SingularClip: Preventing Spectral Collapse to Maintain Plasticity in Continual and Reinforcement Learning

2026-08-18

AI Summary

研究通过定期裁剪权重矩阵的奇异值，解决了神经网络在非平稳任务中因奇异值各向异性增长导致的塑性丧失问题。

Abstract

Neural networks trained on nonstationary tasks frequently lose the ability to fit new targets, a phenomenon referred to as loss of plasticity. We identify a novel source of plasticity loss due to the growing anisotropy of weight matrices' singular values during training, and analyze this phenomenon both empirically and theoretically. To mitigate this issue, we introduce SingularClip, a procedure that periodically clips the singular values of all weight matrices. We show that SingularClip performs strongly against baselines across a range of tasks in both continual supervised learning and deep reinforcement learning.

Problem

• plasticity loss
• nonstationary tasks
• singular values anisotropy
• neural networks

Innovation

• SingularClip
• plasticity
• continual learning
• reinforcement learning
• singular values