PhD Dissertation Proposal by LUO Zilin | Continual Learning for Deep Models

Continual Learning for Deep Models

ABOUT THE TALK

Continual Learning (CL) is a deep learning task aimed at training models to continually learn new knowledge while preventing the forgetting of old knowledge. In this proposal, I will share two research works that discuss this problem from memory-based viewpoints. In the first work, the utilization of memory is considered by compressing exemplars through downsampling non-discriminative pixels, thereby preserving more semantic information. In the second work, the "phase" concept widely used in CL is reviewed and deemed unrealistic for two reasons: in-phase data stationarity and class balance in incoming data. To tackle the issues, per-step distribution shifts are introduced (thus eliminating phases) in the stream generation process. Under this setting, a coreset selector based on stream rates is proposed, presenting an adaptive and robust criterion.

SPEAKER BIOGRAPHY

Zilin LUO is a PhD Candidate in Computer Science at the SMU School of Computing and Information Systems, supervised by Prof. Qianru SUN. His research is focused on Continual Learning and In-Context Learning.