Example forgetting and rehearsal in continual learning

Abstract

A major challenge of training neural networks on different tasks in a sequential manner is catastrophic forgetting, where earlier experiences are forgotten while learning a new one. In recent years, rehearsal-based methods have become popular top-performing alleviation approaches. Rehearsal builds upon maintaining and repeatedly using for training a small buffer of data selected across encountered tasks. In this work, we examine in image classification whether all training examples are forgotten equally and which ones are worth keeping in the memory. Two different statistics of forgettableness are employed to rank examples based on them. We propose a simple strategy for example selection: keeping the least forgettable examples according to precomputed or continually updated forgetting statistics. Despite the simplicity of this method, it achieves better results compared to different memory-management strategies on standard benchmarks.