce-e3e2-4189-84c4-c7617217162a}} , where {{formula:2aa27cb5-08aa-4eda-98b1-eb8581cdc76d}} and {{formula:0441b129-655b-4ce1-b329-c439cf295e7e}} are, respectively, the mean and SD of the process and {{formula:55ea675a-1021-47c4-8d37-77114a54583a}} is a Gaussian white-noise variable. More importantly, many previous studies showed that neuronal activity is intrinsically irregular in the generation of action potentials due to channel noise, synaptic noise and network interactions{{cite:190e8ff3027e46d0de71e52c81790093fb7b18a0}}, {{cite:a7a8739b78ec58a462300e34f267a61b1ebb0ee0}}, {{cite:4fc0d95bddec79c856d8eaed7947b6007ef210bd}}. For ease of comparison with the previous irregular firing, example of intrinsically irregular firing induced by channel noise {{formula:32c2db3a-4384-4d41-a399-561116b324fe}} (note that {{formula:c2268676-05fd-434b-b1b9-329ea093d9af}} and {{formula:8657cccc-8979-46f3-bc6a-5420e2bb296e}} are the number of K+ channels and open channels at equilibrium state), has been shown clearly in [fig:Schematicfigure]Fig.REF B,bottom. As shown in [fig:Schematicfigure]Fig.REF B, the former irregular firing impacted by input noise of continuous-time stochastic processes as Langevin equation, however, the later intrinsically stochastic firing affeced by channel noise in a discrete time set-up of specific open channels. The theoretical underlying mechanisms of generating these significant stochastic neural firing, such as dynamical bifucation, have been so deeply investgated{{cite:3aefc5e264ebdfc9700c111943dffeae313df8ca}}, {{cite:11b2f6a5c4aa8dd2b8d59380364aaad64ab2bb5c}}. This study is to find and ascertain the effects of intrinsically stochastic of interest on collective behaviour (e.g coexisting firing patterns) of balance neural networks.
| m | 95e5360cce53b81bdf81f07b7cf4c4da |
The continuous progress in deep learning and the growing amount of image data have enabled new data-driven CE design that joinly design the optical encoder and computational decoder. More precisely, it consists of simulating the COI system as a fully differentiable image formation model that considers the physics-based propagation of light and its interaction with the CEs. In this model, the CE can be represented by learnable parameters and can be interpreted as an optical layer. In the same way, the overall COI system can be interpreted as an optical encoder composed of different optical layers. The optical encoder can be coupled with a DNN or differentiable algorithm in a deep learning model where the ensemble parameters (CEs and DNN parameters) can be optimized in an end-to-end (E2E) manner using a training dataset and a back-propagation algorithm. The optimized CE provides novel data-driven optical designs that outperform the conventional non-data-driven approaches (see Fig. REF for some visual examples or refer to {{cite:4e6a4ca194f6b4361db47701cb7f008647a23f54}}, {{cite:4aabc4536b2167b255efd296dccee51a2d7b7b2a}}, {{cite:9ce2aa2e1c12be173fc72533161af5071e3a5c34}}, {{cite:4f2dc4ddee809e2e288dd6d57f9cb76dbf3682b9}}, {{cite:6218392e05cadd0fdac086c4e6e2b1ba94492f75}} for more details). Interestingly, it has also been shown that a cascade of multiple optical layers can be trained to perform a specific task such as classification {{cite:d91a90d1228e64de45af9927366198c77408e6ed}}, image-formation {{cite:d91a90d1228e64de45af9927366198c77408e6ed}}, 3D object recognition {{cite:341efdc1ae156f44acb5361133f9c7844030263b}}, or saliency segmentation {{cite:08228ebf3dda8e6205e9fa2fbc69b0cff944a6d8}}. This all-optical deep learning framework is referred to as diffractive deep neural networks (D2NNs). Each optical layer represents a CE, and the transmission coefficient at each spatial location is treated as a learnable parameter. Then the optical model can be trained in an E2E manner to perform a specific task. In this context, the optical system also works as an inference algorithm, and the inference can be achieved from pure optics at the speed of light.
| i | 0c9243796ef2710729730902c1151c84 |
Recall that the core of text generation is language modeling.
Given a set of examples each consisting of a sequence of symbols with variable lengths {{formula:c27bb546-7e65-4e25-be9b-ec32e0b4212a}} , language modeling is to estimate the probability distribution of the corpus.
The joint probabilities over symbols are factorized as the product of conditional probabilities{{cite:f18e14dcbd412386bc51340b9fa47074bf84d66b}}:
{{formula:16c45745-b2fb-4026-b17d-44ee3262c84a}}
| m | 08867e7c0f4858c25d9e2ffc6f8c6828 |
In this section, extensive experimental results are presented to evaluate the denoising performance of the proposed GSRC. For the test images, we use two different test datasets for thorough evaluation. One is a test dataset containing 200 natural images from Berkeley segmentation dataset (BSD200) {{cite:76cc3f6e90d1b354d6095c8069c0d002c18e9497}} and the other one contains 12 images which are shown in Fig. REF . We consider two versions of pre-filtering: (1) a pre-filtered image {{formula:09d62600-e09f-4afd-b0b9-098dad58cd6f}} generated by the BM3D method {{cite:7ee78bcca2846e3917c373d728de447853a5e31f}}, denoted as GSRC-BM3D; (2) a pre-filtered image {{formula:91035604-faea-47b8-a1ed-f682160e49dd}} generated by the EPLL method {{cite:6eb6b925405812a81ed7893cec64dc18cfd8f4b9}}, denoted as GSRC-EPLL. To evaluate the quality of denoised image, both PSNR and SSIM {{cite:d25b5ac0a5b06b679fe8473f631bf0528b39542b}} metrics are used.
| r | ad9be940ab5f8c3ab2f473f50e91b2dd |
Traditionally, unitarity of the scattering matrix is implemented at
the integrated level via dispersion relations (see
e.g. Ref. {{cite:40f7e0ca90403ad28fb5ee21160aebedbbe9046a}}). However, for our purposes, it is
much more convenient to use an integrand-level version of unitarity {{cite:8468a360a0a60071c1f8ba1b8ad86194f07d4ac4}}.
This is based on the concept of a generalized-unitarity cut that reduces an integrand
to a sum of products of tree-level amplitudes. For example, for the
{{formula:453665e5-4ab5-41ee-980e-954ed3d78ec5}} -channel cut displayed in Fig. REF (a),
{{formula:fbc494a4-adf9-4426-b65a-0c7dce835ff4}}
| m | d8cbfe98d26a6cd86b08722c8849a705 |
Experimentally, first evidence of the Gurzhi effect has been found for GaAs constrictions.{{cite:a1c7f6300c3505248ebe939d13c31fc4d06999d2}} More recently, indications of a dominating viscous electron flow appeared in other 2D materials such as graphene {{cite:f100dbeae06c61ea7be6ea5ec7422f6e82a3e445}}, {{cite:b91dd1be077f796b1c6532b63a4fed00aa625bea}}, {{cite:12591f732a180c1c5f15b0c530df74d5365ac279}}, {{cite:4bc0e7cb54187c05fbee392c757ec013988d171b}}, {{cite:ad707198f6f7bc2dcb4e4b090c18b20f19a83e42}}, {{cite:78afb610ed2a52c36b215f92a9f58c384ab50eb4}}, {{cite:778f03e65e28362c58660573687f42db4b50d3ff}}, {{cite:2acbf3bc3c9d0e86fa50e07c2e20950f2ca50117}}, {{cite:7be06e5064086d469b3532b8149569d02d155e6e}} or PdCoO{{formula:c7d55f69-6897-463a-9f3e-ecdb4302d7e1}} {{cite:79d74e385d3afa36ffe1999bb3b916a97b84800d}} as well as in the 3D Dirac- and Weyl-type materials PtSn{{formula:df2bd0d5-be4c-4fee-8796-4883b74d6098}}{{cite:3378f0da57778f45b044a67ca366b944568d6e90}} and WP{{formula:ff379fed-b0d6-42bd-9058-c51f3516e1b7}}{{cite:5a87f6d2066192af0c2d89832f7f31749d0e1f02}}, respectively. These encouraging results using electric or heat transport experiments, partly at optical frequencies {{cite:4bc0e7cb54187c05fbee392c757ec013988d171b}}, {{cite:2dd8080be80ee2fff24bbbb7749aeb43c39e6bf9}}, revealed the viscous flow indirectly via its detailed parameter dependence. A real-space visualization has been accomplished for graphene displaying the Poiseuille charge flow profile {{cite:2953cd366e22b4ddce426957e940a05c42ed8892}} up to room temperature {{cite:ccf378f07b39898202ad8061fdc8929454c68988}} and its transition to ohmic {{cite:2a0e045ffa5fd84b06baeabb440dd995dc0f7536}} or ballistic transport profiles {{cite:2953cd366e22b4ddce426957e940a05c42ed8892}}. Moreover, a GaAs constriction in the viscous regime has been probed by scanning gate microscopy without conclusive interpretation of the observed patterns {{cite:2243bf8dfa038be763067aff95419f29f21d9229}}.
| i | 442de226a7ea328443743cdc82d3c0f7 |
We continued using the noise-reduction techniques combining full-volume low-mode averaging {{cite:97403d47fec758ba236906cce4a2a83770939188}}, {{cite:7b99908f511a99350ecfcecf97c53b4b9b4e1336}}, {{cite:5d0cfd49e27edd57f67a500723780a8cf067d674}}, {{cite:240f612c4484bcc9a14ce7b996d346543d477b87}} and all-mode averaging developed by the RBC and UKQCD collaborations {{cite:8935f38d25480fe73d7c0281f1c519300a4d8284}}, {{cite:d8aa57b45f469d688b07f59bd2b5581527b314d8}}, {{cite:f3974bc541f668299ca483d81b44c34c24dad379}}. We omit the specific details as they are the same as in Ref. {{cite:1609364f994f6ddb09a2ef7fd3ec28a40d20db67}}.
| m | 7346f7a17a7fd5d5a8d03f0e7f1184f8 |
The design of interval observers (a particular form of set-valued observers) has been extensively investigated in the literature for various classes of dynamical systems such as linear time-invariant (LTI) systems {{cite:fd809487c033275634924be1c95a1452c6b140c6}}, linear parameter varying (LPV)/quasi-LPV systems {{cite:cb7f49aee9d59e9779ba2e9a28fdaf4ef8ad8eed}}, {{cite:ada9236f0fc5c0cf8b15f1993ffd42e787657315}}, monotone/cooperative dynamics {{cite:d6867c19e2e577190f8be8aba64f39e5c798d707}}, {{cite:78695f74ffb685723e18e6c8ea3688930bd021a0}}, Metzler systems {{cite:b0a8340b649a052da9b9d035a014acdc84e962c0}} and mixed-monotone dynamics {{cite:6130ea55110c6408ac78ecbeeae2546925efd541}}, {{cite:82bfce2ead0fbed48288b106ed6e0711158a5cdb}}, {{cite:f0c3a533c449afde78f7f5492e7f3c2459312ac0}}. To obtain cooperative observer error dynamics, the design of interval observers has either directly relied on monotone systems theory {{cite:62b1e511b3ddbb557b4d6997b6d94496186bfda9}},
or relatively restrictive assumptions about the existence of certain system properties were imposed to guarantee the applicability of the proposed approaches. However, even for linear systems, it is not easy nor guaranteed to synthesize the framer gain to satisfy correctness and stability at the same time {{cite:18ddcf5e01bd0800880a8c82c67014d66d1ae7e3}}.
