The Perception-Action Loop in a Predictive Agent

AbstractWe propose an agent model consisting of perceptual and proprioceptive pathways. It actively samples a sequence of percepts from its environment using the perception-action loop. The model predicts to complete the partial percept and propriocept sequences observed till each sampling instant, and learns where and what to sample from the prediction error, without supervision or reinforcement. The model is exposed to two kinds of stimuli: images of fully-formed handwritten numerals/alphabets, and videos of gradual formation of numerals. For each object class, the model learns a set of salient locations to attend to in images and a policy consisting of a sequence of eye fixations in videos. Behaviorally, the same model gives rise to saccades while observing images and tracking while observing videos. The proposed agent is the first of its kind to interact with and learn end-to-end from static and dynamic environments to generate realistic handwriting with state-of-the-art performance.

Return to previous page