Skip to Main Content
Neural circuits that route motor activity to sensory structures play a fundamental role in perception. Their purpose is to aid basic cognitive processes by integrating knowledge about an organism's actions and to predict the perceptual consequences of those actions. This work develops a biologically inspired model of a visual stimulus prediction circuit and proposes a mathematical formulation for a computational implementation. We consider an agent with a visual sensory area consisting of an unknown rigid configuration of light-sensitive receptive fields which move with respect to the environment and according to a given number of degrees of freedom. From the agent's perspective, every movement induces an initially unknown change to the recorded stimulus. In line with evidence collected from studies on ontogenetic development and the plasticity of neural circuits, the proposed model adapts its structure with respect to experienced stimuli collected during the execution of a set of exploratory actions. We discuss the tendency of the proposed model to organize such that the prediction function is built using a particularly sparse feedforward network which requires a minimum amount of wiring and computational operations. We also observe a dualism between the organization of an intermediate layer of the network and the concept of self-similarity.