This difficulty to obtain such properties was relaxed for certain classes of systems, by applying time-invariant/varying state transformations {{cite:6130ea55110c6408ac78ecbeeae2546925efd541}}, {{cite:fd809487c033275634924be1c95a1452c6b140c6}}, transformation to a positive system before designing an observer {{cite:5c5a33ca0ae8d06ab73d1938e8624a70251331c7}} (only applicable to linear systems) or leveraging interval arithmetic or Müller's theorem-based approaches {{cite:0f5e5f7fe85fcd6b09a4bc60feff652eaf498179}}.
| i | cd9fd878a78e0e96081d8e36d44cc16e |
Building on pioneering optical
studies {{cite:e1a7f8fb0abd7e0c56316f7adfb24d4f15bd3953}}, {{cite:730731a22bb20b4a2222cf79922e7a761846c52f}}, {{cite:0eefb6c1da60de594ce24fa77d0084a7d7999a1d}}, there is
at present renewed interest in nonlinear effects in solids
arising from broken symmetries {{cite:0f991951f4d91346c6078920a65080a2a971b3da}}. The low-frequency
(transport) effects of current interest include
unidirectional
magnetoresistance (both field-induced {{cite:d5877c5fe7f4a75efd5695d0711a1a0a33ec9337}} and spontaneous {{cite:6dc7dc47373229e80f9438417128e3e64d3aec2d}}),
and various
nonlinear Hall
effects {{cite:dc1a28356b6034a12edc3d963daa46478066a44d}}, {{cite:238351cea5fe9141e04eb4fd0d96833c0c90c99d}}, {{cite:81889bea9989abf9f152c1748a7b1bf6f79210f0}}, {{cite:9573027a8de1755ffa558181b22ad1567fec17df}}, {{cite:19c6ef10c919c8e66c81060ac231d010f746df9c}}, {{cite:757bc1208ef9134b61a0303e434a04945e7ba21b}}, {{cite:d00dc98458f875781a2fdbbf796123f0b594c1e4}}, {{cite:5c322914d20a298570f65d59b6adceb595416a8a}}, {{cite:1ec189478002cd19b1d9cd65ec80aad79dea03c3}}, {{cite:811d760018dd7a2dd75d033b134f1a2cded404a9}}.
However, the basic question of how to extend the Hall vs Ohmic
decomposition of Eqs. (REF ) and (REF ) to the nonlinear regime has
received little attention. With the present work, we aim to fill in
this gap by providing a sharp answer to that question.
| i | 320b4340059da531f4cb5a18dd4fcc81 |
When unrestricted TGDs are used as an ontology language,
ontology-mediated querying is undecidable even for unary queries that
consist of a single atom {{cite:1681b269d9e8c8bc5ea630c0d9edb6a5a6a07552}}. This has led to intense
research on identifying restricted forms of TGDs that regain
decidability, see {{cite:b6bd34360c4d0963c433c243f0bea5e91fcd3f03}}, {{cite:1681b269d9e8c8bc5ea630c0d9edb6a5a6a07552}}, {{cite:76839cda2dc67dc3a1be4c4883102145d9ac4680}}, {{cite:4999698c66bc667e265446659f6c1d24aa9cdd8c}} and references
therein. In this paper, we consider guardedness as a basic and robust
such restriction: a TGD is guarded if some body atom, the guard,
contains all body variables {{cite:1681b269d9e8c8bc5ea630c0d9edb6a5a6a07552}}. Guarded TGDs are useful also for formalizing integrity constraints.
For example, the important class of referential integrity constraints
(also known as inclusion dependencies) is a special case of guarded
TGDs.
While being decidable, both ontology-mediated querying and querying
under constraints with guarded TGDs is computationally
intractable. Let us make this precise for query evaluation, i.e. the
problem to decide, given a database, a query, and a candidate answer,
whether the candidate is indeed an answer. We use
{{formula:17c8970b-c1ad-4ae7-b66b-ea5b9116cd2d}} to denote the language of ontologymediated
queries {{formula:9a0c1223-ab82-4514-a965-c1c9b5860d60}} that consist of an ontology which is a
set of guarded TGDs, a data schema , and a conjunctive query (CQ)
{{formula:2660847a-ad63-4d96-98fa-5261657ad69f}} . As usual, contains the relation names that can be used in
the data while both the ontology and query can also use additional
names. Evaluating ontology-mediated queries (OMQs) from
{{formula:c1f99730-bec4-4c54-9ec5-143d0a042875}} is -complete in combined
complexity. The same holds for {{formula:6a26ee5d-bc4a-4882-9a8f-a7921e9bcc76}} where the
queries are unions of CQs (UCQs) {{cite:1681b269d9e8c8bc5ea630c0d9edb6a5a6a07552}}. For querying under
constraints, we consider constraint query specifications (CQSs) of the
form {{formula:9b9c7b2b-0d49-48e4-a638-da815bfa4ff7}} where is a set of integrity constraints and
{{formula:1d09e2e7-2d6e-4a46-bc18-644a1605fcb4}} is a query, both over schema . Overloading notation, we use
{{formula:2f68e57d-9dce-4e72-9e42-dca07e435cc7}} also to denote the class of CQSs in which
the constraints are guarded TGDs and the queries are (U)CQs; it will
always be clear from the context whether {{formula:494e9d7e-7916-4386-842b-f8dd8a3b236d}}
denotes an OMQ language or a class of CQSs. Query evaluation
for CQSs from {{formula:76d5dadb-905f-46d1-a65e-270c0af14c7e}} and {{formula:548f8d5a-620f-4361-9da3-9a90cd191604}} is
-complete.
| i | 32492e0569f488ee3c02083498d525ee |
In recent years, Deep Learning has gained popularity in computer vision and pattern recognition primarily due to Deep Convolution Neural Network (DCNN) {{cite:c3c450f689f05c3757a86533cf9be697e32b9af6}}.
DCNN includes convolutional layers, fully connected layers, pooling layers for feature extraction, image representation, and dimensionality reduction respectively. DCNN also includes sigmoid or rectified linear units for non-linearity. CNN is based on local receptive fields and shared weights {{cite:34277e8e53bcabda17b80fc4ef58fea052d2e376}}. The local receptive fields are regions that provide local information and are spread over the image. Each local receptive field constitutes a neuron and shares weights in the hidden layer. This helps to get distinct feature characteristics. Pooling layers that follow convolution layers, down samples and assist in overcoming overfitting by reducing the number of parameters.
Additionally, dropout, wherein hidden neurons are randomly removed, also helps to avoid overfitting. Pre-trained CNN models are trained on {{cite:30dbf3af6e306133fd781bc025e5db46274a9343}} that carries a large number of data samples and 1000 class labels. The transfer learning techniques can help to apply the existing CNN models to problems where a large dataset is not available.
| i | c72317788dd811096ad73341812481ac |
Here the performance of our algorithm will be shown with five examples. The first example is an artificial reaction mechanism with two reactions {{cite:574a27f4aa10a299f064f06bfc24cbbfce8434e4}}. The second one is the well-known Michaelis-Menten kinetics {{cite:caa47ed74e0d53e6b238b1ace7127dd276f731fb}} in biochemistry. The third one is the hydrogen oxidation reactions {{cite:bcf5a2ba40511501754fae7ce6a9bd8f51000e21}}, {{cite:47cd6510e57eea7314322dc2e4411659a2f3f2d6}}. The fourth one is the extended Zeldovich mechanism, a typical chemical mechanism describing the oxidation of nitrogen and NOx formation {{cite:395a38f98a20f10234fe8ea0ba17ad387cc52ab9}}. The last one is the simplified GRI-3.0 mechanism, a chemical mechanism describing the methane oxidation {{cite:377eef5a8591aaf43170af7da882a2366875bb8d}}.
| r | 9b91dc54e7d3727375d5a6b23cdc95b3 |
Concerning the expected asymmetric line profiles unique to MHD-driving process, there have been some observational hints of the expected asymmetric Fe K UFO features in NuSTAR and XMM-Newton data (e.g. a nearby Seyfert 2 AGN, MCG-03-58-007 in {{cite:2d522f0967b0895ffc052e0f96a8e68fe3aaa140}}; a lensed broad absorption line quasar, APM 08279+5255 in {{cite:a7bdfeb61575b6fed2b986a6a32cce9de9521f1c}}; a complex Fe K trough feature seen in narrow-line Seyfert, 1H 0707-495 in {{cite:525d97ff049f664ace06e41cdc7d62cbe9e54c1b}}, among others).
Also, it might be relevant for other types of disk-winds such as those of BH XRBs occasionally detected, for example, with Chandra, in Galactic transients {{cite:3a9d144c53d55c927fc47ed4a7ff4536214479cc}}, {{cite:1ae3f841545583a7aa61d542d124b36957683b0b}} as well as AGN warm absorbers {{cite:adeec13d008e7866367f00ab6506be91cb96ea5f}}, {{cite:4e31fd968c16711b3947b527e619d25778e2ebb2}}, {{cite:256741e3c0d62ccf553986f73b21911a29a60d04}}. The observed velocities of the ionized outflows in these systems are much lower than those in UFOs; i.e. a few hundred to at most a thousand km s{{formula:1f0e7f73-af6a-4027-bc1d-8d69705cb5da}} . The expected blueshift is thus only up to {{formula:54eb3892-620f-4c06-9036-a9dc5a5c126c}}{{formula:50c4e2f7-4631-4793-9852-0b8b7e636203}} 0.01{{formula:2717dae3-953f-4a8b-8523-35c3cecaccbc}} E/E {{formula:07f13153-5986-4498-afb1-614b5bf8dea4}}{{formula:9767d4b2-d29a-42c9-9a85-a0f26b482b14}} 0.001{{formula:e6e45827-8d24-4e31-8d92-c591bdf8414e}} In this work, we do not incorporate emission lines also expected to originate from the same outflowing gas or separate regions. For example, Monte Carlo approach has been often employed to calculate self-consistent emission and scattering components (e.g. {{cite:a59f68b00f9b19da17694638068f4083b3a2925f}}, {{cite:65f1760ca06fa41a69ebb8dde6c44aa5b706f273}}, {{cite:525d97ff049f664ace06e41cdc7d62cbe9e54c1b}} with MONACO and {{cite:0fdb6648d5088d24d74407353e780ba8105a5a2d}}, {{cite:009839ea7cadd702201ce564fe8aeca984a80998}}, {{cite:6cd6cdba1b7121b134894718e3ac8fa05389fec2}} also including special relativistic effects), while XCORT has utilized XSTAR runs {{cite:d90a20a4ddd1b24b3604bb4a12d94ea068811197}}. More recently, {{cite:05cc0305a9db2a947f61723dd5b6026acfa7d1f4}} has proposed WINE model with a special focus on detailed special relativistic dimming for UFOs {{cite:d336843af21a7a982a52a3291422617c6dc9979c}}, {{cite:4631ec0c76ae277b51bfb017e33dfb84ff2e26b1}}. Such emission process might be important to accurately assess the physical properties of ionized absorbers because they can “fill in" the intrinsic absorption features. For example, some very luminous quasars accreting at near-Eddington rate such as PDS 456 is found to show a broad P-Cygni profile in Fe K band associated with its prototype UFO where the wind may well be Compton-thick {{cite:6e114ae6a1f7de78a38cd89585c4c8b7cf189d7e}}. On the other hand, one of the well-studied Seyfert 1 AGN showing canonical X-ray warm absorbers, NGC 3783, is also known to exhibit a handful emission lines in UV/soft X-ray band mostly from the He-like triplets (resonance, intercombination, and forbidden lines) of O vii, Ne ix, and Mg xi as well as Ly{{formula:c2d6c408-bfdc-4f4b-9bbf-7a49eca8fec0}} lines from the H-like species of these elements with a very small systematic velocity shift ({{formula:8ed09a9e-6059-48e2-9300-f33afb90d397}} km s{{formula:3271f68f-a28d-4af1-9d4c-c132c07ec851}} ; {{cite:adeec13d008e7866367f00ab6506be91cb96ea5f}}, {{cite:15b588cd567502c6d7109cae630ffd796f6e29b6}}, {{cite:4e31fd968c16711b3947b527e619d25778e2ebb2}}). However, given a moderate accretion rate in Seyfert 1 population with typically very weak (if not none) P-Cygni profiles in X-ray spectra, it is reasonable to speculate that X-ray absorbing gas is most likely optically thin. Scattering within the absorbers might then be less efficient.
