......................................................................

NJIT Mathematical Biology Seminar

Tuesday, February 13, 2007, 4:00pm
Cullimore Hall 611
New Jersey Institute of Technology

......................................................................

The importance of neural diversity in complex cognitive tasks

Stefano Fusi

Center for Neurobiology and Behavior
Department of Physiology and Cellular Biophysics
Columbia University School of Medicine

Our decisions are usually based on a variety of factors, not only sensory inputs and response options, but also behavioral rule in play, previous choices and their consequences. All these factors determine the context which plays a key role in our decision process. Here we propose a model for the internal representation of contexts as attractor states of a neural circuit. Every contextual attractor corresponds to a specific inner mental state, which is self-sustained and reflects a specific disposition to behavior (e.g. how to interpret a stimulus in a specific context and to react to it in order to get reward). Every externally or internally generated event induces a transition from one inner mental state to another. In principle such a dynamical system can reproduce arbitrarily complex behaviors, provided that there are enough inner mental states. We show that unfortunately the fraction of transitions that can actually be implemented in a dynamical system decreases rapidly with the total number of desired transitions Such a limitation does not depend on the representation of the attractors or the external events. The solution that we propose is to harness the huge diversity observed in the neural responses recorded in vivo, which most likely reflects the heterogeneity of the neural cells and their connections. If we start from the neurons which have a particular selectivity (e.g. to a stimulus, or to an intended motor response), other neurons with random connections with them would exhibit mixed selectivity. We prove that if there are enough of these mixed selectivity neurons, then any scheme of attractors and transitions can be implemented in a dynamical system. The recipe that we propose to build such a system is valid for any arbitrary neural representation of the inner mental states (attractors) and the events which trigger the transitions. We then illustrate the theory by building an attractor neural network which reproduces the animal behavior and several features of the neural activity recorded in prefrontal cortex in an experiment in which monkeys perform a Wisconsin card sorting test (Mansouri, Matsumoto, Tanaka 2006).