Weight Clipping for Deep Continual and Reinforcement Learning

TL;DR

This paper proposes a simple weight clipping technique to improve deep continual and reinforcement learning by preventing weight explosion, enhancing generalization, and maintaining plasticity without altering the core optimizer or architecture.

Contribution

Introducing a straightforward weight clipping method that can be integrated into existing systems to address weight growth issues in deep learning.

Findings

Weight clipping improves generalization in deep learning models.

It helps prevent policy collapse in reinforcement learning.

Clipping facilitates learning with high replay ratios.

Abstract

Many failures in deep continual and reinforcement learning are associated with increasing magnitudes of the weights, making them hard to change and potentially causing overfitting. While many methods address these learning failures, they often change the optimizer or the architecture, a complexity that hinders widespread adoption in various systems. In this paper, we focus on learning failures that are associated with increasing weight norm and we propose a simple technique that can be easily added on top of existing learning systems: clipping neural network weights to limit them to a specific range. We study the effectiveness of weight clipping in a series of supervised and reinforcement learning experiments. Our empirical results highlight the benefits of weight clipping for generalization, addressing loss of plasticity and policy collapse, and facilitating learning with a large replay ratio.