| d | 06a8cdf977b49d923ed027bc8c855f36 |
Recent advances in deep learning have rapidly advanced the state-of-the-art object detection algorithms. The best mean average precision score on the popular COCO {{cite:fa15dbc50be239aa07ce96e509be919ffeed2c97}} benchmark has improved from 40 mAP to over 60 mAP in less than 4 years. However, this success required large datasets with annotations at the bounding box level and was a achieved in a closed-world setting, where the number of classes is assumed to be fixed. The closed-world setting restricts the object detector to only discover known annotated objects and annotating all possible objects in the world is infeasible due to high labeling costs. Therefore, research of open-world detectors, which can also discover unmarked objects, has recently come into focus {{cite:dcf380d148fd7d811254bde78aa2c30ab292c895}}, {{cite:bbea9fb8beaadeeda28b5fdfdedbf121ea2a6b09}}, {{cite:8332181f6e3d91dbff57bbd7bfd7ce04452c5738}}.
{{figure:003189ea-7d6a-4451-95d0-13b28977747f}} | i | 5249df410393eb4c25cfa93d6c8a4f2c |
Figure REF shows the inclusive and differential measurements performed for {{formula:15f4b116-7e65-4818-bc09-f9a3817ea2a5}} and {{formula:0e132aa9-a0b6-4b46-b0e7-54e8149984ea}} at parton level in the full phase space. The total uncertainty on the measurements is shown. The main source of uncertainty on the different measurements is the statistical uncertainty, followed by the signal modeling uncertainty. The measurements that involve the reconstruction of the {{formula:e0df600e-433a-4ce3-ab25-90ebe5cd2f1c}} system are also affected by a reconstruction uncertainty which is approximately half of the size of the statistical uncertainty. The uncertainties corresponding to the detector and background modeling do not contribute significantly to the total uncertainty. The results are compatible with the SM predictions {{cite:3c60b118c528360fa2130c431607b6036fa5e9a0}}. A similar behavior is observed on the uncertainties in the measurements performed in the fiducial region, however, there is a reduction in the modeling uncertainties. Figure REF shows the unfolded distribution for the {{formula:d771a801-b8df-47a5-920d-03aece80a0a1}} and {{formula:fda5517d-eff3-4916-87cb-75ddf3ae87be}} observables in the fiducial region. The distribution is in agreement with SM predictions. Figure REF shows the {{formula:8cfefcda-85cc-4aff-9509-b464e675a439}} and {{formula:9c977203-0d37-4959-a44b-65a996fb2e48}} measurements in comparison with several models beyond the SM {{cite:7052c0f36a21b80357d69b663e10bb40e4212878}} in the full phase space. In these models, the values of the asymmetry are expected to be different from the SM expectations. The ellipses correspond to the {{formula:96a8f999-59ba-449e-859e-89c7398199ca}} and {{formula:bd77c742-1b1c-4d51-9bbe-35020aaa428a}} total uncertainty on the measurements. The correlation between {{formula:ef095c24-6ba5-4603-8b44-6900761818a7}} and {{formula:28cc4f32-2bd7-4605-b65d-6c12e515b3fd}} is about 48%. The measurements are compatible with the SM and do not exclude the two sets of BSM models considered.
{{figure:b7dc3831-78f1-4086-8411-08ea78a7f8b4}}{{figure:6e4138ce-730d-4774-b448-a431a986df58}}{{figure:9bb276a5-e8e5-4157-88a6-15d7a700ccd9}} | r | 5ead08f5504e4f2be156ed1a93a19641 |
We follow the evaluation protocol of the study {{cite:3693ee50f8b9c3937de2798c748ff82c9ddf30f9}} by using CUB {{cite:d817b63d8db8e1b94782df9320778ef93ebf29cb}}, mini-ImageNet {{cite:57c6c53416acd7eb605bdc1727d038f543f8477b}}, and tiered-ImageNet {{cite:fda2aa7d17928b6405aa5cfe50a49dc23c92d0f9}} to evaluate the proposed method. The mini-ImageNet dataset comprises 64 classes for training, 16 classes for validation, and 20 classes for testing, and each class has 600 images. Meanwhile, the CUB dataset comprises 11,788 images, among which 100 classes were for training, 50 classes for validation, and 50 classes for testing. The tiered-IamgeNet dataset uses 351 classes from 20 categories as the training set, 97 classes from 6 different categories as the validation set, and 160 classes from 8 different categories as the testing set.
| r | 32f86b5f92d579560f7007ebc8d233cb |
Many methods attempt to model the class-dependent asymmetric noise, such as with a label transition matrix {{cite:4d0d3d059003401530c49d7fa2092768df7922ef}}, by learning
noise-adaption layers and performing loss-correction {{cite:02404ddf89f12dc3f40a7a5f2282b156843325bd}}, {{cite:b9d6b48c0f30caa1f1cd1dca0373338e0bcecd38}},
or by using reconstruction error as a consistency objective {{cite:4a967a34e70965f6acfe7349238a9892ee555649}}. However, these methods are not as competitive as SSL approaches in Sec. REF because they generally do not address instance-dependent noise and have
limited mechanisms to make use of mislabeled samples.
Methods that attempt to handle semantic noise by estimating instance-based transition matrices can in principle deal with semantic noise {{cite:02404ddf89f12dc3f40a7a5f2282b156843325bd}} and can provide guarantees on convergence and bounds on generalization error {{cite:49c6c7db8619a0af351f976c2794311125c41539}}, but they do not provide SOTA results in practice.
| m | 40623b6ca8c22e1744473db4002f7e80 |
Quantum field theories from the same class as QCD are now experiencing dramatic changes and rapid advances in a deeper understanding of anomalies. I want to mention two crucial papers: {{cite:d464d6dce384dfe5fbce43dc688a56b4f8344694}} and {{cite:ba27b461166d4a5f804f31b3766b32bd12f933a4}}. The latter demonstrates that at {{formula:79f6a8f6-e561-461e-93ef-f966fb07cbc2}} there is a discrete 't Hooft anomaly involving time reversal and the center symmetry. It follows that at {{formula:b0b12c95-f9fc-47d5-b46c-bebe3fdae4e0}} the vacuum cannot be a trivial non-degenerate gapped state.
| m | d8f8a7f12f4a8882c89e19fd3496db78 |
Overall Performance: Table REF showcases the various model performances across 10,000 TSP instances. We rerun all models again on our machine to get updated results for a fair comparison. Evidently, moving away from a learnable decoder reduces the model's predictive power; the self-attention transformer is highly capable of encoding partial tours that improve the overall performance. However, we also see that the inference times in our approach has improved tremendously. At the same time, we can achieve better performance than the previous state-of-the-art in {{cite:dbb250efa12c2549a0941ae2395237272243f98f}}. Also, since we can run beam search much faster, this suggests that we can increase the total number of beams, improving the search space and solution quality. With the addition of the optimal transport bias into the decoder, we see a 3-fold improvement in optimality gap, with minimal impact on inference time on TSP50. Likewise, experiments on TSP100 produce similar observations; the optimal transport bias is beneficial to the model and is light.
| r | d520090690d0ac18488ea3a9e189926c |
Several studies, such as {{cite:bbcd1058d564f5764af329f14d5a76bff0069f06}}, {{cite:9bd1c6d0a68764bb2914356c4e999b6d7783b090}}, {{cite:d1b7b9ce649e7646b0e765adb007142ebe81e299}}, have reviewed various oversampling approaches; nevertheless, they are not thorough and have not paid adequate attention to validating the oversampling approach to the problem of class imbalance.
| m | 305253affd4e8a65acb0a78a210bec52 |
Remark 2.1 By (REF ), it is clear that both {{formula:3aa0977d-a754-454c-9b7b-422b777d0754}} and {{formula:f72a5a23-a051-483b-a875-cf679ecfe1f4}} are surface-localized according to Definition REF . In our subsequent proof of Theorem REF , it can be seen that the transmission eigenmode corresponding to higher mode number is more localized around the surface. In other words, in (REF ), {{formula:06df0215-85ae-44c2-8f24-e07df82e7a38}} can be very close to 1 provided {{formula:b6190689-7059-43f2-a937-070ab010a001}} is sufficiently large. It is interesting to point out that {{formula:4f9682a4-e520-4178-83b8-8cf6a62de1da}} is actually the localizing radius of the eigenmode, which defines the super-resolution power of the wave imaging scheme proposed in {{cite:945c502def3c4ddb0d1c05817e563b0ffaaba202}}.
| r | 76d3cd732a94a4ce1ba12972f91d8f55 |
At the algorithm level, existing end-to-end methods {{cite:daff0be54bdd8e3dccf6bd13c471d193d3bf4e2f}}, {{cite:294c3dcf1567eae8e0257690bf550be0d401a32c}}, {{cite:ba42fcec8b38716507dc2d26fb8156361da49261}} score each pair of mentions only based on mention representations from the output layer of a contextualization model. This means that the model lacks the connection between mentions and their contexts. Semantic matching operations between two mentions (and their contexts) are performed only at the output layer and are relatively superficial. Therefore it is hard for their models to capture all the lexical, semantic and syntactic cues in the context.
| i | 4d3bb412e3656225bb90c9d912ed730d |
We now compare the commonly used motion-forecasting datasets, i.e., nuScenes , Waymo-Open-Motion-Dataset (WOMD) {{cite:487e5d57d98d2b65856974d3899810d54886cf6c}} and Argoverse {{cite:0fc6e0679853603ef93b88f630d536a29e3df1ee}}. We individually discuss why Argoverse is best positioned to bring out the benefits of our proposed work.
| d | 4389840e860c920cd39414e9bbe4a66a |
We note that the non-collinearity of spins by itself does not guarantee strong magnetoelectric effect and electromagnon peaks – the crystal structure is equally important. Thus, in the layered Kagomé antiferromagnet, the iron jarosite KFe{{formula:a42993ff-0a05-4776-8fe1-d7b5e4da4fb7}} (OH){{formula:0b9f1d7d-40cd-4dfb-967d-5467d6f2b7d3}} (SO{{formula:1216dc69-53b6-4147-8c31-b295a27859d6}} ){{formula:0682fa2f-eb27-464f-9228-b83997e07f19}} ,{{cite:4d6645b582491907e9f3282259623ca44424042e}} which has the spin ordering shown in Fig. REF , the ligand ions are located outside of both up- and down-triangles, which cancels the magnetoelectric effect due to the Heisenberg exchange striction. The cancellation also occurs in triangular magnets with the {{formula:4f6c6bc7-fbed-4a6a-a460-daa23d399c5f}} spin ordering, as they contain three different spin triangles, such that spins in one triangle are rotated by {{formula:df74c22f-31c6-4dfb-be31-57be6dc55c90}} with respect to spins in two other triangles{{cite:5414fe2153ceade9561f1767770fcc13ed5bec1e}} (more generally, the linear magnetoelectric effect can only be induced by a spin ordering with zero wave vector). We note, however, that the lattice trimerization in hexagonal manganites{{cite:e0fee4bdeba549d87cd0e49e4e38a5de54d90aec}} makes the three types of spin triangles inequivalent and destroys the cancellation. This can be also seen from the symmetry properties of the A{{formula:158a4734-8c76-4a50-8794-e448b13426f4}} and B{{formula:2c1d947e-ef15-45e1-b035-6061b403fd2c}} phases of hexagonal manganites{{cite:86765497aa0007e79f91df8a44329981786a0bff}} allowing for the magnetoelectric term {{formula:47b92e58-0a1b-405d-ae94-84f15fef83f6}} in the A{{formula:2ec73c7f-9749-432a-8a6e-1a38dbd19746}} -phase, which has a toroidal moment, and the term {{formula:f5f393a4-6f3f-421d-89ef-f3ca1231e0d8}} in the A{{formula:a7e3e17a-ead4-4faa-a3fd-927b25a1e00b}} -phase, which has a magnetic monopole moment. Whether electromagnons in these phases can be observed, depends on the magnitude of the trimerization and remains to be explored.{{cite:4eba7d7c253d6a1b61cfb9beb8d7eebc2ef3e202}}
| d | 51a82134c0ea401e2bde56043909c365 |
In this work, we establish the thermodynamics for growing systems.
