Preempting Catastrophic Forgetting in Continual Learning Models by Anticipatory Regularization
Alaa El Khatib, Fakhri Karray · 2019
Neural networks trained on tasks sequentially tend to degrade in performance, on the average, the more tasks they see, as the representations learned for one task get progressively modified while learning subsequent tasks. This phenomenon- known as catastrophic forgetting-is a major obstacle on the road toward designing agents that can continually learn new concepts and tasks the way, say, humans do. A common approach to containing catastrophic forgetting is to use regularization to slow down learning on weights deemed important to previously learned tasks. We argue in this paper that, on their own, such post hoc measures to safeguard what has been learned can, even in their more sophisticated variants, paralyze the network and degrade its capacity to learn and counter forgetting as the number of tasks learned increases. We propose instead- or possibly in conjunction-that, in anticipation of future tasks, regularization be applied to drive the optimization of network weights toward reusable solutions. We show that one way to achieve this is through an auxiliary unsupervised reconstruction loss that encourages the learned representations not only to be useful for solving, say, the current classification task, but also to reflect the content of the data being processed-content that is generally richer than it is discriminative for any one task. We compare our approach to the recent elastic weight consolidation regularization approach, and show that, although we do not explicitly try to preserve important weights or pass on any information about the data distribution of learned tasks, our model is comparable in performance, and in some cases better.