Continual Learning with Scaled Gradient Projection

TL;DR

This paper introduces Scaled Gradient Projection (SGP), a novel continual learning method that balances learning new tasks and retaining old ones by combining orthogonal gradient updates with scaled steps based on task importance.

Contribution

The paper proposes SGP, which improves continual learning by integrating scaled gradient steps with orthogonal projections, using a new importance measure derived from SVD of input representations.

Findings

SGP outperforms state-of-the-art methods in image classification tasks.

SGP achieves better retention of old tasks with less training overhead.

The importance measure effectively guides gradient scaling.

Abstract

In neural networks, continual learning results in gradient interference among sequential tasks, leading to catastrophic forgetting of old tasks while learning new ones. This issue is addressed in recent methods by storing the important gradient spaces for old tasks and updating the model orthogonally during new tasks. However, such restrictive orthogonal gradient updates hamper the learning capability of the new tasks resulting in sub-optimal performance. To improve new learning while minimizing forgetting, in this paper we propose a Scaled Gradient Projection (SGP) method, where we combine the orthogonal gradient projections with scaled gradient steps along the important gradient spaces for the past tasks. The degree of gradient scaling along these spaces depends on the importance of the bases spanning them. We propose an efficient method for computing and accumulating importance of these bases using the singular value decomposition of the input representations for each task. We conduct extensive experiments ranging from continual image classification to reinforcement learning tasks and report better performance with less training overhead than the state-of-the-art approaches.