Rehearsal-Free Modular and Compositional Continual Learning for Language Models

TL;DR

This paper introduces MoCL, a rehearsal-free modular framework for continual learning in language models that adds and composes modules to transfer knowledge without privacy or memory issues.

Contribution

MoCL is a novel modular and compositional approach that enables rehearsal-free continual learning with improved knowledge transfer in language models.

Findings

MoCL outperforms state-of-the-art methods on various benchmarks.

MoCL effectively facilitates knowledge transfer between tasks.

MoCL avoids privacy and memory issues associated with rehearsal-based methods.

Abstract

Continual learning aims at incrementally acquiring new knowledge while not forgetting existing knowledge. To overcome catastrophic forgetting, methods are either rehearsal-based, i.e., store data examples from previous tasks for data replay, or isolate parameters dedicated to each task. However, rehearsal-based methods raise privacy and memory issues, and parameter-isolation continual learning does not consider interaction between tasks, thus hindering knowledge transfer. In this work, we propose MoCL, a rehearsal-free Modular and Compositional Continual Learning framework which continually adds new modules to language models and composes them with existing modules. Experiments on various benchmarks show that MoCL outperforms state of the art and effectively facilitates knowledge transfer.