Understanding Catastrophic Forgetting and Remembering in Continual Learning with Optimal Relevance Mapping

Catastrophic forgetting in neural networks is a significant problem for\ncontinual learning. A majority of the current methods replay previous data\nduring training, which violates the constraints of an ideal continual learning\nsystem. Additionally, current approaches that deal with forgetting ignore the\nproblem of catastrophic remembering, i.e. the worsening ability to discriminate\nbetween data from different tasks. In our work, we introduce Relevance Mapping\nNetworks (RMNs) which are inspired by the Optimal Overlap Hypothesis. The\nmappings reflects the relevance of the weights for the task at hand by\nassigning large weights to essential parameters. We show that RMNs learn an\noptimized representational overlap that overcomes the twin problem of\ncatastrophic forgetting and remembering. Our approach achieves state-of-the-art\nperformance across all common continual learning datasets, even significantly\noutperforming data replay methods while not violating the constraints for an\nideal continual learning system. Moreover, RMNs retain the ability to detect\ndata from new tasks in an unsupervised manner, thus proving their resilience\nagainst catastrophic remembering.\n