Toward durable representations for continual learning

Continual learning models are known to suffer from catastrophic forgetting. Existing regularization methods to countering forgetting operate by penalizing large changes to learned parameters. A significant downside to these methods, however, is that, by effectively freezing model parameters, they gradually suspend the capacity of a model to learn new tasks. In this paper, we explore an alternative approach to the continual learning problem that aims to circumvent this downside. In particular, we ask the question: instead of forcing continual learning models to remember the past, can we modify the learning process from the start, such that the learned representations are less susceptible to forgetting? To this end, we explore multiple methods that could potentially encourage durable representations. We demonstrate empirically that the use of unsupervised auxiliary tasks achieves significant reduction in parameter re-optimization across tasks, and consequently reduces forgetting, without explicitly penalizing forgetting. Moreover, we propose a distance metric to track internal model dynamics across tasks, and use it to gain insight into the workings of our proposed approach, as well as other recently proposed methods.