Visualization of Focal Cues for Visuomotor Coordination by Gradient-based Methods: A Recurrent Neural Network Shifts The Attention Depending on Task Requirements

摘要

For an autonomous robot to flexibly move in response to various tasks or environmental changes, an attention mechanism is required that is based on the robot's behavioral experience. In this paper, we visualize how attention is acquired inside a neural network learned using supervised learning and describe how to acquire a suitable representation for performing a task. Our experimental evaluation shows that the attention was automatically acquired for objects that are needed to perform tasks by learning the time-series of both vision and motor information rather than only vision information. By multimodal learning, the attention is robust against unlearned conditions which background changes or obstacles.