Natural continual learning: success is a journey, not (just) a destination

Biological agents are known to learn many different tasks over the course of\ntheir lives, and to be able to revisit previous tasks and behaviors with little\nto no loss in performance. In contrast, artificial agents are prone to\n'catastrophic forgetting' whereby performance on previous tasks deteriorates\nrapidly as new ones are acquired. This shortcoming has recently been addressed\nusing methods that encourage parameters to stay close to those used for\nprevious tasks. This can be done by (i) using specific parameter regularizers\nthat map out suitable destinations in parameter space, or (ii) guiding the\noptimization journey by projecting gradients into subspaces that do not\ninterfere with previous tasks. However, these methods often exhibit subpar\nperformance in both feedforward and recurrent neural networks, with recurrent\nnetworks being of interest to the study of neural dynamics supporting\nbiological continual learning. In this work, we propose Natural Continual\nLearning (NCL), a new method that unifies weight regularization and projected\ngradient descent. NCL uses Bayesian weight regularization to encourage good\nperformance on all tasks at convergence and combines this with gradient\nprojection using the prior precision, which prevents catastrophic forgetting\nduring optimization. Our method outperforms both standard weight regularization\ntechniques and projection based approaches when applied to continual learning\nproblems in feedforward and recurrent networks. Finally, the trained networks\nevolve task-specific dynamics that are strongly preserved as new tasks are\nlearned, similar to experimental findings in biological circuits.\n