The difficulty in developing it lies in that the change in the volume affects all reactions in it.
In the conventional theory of chemical reactions, reaction fluxes are described as functions of densities of chemicals (concentrations) {{cite:e1a3a8b3f7a9b14a83c8f3cf486610e60d2cd591}}, {{cite:703519e1473c714b42552aa90671df71566c2afa}}, {{cite:f1c1d22335563f8e06556e5657fc8dbbc023ca36}}, {{cite:66d7280807c3d687aab1dda48b519a4d182d80aa}}, {{cite:1425079e73583697465b16f5ab1f15fb435e2f1a}}, {{cite:c72cb15f9ac88607b45d4c0e837706a977078469}}, {{cite:474cb47f91abaa4f15f83ccaa9927a0b5ccf3236}}, {{cite:dc6ee1a7ce6b87f543f0b2247e76e5aa956a4b84}}, which presumes a constant volume.
However, if the volume changes, the densities can change even though the numbers of chemicals remain unchanged.
Hence, it is necessary to return to a thermodynamic formulation in which the numbers of chemicals and the volume are treated separately.
In other words, we have to explicitly account of the extensivity of thermodynamic functions, which is scaled out when the densities alone are considered.
Nevertheless, we should also retain the density representation and its dual representation by the chemical potentials to appropriately characterize steady growing states and the conditions imposed by the intensive variables of the environment.
| i | 61540d1f15e974f8ec47dab5442b5b1c |
Among other things, Luo {{cite:4af449d8dfe200011a7a539de012d7078006f1ba}} further established a variational principle for vertex scaling of PL metrics on surfaces
and proved that there
exists some combinatorial obstructions for the existence of PL metric with constant discrete Gaussian curvature {{formula:9d82922e-f503-43f4-8b78-469a074b8259}}
in a discrete conformal class on a triangulated surface {{formula:32ecd309-6473-4dfd-a4dd-515133716540}} in the sense of Definition REF .
This implies that there exist some combinatorial obstructions for the solvability of the prescribing discrete Gaussian curvature problem
in a discrete conformal class on a triangulated surface {{formula:8dca1859-9edf-42c3-baec-7ef6af224fd5}} in the sense of Definition REF .
To overcome this difficulty, based on Bobenko-Pinkall-Springborn's important work {{cite:5984e50a776ccd924435293ca8f2106a40149727}},
Gu-Luo-Sun-Wu {{cite:5542b20424d7a2e10d4bbc6613f65509a30ba6ad}} introduced the following new definition of discrete conformality of PL metrics on marked surfaces,
which allows the triangulation of the marked surface to be changed under the Delaunay condition.
| r | f9e4437e8a5025cd1551312c6784bb7c |
where {{formula:067861ee-a923-4627-ab3f-e761aea233e7}} , and {{formula:b0705486-20fd-45c9-8cae-e6d0a740a3b8}} indicates that {{formula:2dfc84a6-075b-46fc-8547-6a9ad4610116}} is an {{formula:df6d5f1c-546f-4e02-b90a-ebdfc06e7290}} -tuple with {{formula:08a04a5d-4ec3-4ae6-947f-fb5c6a42e9a1}} . Given {{formula:a1d45da3-b416-4205-b212-ea3d8ebbacc2}} and {{formula:d2bbcf87-75a6-440c-aa21-f799afdc49eb}} , overlap-aware diarization seeks to compute an optimal label sequence {{formula:8df7c78b-f115-40cb-be56-bef663e73b85}} which minimizes the diarization error. Since we do not additionally have information about the number of speakers {{formula:dd78417d-ab15-4719-acb9-b06b28b1000e}} in the recording, we first estimate it using the heuristic described in {{cite:1a02b755cfa0e73158a38dcb3bc087c6cefaa045}}. Subsequently, we perform multi-class spectral clustering to group the {{formula:0d2d8313-53aa-460b-9656-f80673fe17af}} segments into the estimated {{formula:03af3dbe-bdda-4c41-a8d3-579469a6b6b9}} clusters using the optimal discretization procedure proposed in {{cite:a7731ca52659e15b8b04315541b1e4cfdfc08c6d}}, where we make a key modification to constrain the optimization process on the output {{formula:18baeb14-afca-4d30-8289-038f7f10e0a4}} of our overlap detector.
| m | 9750c93894858e024ef2354d52d0f9a1 |
As a macroscopic equilibrium model, et al. {{cite:5841204a86006be520dc5638ed012c5298a87da1}}, {{cite:a8b76ccbac51dd1605114912ab3509ba469f4a5a}} investigated stochastic differential game problems involving infinitely many players under the name “Large Population Stochastic Dynamic Games”; and independently, Lasry and Lions {{cite:23d353be033dbebf4a4fb8a1beb880d549229642}}, {{cite:11cd37dea6e68e14623bd7997e6307bc69693e1b}}, {{cite:86c6692477c5505d2bdc063ec3656b5a23ea4bd3}} studied similar problems from the viewpoint of the mean-field theory in physics and termed “Mean-Field Games (MFGs)”. As an organic combination of mean field theory and theory of stochastic differential games, MFGs provide a more realistic interpretation of individual dynamics at the microscopic level, so that each player will be able to optimize his prescribed objectives, yet with the mathematical tractability in a macroscopic framework. To be more precise, the general theory of MFGs has been built by combining various consistent assumptions on the following modeling aspects: (1) a continuum of players; (2) homogeneity in strategic performance of players; and (3) social interactions through the impact of mean field term. The first aspect is describing the approximation of a game model with a huge number of players by a continuum one yet with a sufficient mathematical tractability. The second aspect is assuming that all players obey the same set of rules of the interactive game, which provide guidance on their own behavior that potentially leads them to optimal decisions. Finally, due to the intrinsic complexity of the society in which the players participate in, the third aspect is explaining the fact that each player is so negligible and can only affect others marginally through his own infinitesimal contribution to the society. In a MFG, each player will base his decision making purely on his own criteria and certain summary statistics (that is, the mean field term) about the community; in other words, in explanation of their interactions, the pair of personal and mean-field characteristics of the whole population is already sufficient and exhaustive. Mathematically, each MFG will possess the following forward-backward structure: (1) a forward dynamic describes the individual strategic behavior; (2) a backward equation describes the evolution of individual optimal strategy, such as those in terms of the individual value function via the usual backward recursive techniques. For the detail of the derivation of this system of equations with forward-backward feature, one can consult from the works of Huang et al. {{cite:a8b76ccbac51dd1605114912ab3509ba469f4a5a}}, Lasry and Lions {{cite:23d353be033dbebf4a4fb8a1beb880d549229642}}, {{cite:11cd37dea6e68e14623bd7997e6307bc69693e1b}}, {{cite:86c6692477c5505d2bdc063ec3656b5a23ea4bd3}} and Bensoussan et al. {{cite:9ca90d9fc006cb569ec921c91101beb00401dcd0}}.
| i | 247ce755ab84d18fa2f7ad9c0a4f33ae |
In the {{formula:93fcacdc-008d-415d-a158-dca615459342}} cold dark matter model, the formation of dark matter structure is hierarchical, in the pattern that small haloes form first, and they subsequently merge to form larger ones {{cite:5714db618a289d40dfa909245669c8584ff27199}}.
The relics of the merging haloes and their stellar components are called as satellite galaxies. Massive galaxy usually sits at the center of the dark matter halo where the gas cooling and mergers with satellites are in progress. The evolution of satellites is more or less passive, as the cold gas is continuously consumed by star formation, and additional gas supply is cut off by ram-pressure stripping {{cite:44d0a3c9a7257896dd7a0b4f773cd7aa0a1e1319}}, {{cite:99c4c78e6a6dc07d8f8e72dbfc0aa40ca94ad83a}}, {{cite:82e28604ab19aaf94d13a7e8889095355fe4a7e9}}, {{cite:589d2557ac4a02da09fdb809cd0c87d7c7c69387}}, {{cite:9a9d6d409a039ed63b7530e552763cd63010aebf}}, {{cite:2ff7b95c10ff2bd2a61f2ebbef4c71c7d5346ff3}}, {{cite:d619ebdf1dcb9743edb34d187eeaeac004efcefc}}, {{cite:1085dac3d5e07c65d1aebe9cb110360174d75c92}}. From this picture, it is natural to expect that galaxy properties such as luminosity, star formation and color, are systematically different between central and satellite galaxies.
| i | af44511be74dc64bed46f1ff6a1434e5 |
Between these extreme limits, theoretically it is possible for the accretion disc to have a state with low viscosity and high advective factor. Since the viscosity is low the efficiency of angular momentum dissipation is low and the fluid particle could stay in an orbit for relatively large time. Since the advective factor is very high the efficiency of radiative cooling is low. Therefore non-radiative mechanisms such as convection will be triggered to shed energy {{cite:8147728c71dbc31a458c26111df36ceaaf941ac9}}. Such a low viscous high advective flow, in other words, the convection dominated accretion flow (CDAF) phase is an intermediate steady state of the accretion disc.
This state is not stable because of the inherent convective and diffusive instabilities,
as the effective buoyancy frequency becomes imaginary {{cite:193d286eaafdfa7355cd84e296337d9b5f678c3a}}.
Outflows followed by spectral transitions are thought to be induced by convection which may
also give rise to quasi periodic oscillations.
| r | 760a92a16236ea72313fab20253245dd |
For the purpose of validation the results Receiver Operating Characteristic (ROC) curve is plotted for all the networks under study, it is actually a probability curve which is used as a performance measurement in a classification problems related to medical field {{cite:cb76dd5013361aaebada0bd98931452b91e8a112}}.
For each class a separate ROC curve is plotted which have True Positive Rate (TPR) and False Positive Rate (FPR) on its x and y axis respectively {{cite:dcef86da64b73eb0bb2a3b49297a8f316afa8160}}. Equation REF and REF below shows the formula to calculate TPR and FPR from the classification results.
