Meta-learning memory representations for continual learning with scalar-output networks

Abstract

Neural networks have achieved remarkable success in various areas, but they still fail when the task changes over time. The subfield of machine learning that focuses on overcoming this problem is called continual learning. In this paper, we applied meta‐learning techniques to train a continual learning agent that utilizes an external memory via a novel scalar‐output architecture. Unlike previous works that use a fixed method to access and update memory, our agent learns the optimal way from scratch. We show that our agent successfully learned what to remember and how to recall, achieving final average accuracy about 10 % higher compared to other continual learning algorithms.