A Reconfigurable Classifier Model of the Visual System

Neurons are costly for an organism to build and maintain, yet at the same time, there has apparently been strong selective pressure on organisms to incorporate additional functions into their nervous systems. Based on the known properties of dendritic trees and synapses (Koch, 2004), I propose a simple, biologically and evolutionary plausible architecture by which a neurons can be used for multiple different computations (time multiplexed). In such a model, signals extrinsic to an area selectively enable and disable parts of the dendritic trees of neurons within an area in a coordinated way. When this mechanism is extended to an entire pathway, like the ventral pathway, it gives rise to a notion of a reconfigurable pathway or a reconfigurable classifier, in which extrinsic signals can have the same pathway perform very different computations at different times (although some basic constraints remain due to overall connectivity; for example, V1 and V2 necessarily perform a kind of “feature extraction”, albeit tunable), and also route inputs and outputs from the pathway from and to different areas. In the rest of the paper, I explore the implications of such models.