Memory Bounds for Continual Learning

TL;DR

This paper explores the fundamental memory requirements for continual learning within a PAC framework, revealing inherent linear memory growth and proposing an efficient algorithm with logarithmic passes.

Contribution

It introduces a complexity-theoretic analysis of continual learning, establishing linear memory lower bounds and presenting a multiplicative weights update algorithm for improved efficiency.

Findings

Memory must grow linearly with the number of tasks

An algorithm with logarithmic passes over tasks is proposed

Improper learning is necessary for the proposed algorithm's performance

Abstract

Continual learning, or lifelong learning, is a formidable current challenge to machine learning. It requires the learner to solve a sequence of $k$ different learning tasks, one after the other, while retaining its aptitude for earlier tasks; the continual learner should scale better than the obvious solution of developing and maintaining a separate learner for each of the $k$ tasks. We embark on a complexity-theoretic study of continual learning in the PAC framework. We make novel uses of communication complexity to establish that any continual learner, even an improper one, needs memory that grows linearly with $k$, strongly suggesting that the problem is intractable. When logarithmically many passes over the learning tasks are allowed, we provide an algorithm based on multiplicative weights update whose memory requirement scales well; we also establish that improper learning is necessary for such performance. We conjecture that these results may lead to new promising approaches to continual learning.