Autonomous, self-calibrating binocular vision based on learned attention and active efficient coding

摘要

We present a self-calibrating binocular vision system that autonomously learns how to encode the visual input and how to move its eyes. The model combines the learning of disparity representations and vergence eye movements through Active Efficient Coding (AEC) and the learning of saccades to interesting targets through two novel attention models. The first model is an extension of the Attention based on Information Maximization (AIM) model by Bruce and Tsotsos to binocular images. The second model aims to directly maximize the learning progress of the AEC model. We demonstrate that both attention models improve learning speed compared to a random gaze control strategy. Notably, the vergence eye movement controller and the two attention mechanisms controlling saccades all use the same learned sparse image encoding. The system represents a step towards building self-calibrating, infant-like robots that autonomously learn how to make sense of their environment and how to interact with it.