{{formula:ee24e3dc-1390-4388-bb1d-700dc0fbecda}}
{{formula:81d111de-27be-4df7-9fb8-a1a8a09e7d87}}
| r | 4e0c6d479198831b1214db84dd372a9b |
The main contribution of this chapter is a dense mapping pipeline which is able to construct a map of the environment based on geometric and photometric information. The evaluation is based on estimating root-mean-square (RMS) error using Eq. (REF ), as well as considering relative pose error (RPE) that measures local accuracy of generated trajectory in a fixed time interval {{formula:0c710d9b-3dae-493d-b739-b1f6e52e08c4}} using Eq. (REF ) {{cite:5f5bbe0d09380d139ed35e06036d47b3616e58c7}}.
{{formula:9d96c3a7-6b1d-4b5a-b27e-78153877c4e0}}
{{formula:3f6b31d4-050f-4144-b393-343e610c318c}}
| r | 38a7d619fedd90dc64ec78aaf28ae902 |
Other than these research, a few cosmologists have studied holographic DE models as other options to DE causing late-time speeding
expansion of the universe {{cite:408761c4b7a761fb59d41af15741f99bc0393ed1}}, {{cite:967fc2439c52d69fc91453fa2cf950c848e89c09}}, {{cite:fdc2c74b83f0f2a57085987cc48a30c44ea87e6f}}, {{cite:75ee1f6c3996b256d7a49b0b704b4e65e0edac3f}}, {{cite:b642194a34f471b99f26e06f3ae90ed5b1ef7efd}}, {{cite:3d7202ae4a204752129897e504931b31a0596658}}, {{cite:e42792b605ddfa7afe4d39ca49fd0c9dee2d1956}}, {{cite:6bbe16060bffdc6392a6d1d10f1e387ef8a182a0}}, {{cite:11de13fab75bab55c70986eac7fcc1c3bc95c53b}}, {{cite:047ebd2a8bd3434fb7b3646077236b13deb74787}}, {{cite:76344268e362c229cb0eca2ebb38430a338e5d11}}, {{cite:b3b6dccc36365962b0ba1538b56095dcc1490453}}. The hypothesis, that
phantom-like DE has an adequate resistant force to cause the formation of black hole in our universe later, is the base of the
proposal of pilgrim dark energy. Depending on the assumption of the generalized ghost version of the pilgrim dark energy this
phenomenon has been explored. We found that the parameter (EoS) varies in both quintessence and phantom regions. Here we use
linearly varying deceleration parameter (LVDP) to test the GGPD dark energy model in the SB scalar-tensor theory of gravity.
We discuss the EoS parameter, squared speed of sound {{formula:8ffa237a-7b0c-4caf-9ca1-494f7a69b8b1}} , {{formula:0e6716cf-8917-435d-8db5-6940aae91997}} trajectories and {{formula:d483b967-0539-4364-bbbc-108f12d4526e}} planes to analyze the
results {{cite:7f1d64f25055d44d9ef227d372529943b78586c9}}. Nowadays, due to the global property of the Universe the holographic dark energy has attracted much attention.
This model depends on the holographic principle {{cite:645ae0e20a7007d66dc0a82dbfa8516415b60f30}}, {{cite:df86369887e9dd66176162556ba0faed1a96dc2e}}, {{cite:3dec5c034b61dda65032c91dfc1e8968e322adc0}}, {{cite:9ff135c00e6ecf5e496c35cec4114521461b78ec}}, {{cite:7ea41fd4fe8aa715be8c2913b85cd008ab9325fb}}, {{cite:1550e38483b9e29697ef4a32799d94349f67126b}}, {{cite:261340319e77aad6db2f1f8362b4a92144ee2ab6}}, {{cite:114736ec83fd7c52f09ed83714c8cd8292c76890}}.
| i | 4b85fd03b658a2b433507bf894675f08 |
As we will argue throughout this article, this last question might in parts be answered by considering yet another apparent fine-tuning of the Higgs sector. The masses of both the Higgs and the top quark lie in an extraordinarily small window corresponding to a metastable electroweak vacuum {{cite:6825cd017e57570588a1d7c20265b8900b77d294}}, {{cite:bab6a860ba47805646196da59be7b89353cf031b}}, {{cite:f051721183bcc9b4f40f8e038e2f315f9694b037}}, {{cite:7110d7ee9dd0a5b8b1223d3d54c69ddcbe6fc3d2}}, {{cite:b7dcdfacd74074f179dca6406a0ced69be49e28d}}, {{cite:779700de685e4bd9b54d9043b5e54a4b1a9c4dab}}, {{cite:c45aaa2c61c530670701a46629529914cebdd658}}, {{cite:6abb60563ec5b043e8810fd35390dc203e8eb04e}}, {{cite:ef1edbb47fb4379c9e3f7a336eed728d5bdc5148}}, {{cite:d3b9fb3b633d2c14d64f58253fb01d1c8e943f1b}}, {{cite:ece139d636952ad19b31872e1318a67e0f063c7f}}, {{cite:5309b443101a4078c27abfde958099bf3f7f4a03}}, {{cite:95607416bd3c1ced91ee234769e70ba382f0b32c}}, {{cite:325226b334dd1f4bcccd8e7c272e505334eab78c}}, {{cite:9a34bdea99af7f8bd436c6c737720fab7dcbf91c}}. When extrapolated to high energies, the Higgs quartic coupling becomes negative at the instability scale {{formula:33358558-de80-45a9-8a65-bd6c6ba19704}} GeV and remains small, which allows our vacuum to decay through bubble nucleation. Using the most recent global averages given in {{cite:2e843c3ffd93ff8b7ff111964dbc95ddcb83ffd6}}, {{cite:3386c127f9bea60ffb9380b0590949b3946405b0}}, its lifetime, defined as the characteristic time to form a bubble a true vacuum within our past light-cone, is found to be at {{formula:8711b1f0-e48a-409e-9424-77e6dee235d4}}We take into account the correlated errors in the top Yukawa, Higgs quartic and strong gauge coupling as given in {{cite:3386c127f9bea60ffb9380b0590949b3946405b0}}. An extensive discussion of the lifetime's sensitivity to other parameters can be found in {{cite:ece139d636952ad19b31872e1318a67e0f063c7f}}.
{{formula:35ccc78c-ec28-4e08-9c7c-1b040949b713}}
| i | f46b9ca9ddf20a17031ac0eb4330a20b |
Post-processing.
For DTU, we get the final point clouds with the depth fusion tool from Gipuma {{cite:9face495b011364cca89f87967bdc42eb95b2550}} with consistent hyper-parameters, i.e., disparity threshold 0.1, number consistent 2, and probability threshold 0.5.
For Tanks-and-Temples, to avoid adjusting hyper-parameters for each case, we follow the dynamic consistency checking proposed in {{cite:9be8c858add657e62563891b0d2040c113ebd34b}}.
| d | 635d3d1077c1be7dca90955bbf21b18b |
Even though our linear contextual bandit setup is different from e.g., {{cite:e47f165b75bcbd325e834b872ad379784cbfe745}} for ranking, the availability of the feedback {{formula:a156c9ac-b1af-4028-b2b0-c62d6a13d9e9}} , which tells us whether item {{formula:e452ccde-8966-4605-b223-6438453efea6}} has been exposed, makes the analysis of the online linear regression similar to the general setup of linear bandits. Our approach builds on the confidence intervals developed by {{cite:d3e701db0a73650f8c0c2afd5ad1863602688ef2}}, which expands the analysis of confidence ellipsoids for linear regression of {{cite:19b26856895ed55b550781681deeaaf30372a1de}} to cascade user models in rankings.
| r | 0822c0fc9648fd0a54f70fa177fb2dbc |
Another crucial origin of the heavy flavor {{formula:35789e3f-a749-4b3a-bd5b-304b9123d9ee}} is the strong electromagnetic fields produced in heavy-ion collisions. It has been estimated that the magnetic field in the early stage of nuclear collisions ({{formula:f7b777a5-2aa4-47c0-a9e6-e456ad337692}} fm) can reach the order of {{formula:f783a4cc-1035-4f53-b4ba-033620d6762d}} Gauss in Au+Au collisions at RHIC and {{formula:088f383a-8e8f-4699-9db0-fc8a25a20bc8}} Gauss in Pb+Pb collisions at LHC {{cite:87239dd2c53855c9fba51a27f48bc3f1731422d6}}, {{cite:6ef23bdf5947237bfaafd7643de10bbce9fb9ac8}}, {{cite:60a1f42b0b7e06a0e2850a2d6630f08aff4bd5aa}}, {{cite:d1b3e256d4d48b7bc0b09cfff516060f4fff36fd}}, {{cite:8dbbda5052f6d3c0289df91afb4ebaca3778c217}}, {{cite:f6a3251f8d69e50973c636e0eee2cc25e2a13113}}. Such strong electromagnetic field can deflect the motion of heavy quarks traversing the medium, causing the separation of {{formula:5f106ad2-d74d-4f68-8ceb-46e1a33e6578}} between {{formula:884a2bf2-9516-4442-a4f1-38064fda83b1}} and {{formula:bda1b92a-39ed-4c43-80ca-ef52dcebe928}} mesons in the end. Interestingly, while the STAR measurement {{cite:2ec8734ef6a17d285ca49d364fb1ad26a45f3945}} observes decreasing {{formula:a392e2dd-562a-456d-b7f3-b09e634efbf0}} with respect to rapidity ({{formula:ba056610-575c-4e26-9d95-cc7ca47eed3a}} ) for both {{formula:2ca27bb3-5f53-47f9-8d29-be2e77f920d5}} and {{formula:00b1db3b-e932-4544-b7eb-8cb651caa4e7}} , with very small difference between their magnitudes, the ALICE measurement {{cite:4aedf6fd68ce94fa050bcf1e0a737c5dc630896d}} presents apparent splitting of the directed flow ({{formula:6ba48fc0-ef1d-4e5f-aac9-e681aa2fed76}} ) between opposite charges, with {{formula:58086169-7ceb-4d50-9974-2828b510ca3b}} increasing and {{formula:6649bde2-2153-4905-98be-eebaaccf238b}} decreasing with {{formula:f5df258c-2993-4da5-b4e4-1284fef148a5}} . This puzzling observation has attracted a series of investigations on heavy quark dynamics in the presence of electromagnetic field {{cite:993c27c4a53f2d7644186a3a0afa7cc72581bcbf}}, {{cite:cdb9f9a1bb0817fc2794c932890fb9a7f2a5ed6a}}, {{cite:af14908ee9ce4274c3ab5aec24b859d4014f2c4a}}, {{cite:68f34e0adf66ed4c7e434675c18e6afe6bcb0935}}.
