Self-organization of Visual Sensor Topologies Based on Spatiotemporal Cross-Correlation

Abstract

In living organisms, the morphology of sensory organs and the behavior of a sensor’s host are strongly tied together. For visual organs, this interrelationship is heavily influenced by the spatial topology of the sensor and how it is moved with respect to an organism’s environment. Here we present a computational approach to the organization of spatial layouts of visual sensors according to given sensor-environment interaction patterns. We propose that prediction and spatiotemporal correlation are key principles for the development of visual sensors well-adapted to an agent’s interaction with its environment. This proposition is first motivated by studying the interdependency of morphology and behavior of a number of visual systems in nature. Subsequently, we encode the characteristics observed in living organisms by formulating an optimization problem which maximizes the average spatiotemporal correlation between actual and predicted stimuli. We demonstrate that the proposed formulation leads to spatial self-organization of visual receptive fields, and leads to different sensor topologies according to different sensor displacement patterns. The obtained results demonstrate the explanatory power of our approach with respect to i) the development of spatially coherent light receptive fields on a visual sensor surface, and ii) the particular topological organization of receptive fields depending on sensorimotor activity.