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

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