ATTENTIVE RECURRENT NEURAL NETWORKS

Abstract

This thesis investigates attentive recurrent neural networks (RNN) and their applications, which allow distillation of valuable information in a similar way to attentions in humans' cognition. It enables RNNs to selectively process different patches of an input sample (intra-sample attention) or distinguish samples that contain more informative resources for learning (inter-sample attention). Particularly, saliency-driven associative memory (SDAM) and content-driven associative memory (CDAM) are designed with visual saliency encoding for attentive storage and retrieval of image patterns. A stochastic attentive RNN called Glance and Glimpse Network (GGNet) is developed for image classification, whose learning rule can be derived by either reinforcement learning with reward shaping or direct probabilistic inference via importance sampling. Finally, Partially Observable Trust Region Policy Optimization (PO-TRPO) algorithm is developed to learn navigation policies for socially concomitant mobile robot navigation. The learning process is facilitated with the Role Playing Learning (RPL) scheme. Both simulative and experimental results are provided.