A5100458749- Domain Adaptation and Few-Shot Learning
- Multimodal Machine Learning Applications
- Advanced Image and Video Retrieval Techniques
- Machine Learning and Data Classification
- Topic Modeling
- Advanced Neural Network Applications
- Data Stream Mining Techniques
- Reinforcement Learning in Robotics
- EEG and Brain-Computer Interfaces
- Neural Networks and Applications
- Advanced Algorithms and Applications
- Adversarial Robustness in Machine Learning
- Natural Language Processing Techniques
- Text and Document Classification Technologies
- Chaos control and synchronization
- Machine Learning and ELM
- Human-Automation Interaction and Safety
- Explainable Artificial Intelligence (XAI)
- Generative Adversarial Networks and Image Synthesis
- Advanced Bandit Algorithms Research
- Neural Networks and Reservoir Computing
- AI in Service Interactions
- Digital Imaging for Blood Diseases
- Anomaly Detection Techniques and Applications
- COVID-19 diagnosis using AI
Beijing Institute of Technology
2022-2025
China University of Mining and Technology
2011-2024
Berkeley College
2024
University of California, Berkeley
2020-2024
Chinese Academy of Sciences
2023-2024
Center for Excellence in Brain Science and Intelligence Technology
2024
University of Chinese Academy of Sciences
2024
Shandong Institute of Automation
2024
Institute of Automation
2023
Fuzhou University
2023
Wei Li, Can Gao, Guocheng Niu, Xinyan Xiao, Hao Liu, Jiachen Hua Wu, Haifeng Wang. Proceedings of the 59th Annual Meeting Association for Computational Linguistics and 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.
As a common problem in the visual world, contextual bias means recognition may depend on co-occurrence context rather than objects themselves, which is even more severe multi-label tasks due to multiple targets and absence of location. Although some studies have focused tackling problem, removing negative effect still challenging because it difficult obtain representation bias. In this paper, we propose simple but effective framework employing causal inference mitigate We first present...
Contrastive learning and supervised have both seen significant progress success. However, thus far they largely been treated as two separate objectives, brought together only by having a shared neural network. In this paper we show that through the perspective of hybrid discriminative-generative training energy-based models can make direct connection between contrastive learning. Beyond presenting unified view, our specific choice approximation loss outperforms existing practice in terms...
Deep Reinforcement Learning (RL) has emerged as a powerful paradigm to solve range of complex yet specific control tasks. Yet training generalist agents that can quickly adapt new tasks remains an outstanding challenge. Recent advances in unsupervised RL have shown pre-training with self-supervised intrinsic rewards result efficient adaptation. However, these algorithms been hard compare and develop due the lack unified benchmark. To this end, we introduce Unsupervised Benchmark (URLB). URLB...
X‐ray photoelectron spectroscopy (XPS) is widely used to determine the composition of elements and oxygen‐containing functional groups. Accurate determination core‐electron binding energies (CEBEs) essential for identifying types contents oxygen‐functional groups present on surface coal. The △SCF method has been carried out CEBEs an extensive series organic systems encompassing most common functionalities. polymers which experimental values are available comparison demonstrate adequacy...
This work considers the out-of-distribution (OOD) prediction problem where (1)~the training data are from multiple domains and (2)~the test domain is unseen in training. DNNs fail OOD because they prone to pick up spurious correlations. Recently, Invariant Risk Minimization (IRM) proposed address this issue. Its effectiveness has been demonstrated colored MNIST experiment. Nevertheless, we find that performance of IRM can be dramatically degraded under \emph{strong $\Lambda$ spuriousness} --...
Learning multiple tasks sequentially without forgetting previous knowledge, called Continual Learning(CL), remains a long-standing challenge for neural networks. Most existing methods rely on additional network capacity or data replay. In contrast, we introduce novel approach which refer to as Recursive Gradient Optimization(RGO). RGO is composed of an iteratively updated optimizer that modifies the gradient minimize replay and virtual Feature Encoding Layer(FEL) represents different...
Meta-learning approaches have recently achieved promising performance in multi-class incremental learning. However, meta-learners still suffer from catastrophic forgetting, i.e., they tend to forget the learned knowledge old tasks when focus on rapidly adapting new classes of current task. To solve this problem, we propose a novel distilled meta-learning (DML) framework for learning that integrates seamlessly with distillation each stage. Specifically, during inner-loop training, is...
This article presents a new adaptive metric distillation approach that can significantly improve the student networks' backbone features, along with better classification results. Previous knowledge (KD) methods usually focus on transferring across classifier logits or feature structure, ignoring excessive sample relations in space. We demonstrated such design greatly limits performance, especially for retrieval task. The proposed collaborative (CAMD) has three main advantages: 1)...
Traditional rehabilitation exercises take a long time and require high training intensity. Moreover, repeating the same boring action will reduce enthusiasm of patient to participate in training. A task-oriented virtual reality system for people with certain athletic ability is designed developed this paper. It can improve efficiency fun The uses MEMS sensors capture human motion real time, body identified by constructing wireless domain network based on Zigbee technology. With Oculus Rift S...
Abstract Sustained attention, as the basis of general cognitive ability, naturally varies across different time scales, spanning from hours, e.g. wakefulness to drowsiness state, seconds, trial-by-trail fluctuation in a task session. Whether there is unified mechanism underneath such trans-scale variability remains unclear. Here we show that cortical excitation/inhibition (E/I) strong modulator sustained attention humans scales. First, observed ability attend varied brain states...
Recently, transformer-based image captioning models have achieved significant performance improvement. However, due to the limitations of region visual features and deterministic projections between space caption space, existing methods still suffer from disentangled rigid sentences. To address these issues, we first introduce panoptic segmentation extract features, which can effectively alleviate confusion caused by widely-adopted features. Then, propose a based sequential conditional...
BERT, as a pre-trained model, can not only greatly improve the performance of task models in field language processing, but also save computational resources and costs. At present, most sentiment classification tasks focus on model structures. But it is very important to explore changes hyper-parameters based so obtain general parameters setting method for improving accuracy models' predictions. In this paper, we conduct two parameter fine-tuning methods: static fine-tune used models, then...
Label noise and class imbalance are two major issues coexisting in real-world datasets. To alleviate the issues, state-of-the-art methods reweight each instance by leveraging a small amount of clean unbiased data. Yet, these overlook class-level information within instance, which can be further utilized to improve performance. this end, paper, we propose Generalized Data Weighting (GDW) simultaneously mitigate label manipulating gradients at level. specific, GDW unrolls loss gradient chain...
As a challenging computer vision task, fine-grained visual categorization (FGVC) has received more and attention. Most classification algorithms are data-hungry time-consuming, on the contrary, humans have ability to recognize different species with only small number of samples. For example, baby can identity kinds dogs few images. Therefore, in this paper, fined-grained model which is based meta learning proposed, it be used for accurate data. Meta uses meta-learners learn relative tasks so...