A5112582835- Domain Adaptation and Few-Shot Learning
- Multimodal Machine Learning Applications
- COVID-19 diagnosis using AI
- Advanced Image and Video Retrieval Techniques
- Topic Modeling
- Video Surveillance and Tracking Methods
- Advanced Graph Neural Networks
- Privacy-Preserving Technologies in Data
- Advanced Neural Network Applications
- Human Pose and Action Recognition
- Cancer-related molecular mechanisms research
- Face recognition and analysis
- Generative Adversarial Networks and Image Synthesis
- Machine Learning and Data Classification
- Machine Learning and ELM
- Anomaly Detection Techniques and Applications
- Adversarial Robustness in Machine Learning
- Explainable Artificial Intelligence (XAI)
- Fire effects on concrete materials
- Particle physics theoretical and experimental studies
- Advanced Vision and Imaging
- Quantum Chromodynamics and Particle Interactions
- High Entropy Alloys Studies
- Hydrogen Storage and Materials
- Natural Language Processing Techniques
Institute of Automation
2015-2024
Guangdong Provincial Center for Disease Control and Prevention
2023-2024
Chinese Academy of Sciences
2006-2024
Shanghai Jiao Tong University
2010-2024
State Key Laboratory of Remote Sensing Science
2024
Jiangxi University of Traditional Chinese Medicine
2023-2024
Qinghai University
2024
South China Normal University
2023-2024
Shenyang University of Technology
2023-2024
Xihua University
2023-2024
Unsupervised domain adaptation (UDA) aims to leverage the knowledge learned from a labeled source dataset solve similar tasks in new unlabeled domain. Prior UDA methods typically require access data when learning adapt model, making them risky and inefficient for decentralized private data. This work tackles practical setting where only trained model is available investigates how we can effectively utilize such without problems. We propose simple yet generic representation framework, named...
Partial person re-identification (re-id) is a challenging problem, where only several partial observations (images) of people are available for matching. However, few studies have provided flexible solutions to identifying in an image containing arbitrary part the body. In this paper, we propose fast and accurate matching method address problem. The proposed leverages Fully Convolutional Network (FCN) generate fix-sized spatial feature maps such that pixel-level features consistent. To match...
Unsupervised domain adaptation (UDA) aims to transfer knowledge from a related but different well-labeled source new unlabeled target domain. Most existing UDA methods require access the data, and thus are not applicable when data confidential shareable due privacy concerns. This paper tackle realistic setting with only classification model available trained over, instead of accessing to, data. To effectively utilize for adaptation, we propose novel approach called Source HypOthesis Transfer...
Domain adaptation (DA) aims to transfer knowledge from a label-rich but heterogeneous domain label-scare domain, which alleviates the labeling efforts and attracts considerable attention. Different previous methods focusing on learning domain-invariant feature representations, some recent present generic semi-supervised (SSL) techniques directly apply them DA tasks, even achieving competitive performance. One of most popular SSL is pseudo-labeling that assigns pseudo labels for each...
A central challenge in training classification models the real-world federated system is learning with non-IID data. To cope this, most of existing works involve enforcing regularization local optimization or improving model aggregation scheme at server. Other also share public datasets synthesized samples to supplement under-represented classes introduce a certain level personalization. Though effective, they lack deep understanding how data heterogeneity affects each layer model. In this...
Domain generalization (DG) is essentially an out-of-distribution problem, aiming to generalize the knowledge learned from multiple source domains unseen target domain. The mainstream leverage statistical models model dependence between data and labels, intending learn representations independent of Nevertheless, are superficial descriptions reality since they only required instead intrinsic causal mechanism. When changes with distribution, statistic may fail generalize. In this regard, we...
DeepFake detection aims to differentiate falsified faces from real ones. Most approaches formulate it as a binary classification problem by solely mining the local artifacts and inconsistencies of face forgery, which neglect relation across regions. Although several recent works explore learning for detection, they overlook propagation relational information lead limited performance gains. To address these issues, this paper provides new perspective formulating graph problem, in each facial...
