Task 008: Closed-Loop Memory GAN for Continual Learning
|Event Date:||July 11, 2019|
|Time:||2:00pm ET/ 11:00am PT
|School or Program:||Electrical and Computer Engineering
Sequential learning of tasks using gradient descent leads to an unremitting decline in the accuracy of tasks for which training data is no longer available, termed catastrophic forgetting. Generative models have been explored as a means to approximate the distribution of old tasks and bypass storage of real data. Here we propose a cumulative closed-loop memory replay GAN (CloGAN) provided with external regularization by a small memory unit selected for maximum sample diversity. We evaluate incremental class learning using a notoriously hard paradigm, “single-headed learning,” in which each task is a disjoint subset of classes in the overall dataset, and performance is evaluated on all previous classes. First, we show that when constructing a dynamic memory unit to preserve sample heterogeneity, model performance asymptotically approaches training on the full dataset. We then show that using a stochastic generator to continuously output fresh new images during training increases performance significantly further meanwhile generating quality images. We compare our approach to several baselines including fine-tuning by gradient descent (FGD), Elastic Weight Consolidation (EWC), Deep Generative Replay (DGR) and Memory Replay GAN(MeRGAN). Our method has very low long-term memory cost, the memory unit, as well as negligible intermediate memory storage.
Amanda Rios is a PhD student studying Computational Neuroscience at the University of Southern California, advised by Dr. Laurent Itti and Dr. Bartlett Mell. Before commencing her PhD she received her bachelor’s degree in Physics and Molecular Biology from the University of Sao Paulo, Brazil. Amanda is currently researching Continual Learning for AI as well as a Neuromorphic inspired Computer Vision. She is very interested in the overarching theme of approximating AI to neurobiological processing behavior.