Semi-Discriminative Representation Loss for Online Continual Learning
This work addresses catastrophic forgetting for machine learning systems that need to learn continuously from streaming data, presenting an incremental improvement over existing gradient-based methods.
The paper tackled the problem of catastrophic forgetting in online continual learning by analyzing the trade-offs between gradient diversity and representation discriminativeness, proposing a Semi-Discriminative Representation Loss (SDRL) method that achieves better performance with low computational cost on multiple benchmark tasks.
The use of episodic memory in continual learning has demonstrated effectiveness for alleviating catastrophic forgetting. In recent studies, gradient-based approaches have been developed to make more efficient use of compact episodic memory. Such approaches refine the gradients resulting from new samples by those from memorized samples, aiming to reduce the diversity of gradients from different tasks. In this paper, we clarify the relation between diversity of gradients and discriminativeness of representations, showing shared as well as conflicting interests between Deep Metric Learning and continual learning, thus demonstrating pros and cons of learning discriminative representations in continual learning. Based on these findings, we propose a simple method -- Semi-Discriminative Representation Loss (SDRL) -- for continual learning. In comparison with state-of-the-art methods, SDRL shows better performance with low computational cost on multiple benchmark tasks in the setting of online continual learning.