Triple-Memory Networks: A Brain-Inspired Method for Continual Learning

• Published in: IEEE transaction on neural networks and learning systems Vol. 33; no. 5; pp. 1925 - 1934
• Main Authors: Wang, Liyuan, Lei, Bo, Li, Qian, Su, Hang, Zhu, Jun, Zhong, Yi
• Format: Journal Article
• Language: English
• Published: United States IEEE 01.05.2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Artificial neural networks; Benchmarks; Biological effects; Biological neural networks; Brain; Brain modeling; Brain-inspired algorithm; catastrophic forgetting; Cerebral cortex; Classifiers; Computer architecture; continual learning; deep learning; Generative adversarial networks; Hippocampus; Information processing; Knowledge; Learning; Life sciences; Neocortex; Neural networks; Neural Networks, Computer; Prefrontal cortex; Synapses; Task analysis; Training data
• ISSN: 2162-237X; 2162-2388; 2162-2388
• DOI: 10.1109/TNNLS.2021.3111019

Continual acquisition of novel experience without interfering with previously learned knowledge, i.e., continual learning, is critical for artificial neural networks, while limited by catastrophic forgetting. A neural network adjusts its parameters when learning a new task but then fails to conduct the old tasks well. By contrast, the biological brain can effectively address catastrophic forgetting through consolidating memories as more specific or more generalized forms to complement each other, which is achieved in the interplay of the hippocampus and neocortex, mediated by the prefrontal cortex. Inspired by such a brain strategy, we propose a novel approach named triple-memory networks (TMNs) for continual learning. TMNs model the interplay of the three brain regions as a triple-network architecture of generative adversarial networks (GANs). The input information is encoded as specific representations of data distributions in a generator, or generalized knowledge of solving tasks in a discriminator and a classifier, with implementing appropriate brain-inspired algorithms to alleviate catastrophic forgetting in each module. Particularly, the generator replays generated data of the learned tasks to the discriminator and the classifier, both of which are implemented with a weight consolidation regularizer to complement the lost information in the generation process. TMNs achieve the state-of-the-art performance of generative memory replay on a variety of class-incremental learning benchmarks on MNIST, SVHN, CIFAR-10, and ImageNet-50.