Xinrui Wang "Beyond Catastrophic Forgetting: Online Continual Learning from Flat Labels to Structured Multimodal Streams"

This talk can be viewed via zoom. (Note: this talk will be recorded)

Abstract:

Today's AI models are trained on static, curated datasets but deployed into dynamic data streams that keep changing. In this talk, I will follow one thread through my PhD work and into my research plan: as the stream's structure grows, from flat labels, through co-occurring and hierarchical labels, to multimodal streams, what does a model learn and what does it forget? I will show that forgetting is not the only failure, since under a single pass a model can also learn too little when the stream moves faster than it can, and can lose old classes abruptly when it relies on too few cues. Once labels carry structure, the problem shifts from acquiring knowledge to organizing it. Furthermore when the stream becomes multimodal, learning on one modality can interfere with another. Overall, I will argue that continual learning is shaped as much by the stream and the model as by the learning rule, and conclude with a plan to study how modern LLMs and agents should evolve online using a dynamic benchmark, a measurement framework, and possible adaptation methods.

Speaker Bio:

Xinrui Wang is a PhD candidate in computer science at Nanjing University of Aeronautics and Astronautics and a visiting researcher in the LAMP group at the Computer Vision Center in Barcelona. His research studies how models safely acquire, organize, and retain knowledge in non-stationary data streams and incomplete supervisions. He has published over 10 papers in major machine-learning venues like ICML, NeurIPS, AISTATS and AAAI.