PhD Proposal Exam-Colleen Gillon -Tuesday, June 12, 2018

PhD Proposal Exam

Tuesday, June 12, 2018 at 10:10am, MW 229- University of Toronto at Scarborough

Colleen Gillon (Richards Lab)

“How does the brain learn about the statistical structure of the environment?”

Abstract

Over the past decade, artificial intelligence has progressed at great speed, with impressive breakthroughs in fields like computer vision and speech processing using neural network algorithms. These can be broadly divided into two classes: (1) discriminative models, like feedforward, convolutional and recurrent neural nets, which learn to map inputs, like images, to specific outputs, like categories or classes and (2) generative models, like bidirectional Helmholtz machines, generative-adversarial networks and expectation maximization models, which are learn the underlying joint structure of the data. Studies of visual processing in the cortex strongly suggest that in learning to process environmental stimuli, our brains behave like generative models, developing internal models of the joint distribution over sensory stimuli in the environment. Thus, these algorithms could shed light on our brain’s remarkable ability to represent and process sensory information efficiently and accurately. We propose to investigate this by comparing how the brain and different algorithms trained on visual tasks process and adapt to major changes in the relationship between incoming visual stimuli and somatosensory or motor inputs. Specifically, we will record and analyse changes in the activity of layer 2/3 pyramidal neurons in primary visual cortex (V1) in response to a shift in the relationship between visual stimuli and sensory stimuli or motor commands. We predict that this shift will transiently increase activity in the apical dendrites, and alter the rate of apical trunk calcium spikes of these neurons while the system adapts. In parallel, we will train different generative algorithms on this same task, and analyze changes in network activity in order to identify those algorithms that show the greatest potential for explaining how our brains process sensory information.