Cortical Mechanisms of Visual Recognition and Learning: A Hierarchical Kalman Filter Model

We describe a biologically plausible model of dynamic recognition and learning in the visual cortex based on the statistical theory of Kalman filtering from optimal control theory. The model utilizes a hierarchical network whose successive levels implement Kalman filters operating over successively larger spatial and temporal scales. Each hierarchical level in the network predicts the current visual recognition state at a lower level and adapts its own recognition state using the residual error between the prediction and the actual lower-level state. Simultaneously, the network also learns an internal model of the spatiotemporal dynamics of the input stream by adapting the synaptic weights at each hierarchical level in order to minimize prediction errors. The Kalman filter model respects key neuroanatomical data such as the reciprocity of connections between visual cortical areas, and assigns specific computational roles to the inter-laminar connections known to exist between neurons in the visual cortex. Previous work elucidated the usefulness of this model in explaining neurophysiological phenomena such as endstopping and other related extra-classical receptive field effects. In this paper, in addition to providing a more detailed exposition of the model, we present a variety of experimental results demonstrating the ability of this model to perform robust spatiotemporal segmentation and recognition of objects and image sequences in the presence of varying amounts of occlusion, background clutter, and noise.