Regularizing Trajectories to Mitigate Catastrophic Forgetting
Regularization-based continual learning approaches generally prevent catastrophic forgetting by augmenting the training loss with an auxiliary objective. However in most practical optimization scenarios with noisy data and/or gradients, it is possible that stochastic gradient descent can inadvertently change critical parameters. In this paper, we argue for the importance of regularizing optimization trajectories directly. We derive a new co-natural gradient update rule for continual learning whereby the new task gradients are preconditioned with the empirical Fisher information of previously learnt tasks. We show that using the co-natural gradient systematically reduces forgetting in continual learning. Moreover, it helps combat overfitting when learning a new task in a low resource scenario.
PDF Abstract