Training an improved recurrent attention model using an alternative reward function

摘要

A recurrent attention model is considered for application to image classification tasks. Modifications to the approach to improve its accuracy are described. Utilization of the reward in the form of negative cross entropy, which increases the informative value of the reinforcement signal, is proposed for neural network training. A deeper architecture of the attention control subnetwork and an asynchronous actor-critic algorithm are also used. Experiments based on the MNIST and CIFAR datasets are conducted, which confirm the effectiveness of the proposed modifications. Experiments using learned classifiers are also conducted, which demonstrate the complexity of the simultaneous attention control and selection of an embedded classifier for analysis of the chosen patch. (C) 2021 Optical Society of America