Unsupervised domain adaptation aims to leverage the labeled source data learn with unlabeled target data. Previous trandusctive methods tackle it by iteratively seeking a low-dimensional projection extract invariant features and obtaining pseudo labels via building classifier on However, they merely concentrate minimizing cross-domain distribution divergence, while ignoring intra-domain structure especially for domain. Even after projection, possible risk factors like imbalanced may still...
Unsupervised domain adaptation challenges the problem of transferring knowledge from a well-labelled source to an unlabelled target domain. Recently, adversarial learning with bi-classifier has been proven effective in pushing cross-domain distributions close. Prior approaches typically leverage disagreement between learn transferable representations, however, they often neglect classifier determinacy domain, which could result lack feature discriminability. In this paper, we present simple...
Conventional domain adaptation methods usually resort to deep neural networks or subspace learning find invariant representations across domains. However, most highly rely on large-size source domains and are computationally expensive train, while always have a quadratic time complexity that suffers from the large size. This paper provides simple efficient solution, which could be regarded as well-performing baseline for tasks. Our method is built upon nearest centroid classifier, seeking...
Domain adaptation (DA) paves the way for label annotation and dataset bias issues by knowledge transfer from a label-rich source domain to related but unlabeled target domain. A mainstream of DA methods is align feature distributions two domains. However, majority them focus on entire image features where irrelevant semantic information, e.g., messy background, inevitably embedded. Enforcing alignments in such case will negatively influence correct matching objects consequently lead...
To ease the burden of labeling, unsupervised domain adaptation (UDA) aims to transfer knowledge in previous and related labeled datasets (sources) a new unlabeled dataset (target). Despite impressive progress, prior methods always need access raw source data develop data-dependent alignment approaches recognize target samples transductive learning manner, which may raise privacy concerns from individuals. Several recent studies resort an alternative solution by exploiting well-trained...
One major issue that challenges person re-identification (Re-ID) is the ubiquitous occlusion over captured persons. There are two main for occluded Re-ID problem, i.e. , interference of noise during feature matching and loss pedestrian information brought by occlusions. In this paper, we propose a new approach called Feature Recovery Transformer (FRT) to address simultaneously, which mainly consists visibility graph recovery transformer. To reduce matching, focus on visible regions appear in...
Class Incremental Learning (CIL) aims at learning a classifier in phase-by-phase manner, which only data of subset the classes are provided each phase. Previous works mainly focus on mitigating forgetting phases after initial one. However, we find that improving CIL its phase is also promising direction. Specifically, experimentally show directly encouraging Learner to output similar representations as model jointly trained all can greatly boost performance. Motivated by this, study...
We present a lattice QCD calculation of the transverse-momentum-dependent wave function (TMDWF) pion using large-momentum effective theory. Numerical simulations are based on one ensemble with <a:math xmlns:a="http://www.w3.org/1998/Math/MathML" display="inline"><a:mrow><a:mn>2</a:mn><a:mo>+</a:mo><a:mn>1</a:mn><a:mo>+</a:mo><a:mn>1</a:mn></a:mrow></a:math> flavors highly improved staggered quarks spacing <c:math xmlns:c="http://www.w3.org/1998/Math/MathML"...
Structure-dynamics correlation is one of the major ongoing debates in glass transition, although a number structural features have been found connected to dynamic heterogeneity different glass-forming colloidal systems. Here, using experiments combined with coarse-grained molecular dynamics simulations, we investigate transition monolayers rough ellipsoids. Compared smooth ellipsoids, surface roughness ellipsoids significantly change nature transition. In particular, find that induced by...
Subretinal fibrosis is a major cause of the poor visual prognosis for patients with neovascular age-related macular degeneration (nAMD). Myofibroblasts originated from retinal pigment epithelial (RPE) cells through epithelial-mesenchymal transition (EMT) contribute to formation. N6-Methyladenosine (m6A) modification has been implicated in EMT process and multiple fibrotic diseases. The role m6A EMT-related subretinal not yet elucidated. In this study, we found that during mouse model...