| i | 4f0cff2734e1284084bf1be0700863fd |
Besides semantic anomaly detection, self-supervised methods show satisfactory performance for defect detection and spotting sensory anomalies {{cite:e14cd3f245183e99ee62a3fe36c689523e3cc249}}, {{cite:04f6b4c27462a73d5fa1e32bddc69a10eebc62bd}}, {{cite:5dcea0b2f2de9d98d1a53298d74a89d403a0138d}}. Fig. REF shows the performance of the self-supervised models on the MVTecAD dataset against other widely-used algorithms including shallow models, deep models and generative models. More specifically, the compared shallow models are Gaussian {{cite:f60ed9f05290b5c6affc389a45cc66bb929f650f}}, MVE {{cite:f60ed9f05290b5c6affc389a45cc66bb929f650f}}, SVDD {{cite:f5005cc794dd76e76b199e08c66da6174aadc170}}, KDE {{cite:f60ed9f05290b5c6affc389a45cc66bb929f650f}}, kPCA {{cite:f60ed9f05290b5c6affc389a45cc66bb929f650f}}, patch-SVDD {{cite:e3ab88985dda82f5d9a91ab0e72aed2adf62151c}} and IGD {{cite:93954cb97cbaa3d9ea82c8917a37168f88e91708}}. The compared deep models are CAVGA {{cite:c974df4f7b39762c2f8a0b5758d969ba90ca7624}}, ARNet {{cite:44ba16f2d4d76508f13756be6aae94a921a2acd2}}, SPADE {{cite:dd9f6507f7248ea48b4a4a30637b1904fe1b69fd}}, MOCCA {{cite:ad8a77774a9177d466f792bc989fb40b1bd703db}}, DSVDD {{cite:0b742c265240289882979a48ae856b173bdfdfbe}}, FCDD {{cite:ca7546317b3e0bdbce934537a9abd880320e4416}}, DFR {{cite:bafe42db124e1e7de2505aadd1b0e8050e3e29b2}}, STFPM {{cite:9d55ce08c4ffef8f6ce9dc730dff25c7f51cfdad}}, Gaussian-AD {{cite:e2555a125015d0704b2ee60c8a9fd7fa9d98f5c9}}, InTra {{cite:de7a176a1ea04e4da144f2521dcaf9693398ae77}}, PaDiM {{cite:1967ca3a0dc63b475a11d8095e1980481c5734cc}} and DREAM {{cite:de58aa58508653fe1327175bd276d036059d6750}}. The included generative models in Fig. REF are AnoGAN {{cite:7cf5981046b4247a6738407d4c2687005aca394f}}, LSA {{cite:ab112266ff01863631f55b78e79e02e048dd10e4}}, GANomaly {{cite:ae4eb0b19e9ea725cebb0fae155adf6fc437cae1}}, AGAN {{cite:f60ed9f05290b5c6affc389a45cc66bb929f650f}}, Normalizing Flows-based DifferNet {{cite:dae86f7a4e41e7e24800b47ecadf56160b751a16}}, CFLOW {{cite:e128384cf6e17b4d48be786ed3d03202af724b62}} and CS-Flow {{cite:f13488a257f271fe1b00a21b48f762867a9b10c5}}. Looking at the figure, we can infer that SSL-based models can achieve a good performance on this dataset. However, the superiority of self-supervised algorithms over other baselines is less evident in this task than in one-class AD. Also, some algorithms such as GEOM, and CSI, which show state-of-the-art performance on CIFAR-10, achieve a weak accuracy in this anomaly detection task.
| d | 72781893771aa1477b71d9db179a9afb |
Besides, to analyze the mechanism about ParC, we provide the result of two commonly used visualization scheme, Effective Receptive Field (ERF){{cite:0d2520cb6529fecb9b994b191fc89aa1a5d30dee}} and Grad-CAM{{cite:9b56d7f4a80bb1f4257011e4ba32201b6ce01f56}}. Result shows that ParC-based models are able to grab global feature even under high resolution, and this makes model more comprehensive when capture instance related semantic information.
| d | abf07cae4ce9c5d7322ee56ceba944ee |
We gave further evidence, and developed the understanding around the filtration, by studying also attractor flows for multi-centred black holes. These should capture the spectrum of BPS states, and we find agreement with the filtration proposal. So a decay which would violate the filtration, manifested as a split attractor flow, leads to an unphysical attractor locus for one of the constituents. Also, the filtration can be understood as the statement that the wall crossing formula {{cite:1ae65b45e95a38f03123dc3c8fb4514204e5efb3}} should apply to all states in the theory, everywhere on the wall (so imposing the absence of non-stable states anywhere on the wall).
| d | d457a11bb0ffe50aeb293bf4a7e89e24 |
where (REF ) follows from the tower property {{cite:07dbe9b4063c132485b12dda9f2957fed164f263}}, and () follows from the time-homogeneity of Markov process {{formula:5616afbe-5f38-4916-8535-738cac8e12fa}} .
| r | 4952e322e947ef36307916e25ff237ce |
We also conducted experiments on the VGGFace2 dataset using similar parameters with Resnet64 in SphereFace {{cite:385fd44ac9f1166c02632bcda21feb086704f5b3}} to show the results of our part fViT in Table REF . Despite adding a large amount of data augmentation, our baseline fViT perform worse than the results provided by Resnet64 which is similar to the Face Transformer when training on a small scale dataset such as CASIA-webface {{cite:498620e3905e38f54276d6e4d7fcf6a839221244}}. Our part fViT also achieves a better result than the baseline fViT when training on MS1M, while it is still a little worse than the Resnet64 with advanced losses(e.g. ArcFace {{cite:394186247352d765855a570a1768bfb580ebfc00}}). Our future work will investigate how our method works on other large scale benchmarks like Glink360 {{cite:fde9f557bed5647de89158ae59c4e82028a4fd2c}}.
{{figure:1bf3facf-f103-435b-a3c5-e2b41ab138ee}}{{figure:bd259e62-d079-4479-b5b3-31cab281c011}} | r | d6e8c4ccbc97b9e4f2e0b932c03dfca4 |
Spin-orbit coupling is a universal phenomenon that can be observed in a variety of fields in physics including classical mechanics {{cite:085c34c9df0057fd15151e60a9681229f6b2d835}}, quantum physics {{cite:210fd26fea20b96e68d1f85afe7b77d18b09964e}}, and photonics {{cite:532393182098e10455310df38ef899a5019a8246}}. In particular, there exist diverse kinds of spin-orbit interactions in photonics because of rich physics enabled by the two electric and magnetic vector fields {{cite:fdf2a4f38e89a7875385fd8e6dd6439f121693aa}}. Among them, one interesting spin-orbit related feature is the spin Hall effect of light (SHEL) {{cite:3392a58155f7d2ef7a613c62cf848a9d66a9023c}}, {{cite:12672a74554904fd115657ccd8e2be520b6477e0}}, {{cite:28c684330b950ac9abd1df001ab3f491ce0e3105}}, also known as an Imbert-Fedorov shift {{cite:b3b0275699232e4b12e100a44d26817967263d21}}, {{cite:ab4d6b78efe977841b8e20f1e189952b26b89caa}}, which reveals itself as a transverse and spin-dependent splitting of a finitely-thick beam at an optical interface. A physical mechanism that underpins the spin-dependent shift is the transverse nature of light, {{formula:fc60cb16-95bc-4a4e-ad69-755d82cc5ee6}} , that makes the incidence contain a bundle of slightly differently defined polarization bases {{cite:d3b3b90d72f351240490cbc7c230512ad2a3e8dd}}.
| i | 0a75e3ab056bae88f1d5ae8bb7c0611c |
The performance of a machine learning system in practice depends on a large number of considerations {{cite:423387c0737c5897fc5cafe0403819ffdb23dcd6}}. State-of-art neural network architectures for image classification contain millions of trainable parameters, and they are trained on hundreds of thousands of scans {{cite:8a2e856c8a6173da8fd8e322583bf07ee3012b80}}. On the other hand, our network and the dataset it was trained on were both far smaller in size. We replaced only a single layer consisting of 6 trainable parameters with a variational quantum circuit to create the hybrid network. As quantum computing hardware becomes more accessible and the software ecosystem around it matures, researchers and developers would be able to incorporate larger quantum circuits within state-of-art classical neural networks and train these on large-scale datasets. We expect hybrid networks will begin outperforming the best-performing classical networks at that point. We see our work as a stepping stone leading towards that milestone.
| d | ed172e8eb1ec61dd13732f97729595d0 |
which is similar to Equation (6) in {{cite:5e457df54a0457fdbc20d54cc4d62e33b18fa955}}, where a tight correlation was found between the widths and redward shifts of the Fe iii{{formula:43e0ae22-9535-4d07-997c-2a0ea3415cd4}} 2039–2113 blend for their quasars, and this correlation supports the gravitational interpretation of the Fe iii{{formula:772132bd-8d9f-48cf-9ca4-492fb38c695d}} 2039–2113 redward shifts. The Spearman's rank correlation test shows a positive correlation between the line width and velocity shift of the H{{formula:bb93dd7d-c57a-4902-b873-fdcbe94fb658}} line for these 1973 quasars (see Table 2). A series of lines based on Equation (6) with different {{formula:281ae276-ee33-43f4-8077-4d8134ad444a}} are compared to the observational data points (see Figure 6). From top to bottom, the corresponding {{formula:02811e29-bd75-4aa4-9915-beba5892ea2a}} increases. Because of the co-dependence between the Eddington ratio, dimensionless accretion rate and {{formula:86366e70-c3d0-45d3-ab4a-1d182e228ce4}} , the large ranges of the former two quantities may lead to the large span in the direction roughly perpendicular to these lines (see Figure 6). Also, the metallicity difference of BLR might decrease correlations in Figure 1. Micro-turbulence within the BLR clouds can act as an apparent metallicity controller for the Fe ii, and the reduction in the value of the metallicity can be up to a factor of ten for the Fe ii/H{{formula:5a749f90-7cd4-4c00-8b37-aa4fdb73a5ff}} line ratio {{formula:3d9430ee-3e26-44b5-8568-6ea68c561369}} when the micro-turbulence is invoked {{cite:e11d795a5818741b8c8052ca6f533186f9000c8c}}. In addition, spectral simulations show that {{formula:13c37dd9-0e00-42f2-b32a-5ed46a99f01d}} depends clearly on cloud column density {{cite:a2059ce4c59c7994ede43b45db098cd426826b2c}}. The combination of the column density, metallicity and internal physical processes may further decrease these correlations in Figure 1.
{{figure:9133467f-be31-4297-881e-c28bc1845fff}} | d | ef31c64de17b3642caf34fe5794850b1 |
Recently jets and jet substructure have been proposed as alternative probes for
portraying the full three-dimensional (3D) image of a nucleon and enriched the content of the transverse-momentum-dependent (TMD) spin physics {{cite:b07ddc39b83ab9d93c4939b6dc0771e5d9789d39}}, {{cite:98c4f97874540ce3c2d777e004a730cea8b8da84}}, {{cite:a14c2bd5c9a4a7c8f4b0dc05bc4bd0fb38e22553}}, {{cite:c3641888baf15f587c83caa661d5a584d5415816}}, {{cite:28d319c825a400f1970bf12fc3e439fb2f6abab4}}, {{cite:9a38f8b1d84d4f3527cd4b96d89acaeb8c371275}}, {{cite:0cbd639feb1a8d0f7ce8d0b1b7df9f1abd920ffc}}, {{cite:0a1602e52eb8ee780ff2efcb54fef3e73326362a}}, {{cite:9be19844c7c2b099105385a4271534857f2622db}}. The jet probe into the nucleon structure has been shown to be able to access the TMD parton distribution functions, including the Sivers function of a transversely polarized nucleon. Conventionally, we require the jets to acquire large transverse momenta, and therefore jets are regarded only feasible for high-energy colliders such as the LHC but practically challenging for machines with a relatively low center-of-mass energy such as the Electron-Ion Collider in China (EicC) {{cite:7d14257af6f83e451852fdd0d7c2df2e0a3fc983}} or detectors more optimized for low energy scales, such as the EIC Comprehensive Chromodynamics Experiment (ECCE) {{cite:bdea1de6469951456042a03e82e2501582f52c22}}. However, in this work, we will argue that this is not the case by reinterpreting jet clustering as an axis-finding procedure to measure the virtual photon {{formula:9a1a755f-7c70-4223-9344-89684a941e78}} , which allows an inclusive probe of the TMD spin physics suitable also for low energy machines {{cite:69cdafd4ff16436d201109bf42eb61b182b6ccc0}}.
