Semi-Discriminative Representation Loss for Online Continual Learning

TL;DR

This paper introduces SDRL, a simple and efficient loss function that improves online continual learning by balancing discriminative representation learning and gradient diversity, reducing catastrophic forgetting.

Contribution

The paper clarifies the relationship between gradient diversity and discriminative representations, and proposes SDRL to enhance continual learning performance with low computational cost.

Findings

SDRL outperforms state-of-the-art methods on benchmark tasks.

SDRL achieves better accuracy with lower computational cost.

The method effectively balances gradient diversity and discriminativeness.

Abstract

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.