Large-Scale Models of Visual Cortex

Dana Ballard

Models have provided enormous insight into the understanding of the operation of a single neuron, but their complexity has precluded their use in the modeling of the large hiearchical networks seen in mammalian visual cortex. In order to progress, it may be the case that simpler models that can be readily extended are required. Such models must be motivated by an understanding of the purpose of the cortex, and how it codes its input. We propose such a model based on the Minimum Description Length principle. The main virtue of the model is that the equations between cortical areas are linear, allowing large networks to be combined in a prescriptive way that can predict their functioning. In particular experimentally observed response properties of neurons in striate cortex can be interpreted as network phenomena. 1) Endstopping can be shown to result from connections to neurons in higher cortical areas. 2) The translation invariance of complex cells can be shown to result from network connections to simple cells.

Back to the Natural Scenes Meeting Agenda.