Lifelong Learning Process: Self-Memory Supervising and Dynamically Growing Networks

From childhood to youth, human gradually come to know the world. But for neural networks, this growing process seems difficult. Trapped in catastrophic forgetting, current researchers feed data of all categories to a neural network which remains the same structure in the whole training process. We compare this training process with human learing patterns, and find two major conflicts. In this paper, we study how to solve these conflicts on generative models based on the conditional variational autoencoder(CVAE) model. To solve the uncontinuous conflict, we apply memory playback strategy to maintain the model's recognizing and generating ability on invisible used categories. And we extend the traditional one-way CVAE to a circulatory mode to better accomplish memory playback strategy. To solve the `dead' structure conflict, we rewrite the CVAE formula then are able to make a novel interpretation about the funtions of different parts in CVAE models. Based on the new understanding, we find ways to dynamically extend the network structure when training on new categories. We verify the effectiveness of our methods on MNIST and Fashion MNIST and display some very insteresting results.
Comment: 8 pages, 12 figures