| i | a757cb479b5055a528bd03170f7edfa7 |
The first term accounts for local anisotropic drag that depends on the aspect ratio of the filament {{cite:b856adae0f3bc4d0fc3b357bbb8e1eb23c2cb51c}} and the second term accounts for non-local interactions {{cite:161bc5798ed767b1f6fb8d346c4c18d5a9e01e36}}. In (REF ) {{formula:31a2bb9e-04a3-4a96-9b33-f7e287d004f6}} is the disturbance velocity generated by all the other filaments at {{formula:ab93cc36-14be-47b7-a77f-cdb4936f0064}} . This accounts for the long-range hydrodynamic interactions and is given by
{{formula:dd38acd2-376c-40b4-af03-859d728d7b08}}
| m | 311fe9a198122d5d40475bd61f0ad256 |
More generally, we believe that continual learning with nonstationary time series is a relevant area of research. Although transfer learning is primarily concerned with transferring knowledge from a source domain to a target domain, and therefore may be used in either an offline or online setting, an increasing number papers focus on online transfer learning. See for example {{cite:f7f03d9dcbefa56829630a59c7fc71cfe54192f1}}, {{cite:d9e90e83beb1cfdb37397d8ee03cc315fb911c6b}} and {{cite:1c77567ef87615c30f71add01c7cee7ab84e8b88}}. Our paper has contributed to the research of continual learning in financial time series. This is important as our experiment shows that continual learning provides a benefit with multi-step forecasting, above and beyond sequential learning. If we compare the local learning of the radial basis function network with the global learning technique of the feed-forward neural network, the latter suffers from catastrophic forgetting. {{cite:c09e69686c30566d13e096eb7bea5332dc432c3e}} and {{cite:257120e9d80d36efafcc7e4bacf51b058d8066dd}} look at ways of improving this issue, specifically at training networks that can maintain expertise on tasks that they have not experienced for a long time. The radial basis function network that we formulate, is naturally designed to measure the similarity between test samples and prototypes that capture past characteristics of the feature space. As such, the model is somewhat robust in mitigating catastrophic forgetting.
| d | 0b8028f0e81d34892a69bc3270442320 |
enabling them to implement minimal sufficient statistics for different {{formula:2543f8c9-7bd0-4b36-bc6e-a6ff92253f84}} -constraints on the error.To what extent the relationship between IB and deep learning holds in general is
the subject of ongoing debate {{cite:9a6d46032577b493d55b96e630a8c5d9005ae7a7}}, {{cite:1bfc541c1d749e020bf4a1a0e8515a7bcf1335fd}}, {{cite:2ec65b112397fe19b35e602873c2aacc11f81411}}.
| m | 45db2df07524519fae4c228f15c91b3b |
HiPe is able to create model-agnostic saliency maps so quickly because it is content-aware in a way that existing perturbation-based saliency mapping algorithms are not. {{cite:e158f009635edce5771202dcb7b4d816a858d96e}}, {{cite:418d8983b0140c0124f3f798768d18b3f2667d13}}, {{cite:22be4e471c0e27ead0d565a98f3d138dea8da96d}} require a pre-specified number of iterations – whether that is epochs, number of random masks generated, or occlusion kernel size and step – which fix the amount of computation required for an input of given size irrespective of the proportion of the input that is actually salient. It is also impossible to know ahead of time what the optimal value for these parameters might be in order to trade-off accuracy and efficiency, and finding the optimal parameters (for one input sample, or across an entire dataset) may require many trials. Additionally, the these parameters limit the size of salient region that can be detected, which can lead to omissions as in Figure REF . Our method, by contrast, continually disregards regions which have little affect on the model output, and by so doing so inherently limits the amount of computation required without imposing restrictions on the size of the salient region it is possible to detect. We also note that HiPe is data-agnostic, as well as model-agnostic – although we benchmark it on images here to allow for comparison with existing methods, it may be applied to data of arbitrary dimension.
| d | 5784612b77e23c2f29851815671f6982 |
In order to explore the impact of non-uniform scaled abundances on stellar evolutionary tracks, we construct the Stromlo Stellar TracksThe Stromlo Stellar Tracks are publicly available online at https://sites.google.com/view/stromlotracks, self-consistent stellar evolution models using the Modules for Experiments in Stellar Astrophysics {{cite:1f4d5d2c9fa287760196bb69f84a5add76619a85}}, {{cite:6042bd2e9f9ad09af29e82275f020a8eb81512d1}}, {{cite:1d682e335e6aaade7b23a579529e12d120b10232}}, {{cite:f24c78cec7d616d6860468fb2ef39b31257c69a3}}, {{cite:ea1ce935f6fd52207647d30c8eca3e99a54d7de7}} stellar evolution code. We build upon the stellar evolutionary models used in the MESA Isochrones and Stellar Tracks (MIST) by {{cite:da3ac57ca8c920459f457392319b6469e01c3b3f}} to present models with the scaling abundances based on Milky Way stellar abundance data referred to as `Galactic Concordance' {{cite:f7183ad66ae0a73a66194c1d3e2159d50c9ebb42}}. We focus on massive ({{formula:dd52a145-e94a-4129-b345-e87dd73b6e62}} 10 {{formula:fc24d803-278e-4162-9b54-88ca23ed6875}} ) hot stars that dominate the ionizing budget that power H ii regions and provides the feedback that regulates the efficiency of star formation {{cite:ab2c7d10f71dd7c4635eaa3ed7859b30e8ad61b1}}, {{cite:909eed540a4530c3763aa347f37a2da1adbea5e3}}, {{cite:ca6917a4d22f4770e06d2a412a2c88cebfad8b88}} and drives turbulence and wind outflows.
| m | d9046e79e7c48253216fc1a91802b114 |
The requirement of differentiation of the log-likelihood or surrogate
functions for some of the bias-reduction methods in
Table is also another area where considerable
analytical effort has been devoted to (see, for example,
{{cite:a7ea71cc627d35a28f3c72c78053a0db08cf7339}} for multivariate generalized nonlinear
models, and {{cite:e89c0f09d5cf5d4b3164f09335cd1f7ebae89229}} for Beta
regression models). Nevertheless, differentiation is nowadays a task
requiring increasingly less analytical effort because of the
availability of comprehensive automatic differentiation routines
{{cite:2076574f1e06ce5b2d4ccda5432f5785810fdda5}} in popular computing environments; such
routines can be found, for example, in the FowardDiff Julia package
{{cite:0499cc7e53426c88a08cc4904e96534acf20f23d}}, and the CppAD package for C++
{{cite:59b1b0f5c8fde51a3879909ad00df053775ea3c2}} that enabled the development of software like the
TMB package {{cite:0b957aab04e4ebd1d38ec11da8e97ada126a9033}} for R
{{cite:457999f95e13bbcd75e1ddf914b534e47582bf8e}} which is a generic framework for fitting and
inference from complex random effects models.
| i | 88c1e24fe62d630e652a930f1a9db335 |
As mentioned earlier, in the solid state physics interpretation of the partition function for our (0+1) dimensional {{formula:e0d9287a-d6f5-4377-870c-8d322dfb7973}} model, we consider the Euclidean time-direction as spatial dimension and let, instead of {{formula:29b0b93a-e5e0-4f73-b854-5c3f5109035a}} , the paramete {{formula:6fbdc32f-7efe-487f-9787-33abed7881a5}} play the role of inverse temperatuere: {{formula:50a9a6ff-6af8-4a83-9ad3-e3ac1208054d}} . The Euclidean action then turns into {{formula:e7abeb75-2455-4b40-a7c9-acf2c0cb4d39}} , with {{formula:87605889-bc98-49da-a048-ee950f5809bd}} being the Hamiltonian of a corresponding one-dimensional solid state physics {{formula:85144b25-3a4c-4496-bfcc-61709dfa7930}} -spin system. In this context, the above mentioned findings of long range order and a non-zero condensate in this 1D {{formula:6ee71e47-f843-427f-af28-011360d33e8a}} spin system seem to conflict with the findings of Mermin-Wagner {{cite:3d31f942b32f4c3d081be0adca4f4a78cbf8bbd2}} and Wegner {{cite:5e44965007d9a2c10b01a0cf6cf93e4f5745b4ef}}, that an {{formula:093c56c6-2249-465c-9935-501f30c2f14d}} -symmetric spin system with finite-range interactions can in one or two dimensions not undergo spontaneous magnetization, and not develop long-range order. Why does our model nevertheless show signs of long-range order at a discrete but infinite set of {{formula:9453d58a-e225-414c-9a19-897b91a8c20a}} -values?
| d | 74693a023fa869fab356354b0578288a |
Ablation on losses of THAT. We show results of THAT with different loss functions in Table REF . We observe that compared to our full framework, removing either the contrastive loss or the normalized softmax degrades the performance slightly, yet the resulting models still outperform the standard adversarial training baseline by at least 1% in clean accuracy and 0.6% against PGD-1000. This highlights the importance of both components for adversarial training. Furthermore, without the contrastive learning branch, our framework produces 58.92% and 42.78% accuracy when evaluated on clean images and against PGD-1000 attacks, respectively. The result is slightly worse compared to removing the normalized softmax (, a standard cross-entropy loss is used in the pipeline), suggesting that contrastive learning is relatively important.
Note that we did not compare to TRADES {{cite:42d72ee0e63947bf2e9c69467749c4e05c22593c}} because we found it to be unstable and non-convergent, even after hyper-parameter searching. Similar attempts and failures to train TRADES on ImageNet are mentioned in {{cite:45d6042f317ceb521826ea7bb425c858d0381c40}}. See Section .
As a workaround, we compare with standard adversarial training with KL divergence loss to force probability distributions from clean images to be close to those of adversarial images Note that this is different from TRADES as TRADES maximizes the KL divergence to generate adversarial examples, while we instead maximize classification loss as in our approach.. We see that it achieves high clean accuracy but is extremely vulnerable to strong PGD attacks. This is likely because of the strong emphasis that the KL loss puts on unlikely class labels (see Section ).
{{table:2854f736-6175-4faa-8c50-ae0e3e380a8d}} | d | 111bb60a9945a3b7d20f81997c379cb8 |
We compared the proposed method with four well-known adversarial attack methods which are described as follows. FGSM {{cite:b1c53d022ccd058f4ad2eb4b1c5d8f907d410068}} is a gradient-based one-step {{formula:9165f77e-94fd-4a51-82d8-60051f548c78}} attack. BIM-10 {{cite:6bc778b8da4538d9c459329145a552d4e05db55a}} is an iterative version of FGSM with 10 iterations. C{{formula:0b5e8c18-7dc8-4117-8dfa-297fc24993b3}} W {{cite:494f1641edc09424ccd412226ff12d01b9b7112b}} is an optimization-based attack. UAPs {{cite:cd5b92ab294961092298f262cf0be340597d1ac3}} is a universal generation-network-based attack approach.
| m | 784ff55c5505d5a617b034cfde7c6934 |
To test the transferability of the found pattern, we also transfer it to three other tasks: 1) language modeling on WikiText-2 {{cite:d28e2a31cfb78a22c2ed7961b13fa7a648b3e849}}, 2) German-English translation on the IWSLT-14 dataset, and 3) English-French translation on the WMT-14 dataset. On Wiki-Text-2, we compare AutoDropout's dropout pattern against Variational Dropout because we find that it works better than vanilla Dropout on this task. On translation tasks, we compare AutoDropout's dropout pattern against the vanilla Dropout configurations that are typically applied in Transformer models {{cite:038e37807d621ae1e011a257f4b2883711902d42}}.
| r | ac5cdff81f8a7dcb56df6529ba606875 |
Theorem REF says that (REF ) are sufficient conditions
for weak recovery. Meanwhile, (REF ) are almost necessary conditions for weak recovery, since if (REF ) fails,
then weak recovery is impossible.
