PhD Transfer Exam – Jordan Guerguiev (Richards lab)

PhD Transfer Exam

Tuesday October 6th, 2:30 pm – Room 432, Ramsay Wright Building, University of Toronto

Jordan Guerguiev (Richards lab)

“Biologically Realistic Supervised Learning”

Abstract

Supervised learning refers to learning in neural networks under the guidance of an outside teaching signal. It is an important concept to consider, since whenever the goal of learning is to match some ideal behavioural target, supervised learning provides the best mechanism for achieving this goal, especially in complicated tasks. However, the most successful learning algorithms for supervised learning in neural networks invoke mechanisms that are not biologically realistic. For example, learning with backpropagation of error requires a neuron in a network to know the synaptic connectivity of all the neurons downstream of it. Thus, a biologically plausible mechanism for supervised learning has not yet been proposed. The goal of this Ph.D. project is to develop such a mechanism for supervised learning in neural networks, by combining two results from previous papers – a difference target propagation algorithm proposed by Dong-Hyun Lee et al. and a dendritic prediction algorithm described by Robert Urbanczik and Walter Senn. Using these results, we have developed a spiking neural network model which can engage in biologically realistic supervised learning. In addition to generating a biologically plausible model, the goal of this research is to generate predictions about neuronal behaviour, and to test these predictions in the living brain using two-photon microscopy paired with optogenetics.
Ramsay Wright is a wheelchair accessible building.