EVENT DETAILS AND ABSTRACT

Title: Neural computation in dynamic recurrent networks
Speaker: Eberhard Fetz
Speaker Info: University of Washington
Brief Description:
Special Note: More current information may be available at Plan-it Purple
Abstract:

Neural network modeling represents a significant tool for systems neurophysiologists to investigate the mechanisms underlying computation in complex networks. Recordings of neurons in behaving animals have provided useful insights, but never lead to complete network solutions because the connectivity of the recorded cells is inevitably missing. Dynamic recurrent neural networks can provide complete model solutions that simulate behaviors; all the connections and activities are known and available for analysis. Such networks can be derived simply from examples of the behavior by gradient descent methods and can incorporate many physiological and behavioral "constraints". This talk will review some basic applications of dynamic recurrent neural networks to model premotor circuits of the primate, oscillating networks, integrators and differentiators, instructed delay tasks and short-term memory. Issues include [1] simulation of lesions and stimulation, [2] pruning techniques to derive minimal functional networks, [3] generalization beyond the training set, and [4] converting networks of sigmoidal units with continuous activation functions (necessary for analytic derivation of connectivity) into networks of integrate-and-fire units performing the same functions.