A5055359094- Machine Learning and Data Classification
- scientometrics and bibliometrics research
- Text and Document Classification Technologies
- Domain Adaptation and Few-Shot Learning
- Anomaly Detection Techniques and Applications
- Advanced Graph Neural Networks
- Meta-analysis and systematic reviews
- Reinforcement Learning in Robotics
- Adversarial Robustness in Machine Learning
- Topic Modeling
- Data Stream Mining Techniques
- Mobile Crowdsensing and Crowdsourcing
- Scientific Computing and Data Management
- Electricity Theft Detection Techniques
- Simulation Techniques and Applications
- Traffic Prediction and Management Techniques
- Cognitive Science and Mapping
- Forecasting Techniques and Applications
- Robotic Path Planning Algorithms
- Imbalanced Data Classification Techniques
- Autonomous Vehicle Technology and Safety
- Machine Learning and Algorithms
- Bioinformatics and Genomic Networks
- Graph Theory and Algorithms
- Artificial Intelligence in Games
Tsinghua University
2025
National University of Defense Technology
2021-2024
The foundation model enables fast problem-solving without learning from scratch, and such a desirable adaptation property benefits its adopted cross-task generalization paradigms, e.g., pretraining, meta-training, or finetuning. Recent trends have focused on the curation of task datasets during optimization, which includes selection as an indispensable consideration for either robustness sampling efficiency purposes. Despite some progress, selecting crucial batches to optimize over iteration...
Task robust adaptation is a long-standing pursuit in sequential decision-making. Some risk-averse strategies, e.g., the conditional value-at-risk principle, are incorporated domain randomization or meta reinforcement learning to prioritize difficult tasks optimization, which demand costly intensive evaluations. The efficiency issue prompts development of active task sampling train adaptive policies, where risk-predictive models used surrogate policy evaluation. This work characterizes...
<title>Abstract</title> Foundation models have revolutionized general-purpose problem-solving, offering rapidtask adaptation through pretraining, meta-training, and finetuning. Recent crucial advances in these paradigms reveal the importance of challenging taskprioritized sampling to enhance robustness under distribution shifts. However, ranking task difficulties over iteration as a preliminary step typically requiresexhaustive evaluation, which is practically unaffordable computation...
The recent progress in graph representation learning boosts the development of many classification tasks, such as protein and social network classification. One mainstream approaches for is hierarchical pooling method. It learns by gradually reducing scale graph, so it can be easily adapted to large-scale graphs. However, existing methods discard original structure during downsizing resulting a lack topological structure. In this paper, we propose multi-scale neural (MSGNN) model that not...
Meta learning is a promising paradigm to enable skill transfer across tasks. Most previous methods employ the empirical risk minimization principle in optimization. However, resulting worst fast adaptation subset of tasks can be catastrophic risk-sensitive scenarios. To robustify adaptation, this paper optimizes meta pipelines from distributionally robust perspective and trains models with measure expected tail risk. We take two-stage strategy as heuristics solve problem, controlling cases...
Meta-learning is a practical learning paradigm to transfer skills across tasks from few examples. Nevertheless, the existence of task distribution shifts tends weaken meta-learners' generalization capability, particularly when naively hand-crafted or based on simple priors that fail cover typical scenarios sufficiently. Here, we consider explicitly generative modeling distributions placed over identifiers and propose robustifying fast adaptation adversarial training. Our approach, which can...
Internet services have led to the eruption of traffic, and machine learning on these data has become an indispensable tool, especially when application is risk-sensitive. This paper focuses network traffic classification in presence class imbalance, which fundamentally ubiquitously exists analysis. existence imbalance mostly drifts optimal decision boundary, resulting a less solution for models. To alleviate effect, we propose design strategies alleviating through lens group distributionally...
Meta learning is a promising paradigm in the era of large models and task distributional robustness has become an indispensable consideration real-world scenarios. Recent advances have examined effectiveness tail risk minimization fast adaptation improvement \citep{wang2023simple}. This work contributes to more theoretical investigations practical enhancements field. Specifically, we reduce distributionally robust strategy max-min optimization problem, constitute Stackelberg equilibrium as...
The concept of GenAI has been developed for decades. Until recently, it impressed us with substantial breakthroughs in natural language processing and computer vision, actively engaging industrial scenarios. Noticing the practical challenges, e.g., limited learning resources, overly dependencies on scientific discovery empiricism, we nominate large-scale generative simulation artificial intelligence (LS-GenAI) as next hotspot to connect.
Academic administrators and funding agencies must predict the publication productivity of research groups individuals to assess authors' abilities. However, such prediction remains an elusive task due randomness individual diversity patterns. We applied two kinds approaches this task: deep neural network learning model-based approaches. found that a cannot give good long-term for groups, while provide short-term predictions individuals. proposed model integrates advantages both data-driven...