Reversible and mergeable learning for continual learning of artificial neural networks

Abstract

Deep learning with artificial neural networks is one of the most powerful AI technologies. It shows state-of-the-art performance in various fields. To save network training time and improve performance, transfer learning is commonly used. Transfer learning achieves good target performance, but source performance is poor due to catastrophic forgetting. We propose a novel continual learning method called Reversible and Mergeable Learning (RML). The RML method keeps all the original parameters to enable the reversible process of additional training on the target task. After training, the original network and newly learned network can be merged into a network of the same size with the original one. A new network trained with the RML method shows better performance than the finetuning method, and surprisingly, can even outperform the original one on the source task without seeing any source data when the source and target tasks are similar. The RML method can also be used for data incremental learning.