Learn to Grow: A Continual Structure Learning Framework for Overcoming\n Catastrophic Forgetting

Abstract

Addressing catastrophic forgetting is one of the key challenges in continual\nlearning where machine learning systems are trained with sequential or\nstreaming tasks. Despite recent remarkable progress in state-of-the-art deep\nlearning, deep neural networks (DNNs) are still plagued with the catastrophic\nforgetting problem. This paper presents a conceptually simple yet general and\neffective framework for handling catastrophic forgetting in continual learning\nwith DNNs. The proposed method consists of two components: a neural structure\noptimization component and a parameter learning and/or fine-tuning component.\nBy separating the explicit neural structure learning and the parameter\nestimation, not only is the proposed method capable of evolving neural\nstructures in an intuitively meaningful way, but also shows strong capabilities\nof alleviating catastrophic forgetting in experiments. Furthermore, the\nproposed method outperforms all other baselines on the permuted MNIST dataset,\nthe split CIFAR100 dataset and the Visual Domain Decathlon dataset in continual\nlearning setting.\n