Vergence control learning through real V1 disparity tuning curves

Abstract

A neural network architecture able to autonomously learn effective disparity-vergence responses and drive the vergence eye movements of a simulated binocular active vision system is proposed. The proposed approach, instead of exploiting purposely designed resources, relies on the direct use of a set of real disparity tuning curves, measured in the primary visual cortex of two macaque monkeys and courteously made available by (Prince et al., 2002), that provides a distributed representation of binocular disparity. The network evolves following a differential Hebbian rule that exploits the overall population activity to measure the state of the system, i.e. the deviation from the desired vergence position, so as its modification as a consequence of the action performed. Accordingly, the signal provides an effective intrinsic reward to develop a stable and accurate vergence behaviour. Emerging from a direct interaction of the sensing system with the environment, the resulting control provides a precise and accurate control for small disparities, as well as a raw control on a broader working range when large disparities are experienced. The developed control converges to a stable state that intrinsically and continuously adapts to the characteristics of the ongoing stimulation. The results proved how actually naturally distributed resources allows for robust and flexible learning capabilities in changeable situations.