CGR: Confidence-Guided Replay for Buffer-Based Continual Learning

Abstract: Continual Learning (CL) aims to acquire new knowledge while preserving previously learned information without catastrophic forgetting. Buffer-based methods, which retain samples from past tasks, have demonstrated promising results; however, efficiently allocating limited buffer space remains a significant challenge. Recent studies often either neglect the varying impact individual samples have on the learning process or incur high computational costs to identify informative replay samples. To overcome these limitations, we propose a novel approach called Confidence-Guided Replay (CGR), a lightweight buffer policy for offline, task-aware continual supervised classification that dynamically allocates the buffer by monitoring confidence fluctuations in the main continual learner model. Leveraging measures of sample contribution and difficulty, CGR adaptively prioritizes highly informative samples within the buffer, enhancing knowledge retention and utilization efficiency. Our approach provides a flexible solution for dynamic buffer allocation, effectively addressing the varying importance and learning complexity of samples over time, and improves CL performance.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Martha_White1

Submission Number: 7632