Accounting for the Effect of Inter-Task Similarity in Continual Learning Models

Catastrophic forgetting has long been a major obstacle to continual learning. In this paper, we explore the effect of certain task characteristics, in particular inter-task similarity, on the extent of forgetting. Moreover, we experimentally study the effect of these characteristics on the effectiveness of recent state-of-the-art continual learning approaches. We show that the performance of some methods, for example, the recently proposed Learning without Forgetting (LwF) and elastic weight consolidation (EWC) models, is significantly dependent on these characteristics. We propose a rehearsal-based extension to continual learning models to address this vulnerability. We develop this extension first in the context of LwF and later demonstrate its effectiveness with other models such as EWC. We show that a memory budget of 1% of training data is sufficient to significantly improve on performance in cases of low inter-task similarity.

I. INTRODUCTION

As the field of deep learning matures, efforts seeking to move deep models beyond learning narrowly defined, isolated tasks and into more general forms of intelligence are increasing. One such form of more general intelligence is the so-called continual learning or lifelong learning frame-work [1]. In this framework, a model learns tasks sequentially over time, and is expected to accumulate knowledge with each new task, much like, say, humans do.