Cortical integration of bottom-up, top-down and horizontal information in biological motion recognition

We propose a cortical architecture for integrating top-down, bottom-up and horizontal inputs, and apply this architecture to the problem of biological motion recognition. The mechanism makes use of spike-based computation in order to represent the relative phases of limb motion in terms of precise neuronal spike times. We propose that information regarding spatial relations between features is represented in the system of precise horizontal connections in cortex, whereas the feedforward and feedback pathways are chiefly concerned with the statistics of feature distributions. The architecture allows cellular biophysical-level integration mechanisms to carry out probabilistic inference on streams of afferent input.