Dual network based complementary learning system for continual learning

Abstract

Neural networks have surpassed human performance at many individual tasks. However, when it comes to learning in a continual setting, where data comes in non-independent and identically distributed streams, they suffer heavily from a phenomenon called catastrophic forgetting, the loss of the information of past tasks with the learning of a new task. This is in contrast with how humans learn as they utilize past knowledge to learn new knowledge more efficiently. Overcoming catastrophic forgetting while facilitating future learning is one of the most challenging issues in machine learning currently. To address this issue, we propose a brain-inspired complementary dual network model, comprising a fast learner and a slow consolidator. The fast learner first adapts to a new task that is seen only once, and the slow consolidator then distills the new task information from the fast learner using knowledge distillation. The two networks are trained in an alternate manner. To consolidate the learning of new task with the learning of past tasks, we employ a small memory of each task for replay during the training of the slow consolidator. In addition, we implement a context-based gating mechanism on the slow consolidator, and empirically prove its positive impact on the performance of the proposed model. We show the results of our model on MNIST Permutations, MNIST Rotations, MNIST Split and Incremental CIFAR100 datasets.