Restricted Orthogonal Gradient Projection for Continual Learning

TL;DR

This paper introduces ROGO, a novel gradient projection method for continual learning that improves forward knowledge transfer using a fixed network, without additional data buffers or parameters.

Contribution

ROGO employs a restricted orthogonal constraint to enhance forward transfer in continual learning with a fixed architecture, providing theoretical guarantees.

Findings

ROGO outperforms several strong baselines in experiments.

The framework requires no data buffers or extra parameters.

Theoretical guarantees support the effectiveness of the relaxing strategy.

Abstract

Continual learning aims to avoid catastrophic forgetting and effectively leverage learned experiences to master new knowledge. Existing gradient projection approaches impose hard constraints on the optimization space for new tasks to minimize interference, which simultaneously hinders forward knowledge transfer. To address this issue, recent methods reuse frozen parameters with a growing network, resulting in high computational costs. Thus, it remains a challenge whether we can improve forward knowledge transfer for gradient projection approaches using a fixed network architecture. In this work, we propose the Restricted Orthogonal Gradient prOjection (ROGO) framework. The basic idea is to adopt a restricted orthogonal constraint allowing parameters optimized in the direction oblique to the whole frozen space to facilitate forward knowledge transfer while consolidating previous knowledge. Our framework requires neither data buffers nor extra parameters. Extensive experiments have demonstrated the superiority of our framework over several strong baselines. We also provide theoretical guarantees for our relaxing strategy.