These conditions are derived under model (REF ),
and are dramatically different from the ones derived under
stochastic block models ({{cite:575d515aaf7483d6c338f3158b3e68d6bcdc3f54}}, {{cite:21c1d2ed3d44077982ebd13a2bc794087ef826ae}}, {{cite:3fded0ac302c3e9099ae9169b47e94695dd3aa19}})
or generalized censored block models ({{cite:034bc3d2e5200e8e154fb46396c72e0d87d735b3}}).
The theorem holds for either {{formula:70353b57-6a6b-4453-be2b-d4162752a94b}} or {{formula:f9767838-a07f-4a31-b350-ed4bf753a438}} , i.e.,
the cardinality of the underlying subhypergraph is either significantly smaller than {{formula:7067bb59-5948-4928-a6b1-e819e16361b1}} or of the same order as {{formula:77a1b9ce-ad7b-448e-8882-7fbd30aa8835}} .
| r | 4907a92b10a7fc171f946a3d814d4c25 |
We propose our algorithm EigenFind (alg:EigenFind) for a more efficient counterfactual search. In the previous paper, {{cite:4eeb0fb9d0f3ee1904d81485e57422e9e98a724f}} presented the AttFind algorithm which iterates over all coordinates in the StyleSpace while changing them one by one, searching for coordinates with the largest affect on the classifier decision. We factorize the StyleSpace with PCA {{cite:55b11a7d60b8e7104f43129e763d225d3ee3aa06}} and modify the algorithm to iterate over Eigenvectors instead.
| m | cd590efd3b8c44f287769df37fc97ccd |
Nevertheless, the computer vision community has attempted to solve the point set registration problem through consideration of outliers and missing correspondences, which are typically encountered in real-world applications. A common technique used in point set registration to robustify the optimization against outliers is to employ random sampling consensus (RANSAC) subroutines {{cite:6d37c250951ddf9d2ba71f24f30ca8f62c0f865a}}, {{cite:84685d472498a4e9debfe0946d49b4cab70d1737}}, {{cite:8572c6068c261c545bcc58c5a80ee02783b0cc9f}}. The main advantages of RANSAC are that the randomization procedure employed can severely reduce the computational cost of an otherwise combinatorial search.
| i | cb9518c2cfd18a28052f0605c88fdeb2 |
Starburst galaxies, that are characterized by a high star-formation rate leading to a large number of cosmic-ray-accelerating supernova remnants, and the accretion disk-fed black hole regions in galactic nuclei, so-called active galactic nuclei (AGN), are both phenomena on their own that generate high-energetic cosmic rays (HECRs).
Associated to these cosmic-rays, multiwavelength observations have shown that AGN and starburst galaxies provide non-thermal emission over a broad range of energies. With first Fermi detections {{cite:08caa9686834ca6fdf680757228bf8e6114d5018}}, starburst galaxies started to become visible in high-energy gamma rays. Moreover, some starburst galaxies, such as NGC 1068, NGC 4945 or Arp 220, show a dominant non-thermal contribution from their central regions, indicating the presence of an active black hole. These active galactic nuclei (AGN)-starburst composites are of special interest, as the simultaneous energy release by multiple supernova events as well as by accretion and possible jet activity of the central black hole makes AGN-starburst composites very promising candidates for the direct detection of HECR and high-energy astrophysical neutrino sources. The latter is in particular the case for NGC 1068, since a recent search by the IceCube experiment for sources of high-energy neutrino emission {{cite:753c0e8be5648c746da64287ea55fead4a2e7365}} exposed the direction of NGC 1068 as the most significant one in the sky, with {{formula:a0d3dd38-6ecc-45b5-821b-f90c5a00fdbd}} . With respect to the non-thermal emission of photons, neutrinos, and HECRs from these astrophysical objects, the AGN and the surrounding starburst are usually discussed individually, e.g. {{cite:1a41eeb230c0ba151ccf049af25bc0bb6e8f7c84}}, {{cite:6cb3cca6600a6d47e1e276a9d54d17a1a547fe43}}, {{cite:7608103adc90220d0f0066777d54769c314b8f90}}, leaving many open questions on the leading interaction processes within the AGN corona compared to the surrounding starburst, on their interplay, and on the impact of the torus that is typically in between these environments.
The observed large-scale motions, in particular the strong ionized gas outflow in the inner hundreds of parsec or the even larger galactic superwinds as, e.g., observed in NGC 3079 {{cite:3897c66e67fe8383df9fdad3543af149be949dea}},
indicate that the starburst and the AGN activity are connected, though the particle and photon target fields of high-energy CRs change significantly. For instance, continuous and line emission at infrared wavelength dominate at the torus and the starburst region {{cite:1a41eeb230c0ba151ccf049af25bc0bb6e8f7c84}}, whereas the optical/UV emission by the accretion disk that gets comptonized to X-ray energies becomes important within the AGN corona {{cite:7608103adc90220d0f0066777d54769c314b8f90}}.
| i | 8d72edb8e03ed3f741f6184c864e8bb6 |
The electronic structure calculations {{cite:c6722ce941810f1b9fae01622dccb54901b9af0b}} are performed using first-principles methods within the frame-work of DFT with Perdew-Burke Ernzerhof exchange correlation energy functional{{cite:8038bb126d5bb0cf2d83828574d3e1d09aa23a1e}} based on a generalized gradient approximation. We used a projector augmented wave method as implemented in Vienna ab-initio simulation package (VASP){{cite:d7aed27c650c7cc9400af0d037ee240ff5b726a4}}, {{cite:9a9e565e003ecd1a7732baf5af4f256f7b72db4c}}, {{cite:7870dc14957650861871babd1240e1ca55b044bf}}. The Kohn-Sham wave functions of the valence electrons were expanded in plane wave basis with energy cut-off of 500 eV. Ionic relaxation was performed using conjugate-gradient method, until forces were reduced to within 0.01 eV/Angstrom. The Brillouin zone sampling was carried out using Monkhorst Pack grid of 11x11x11 k-points. The band structure is computed along the high-symmetry k-points in the irreducible Brillouin zone, with 100 k-points between each pair of high-symmetry points. Computed band structure with the self-consistent density of states (DOS) is shown in the Fig.REF . Since as an input for the transport calculation within Rode's method, only band structure for one valley is needed, we have performed non-self consistent calculations of the band energies in a special k-point mesh around the {{formula:ebb6b16e-6608-41c0-906b-27fc9bb46d26}} point with 8531 k-points. Using such a dense mesh we have obtained very accurate group velocity and effective mass.
| m | 5c74f1f9c8282014898c5b70a3062667 |
To draw conclusions about individual variables of interest in a task, one builds structured probabilistic graphical models to describe statistical relations among all variables and marginalizes out all other unobserved variables. Such exact inference computations are infeasible in general, due to exponential complexity in practice when the latent space is large. Often the latent space to be marginalized is decomposable due to conditional independencies between variables, allowing us to represent the full distribution as a probabilistic graphical model (PGM) {{cite:8c992dcd676a293d74ee7f503f979c10a6b2fb83}}. This may allow us to perform difficult global calculations using simpler integrations on subsets of variables. Such approaches are used by message-passing algorithms like Belief Propagation (BP) {{cite:7774ed4f33b445ec87c97c6e2c826542c6cb3fab}} and Expectation Propagation (EP) {{cite:dd357f100df19b79fc37bc0599b5a3c6c46696bd}}, widely used approaches to computing or approximating marginal probabilities using distributed computation. BP is guaranteed to yield exact results if the graph has a tree structure. However, on general graphs with loops, which are likely to be better descriptors of real data, BP can give approximate posteriors or even fail to converge.
| i | da40a08653e084f06bf54a111f2c3b83 |
To the best of my knowledge, this is the first time a TSF has been shown to resolve the Einstein's Boxes paradox. The TSF resolves the paradox in the ways that Einstein and de Broglie envisioned. The transition amplitude density {{formula:bd36000b-9eaf-4be4-9c39-0e82eb505037}} “localises the particle during the propagation {{cite:75abbcb93ebd7bdf3ce7825b6aa898f8a472b0e2}},” and {{formula:d56bf31c-9611-45ce-8546-05601afe6bde}} “was already in Paris in box {{formula:0e80e4e1-7894-4bcb-847f-0e6165c7b4ef}} prior to the drainage experiment made in Tokyo in box {{formula:89f087c3-099f-498e-b2ae-7bf325ffa291}} {{cite:597266335849d9aacae286821836a9338afa5140}}.” None of the problems associated with wavefunction collapse occur. The TSF appears to give the sought-after exact description of the probabilities and a complete description of the physical reality.
| d | 8223a947a744ba18250dc0d9f25e7708 |
Nowadays, the physicists are mesmerized by the natural beauty of electron-positron-ion (e-p-i)
plasmas because many painstaking observations disclosed the existence of e-p-i plasmas in
various regions of our universe (such as supernovas, pulsar environments, cluster
explosions {{cite:fa6b6b34414f3251f86a2fdecb60c95b31493259}}, {{cite:1e77e380d1b3c330a23c5013bb58862f8b723e28}}, {{cite:58d01488e3f66ce8380676444c36c8623a3cc0d1}}, etc.), polar regions of neutron stars
{{cite:cdb904937d872276c4122c800420fe051331716e}}, white dwarfs {{cite:59b4c9381a16d744cd0a10f08eb39d4a89379507}}, {{cite:827af720f9d9306971ecb0985ed986699a9b6c3b}}, early universe {{cite:6f82d17a15cc1968382c7ecf70e541a834c4a09f}},
inner regions of the accretion disc surrounding black holes {{cite:709cddeb323ec1791fa41b424cd3e4ed5c34ed88}}, pulsar
magnetosphere {{cite:f82ef6b35c347860c4dffd45ca76af6af512bfa5}}, {{cite:652e84ae448e0010d9efbcc51f94b962b3e6fb59}}, center of our galaxy {{cite:3390709d23e7df06d341f1c18756a4c3e260c802}},
and solar atmospheres {{cite:a3489f4bbfcd9377787bcbd907d842c66557f3ef}}, {{cite:b787e18dd5bfa343297b94c55bf5fddd96b7a2e1}}.
| i | 505d4bb850fbe3bf7f0701a4f6bf1045 |
The effect of acetylcholine, through the M-current, on the synchronization of cells has been explored using numerical simulations and phase response curves {{cite:22741ae1c7909928cd08281f8dc193b32d7d9562}}, {{cite:69514fbdd6b5b8d34693fa60484bd537670dfc88}}, {{cite:e48aa56bcd8b377bd4a33398b6f2c7fffbba58c6}}, {{cite:fa1223b48445e67feec6db8e4dfe698a646fb93a}}. We have linked these effects to a particular bifurcation structure of conductance-based models with an M-current and given conditions for this to occur in any conductance-based model. This approach allows us to generalize previous results and to easily explore the effect of multiple parameters in these models.
| d | b78d87aa86bd7366eaa9d535f9ef3147 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.