Sensori-motor networks vs neural networks for visual stimulus prediction

This paper focuses on a recently developed special type of biologically inspired architecture, which we denote as a sensori-motor network, able to co-develop sensori-motor structures directly from the data acquired by a robot interacting with its environment. Such networks learn efficient internal models of the sensori-motor system, developing simultaneously sensor and motor representations (receptive fields) adapted to the robot and surrounding environment. In this paper we compare this sensori-motor network with a conventional neural network in the ability to create efficient predictors of visuomotor relationships. We confirm that the sensori-motor network is significantly more efficient in terms of required computations and is more precise (less prediction error) than the linear neural network in predicting self induced visual stimuli.