Our research centers on Edge AI, aiming to build personalized, trustworthy, and efficient AI that meets the needs of diverse users in real-world environments. These directions are driven by our interest in interpreting neural representations to understand and control model behavior. We develop learning and inference algorithms that enable AI services across edge devices and collaborative computing systems, addressing challenges arising from heterogeneous data, evolving user needs, safety requirements, and limited resources.
We pursue these directions across diverse domains, including large language and multimodal models, image/video understanding and generation, and embodied AI (e.g., vision-language-action models).
Our Approach
Our approach brings together neural representation interpretability, mathematical foundations, and AI systems. We investigate how neural networks represent and process information, and how their internal representations influence model behavior and generalization. By connecting theoretical analysis with empirical investigation, we use these insights to develop principled learning and inference algorithms for personalized, trustworthy, and efficient AI.
AI systems must adapt to diverse users, downstream tasks, and continually changing environments. We study how models can learn collaboratively from distributed and heterogeneous data without requiring users or institutions to share their private data. We also investigate how foundation models can adapt to individual needs and acquire new knowledge while retaining previously learned capabilities. Our current research focuses on collaborative/federated learning, continual learning, and LLM post-training, enabling AI to evolve with users and their environments.
AI systems should remain safe and reliable as they are trained, adapted, and deployed. We investigate how to preserve safety during fine-tuning and compression, align predictive confidence with actual correctness, and selectively remove unwanted knowledge. Our current research spans LLM safety, LLM confidence/uncertainty, and machine unlearning, with an emphasis on maintaining trustworthiness as models and data evolve.
Deploying capable AI in resource-constrained environments requires efficient use of data, computation, memory, and communication. We develop algorithms and systems that reduce the cost of training and inference while preserving model capabilities, with particular attention to the constraints of edge devices. Our research includes parameter-efficient fine-tuning, model compression/quantization, and collaborative/adaptive inference, targeting on-device AI.
Finally, we have also been exploring the above aspects of AI/ML over edge/communication/wireless networks. We explore how communication and computation can jointly support personalized, trustworthy, and efficient AI services at the network edge, addressing the growing demand for intelligent services in real-world environments.