Hebbian motor control in a robot-embedded model of habituation

摘要

Two experiments using a mobile robot examine the performance of a neural network model of habituation. The input to the network is the video feed from the robot's camera, preprocessed to model the visual system. Images, after retinal processing, are translated in the frequency domain and Gabor-filtered. The output of the network controls the robot's motors and thus where it looks. In one condition, network output directly controls the motors. In a second condition, network outputs are connected to control units via weights that are modified with simple Hebbian learning. In both cases, the robot's behavior reproduces the important familiarity-to-novelty shift observed in human infants. Hebbian learning, however, helps to increase and stabilize novelty preference.