Fall 2016
Seminars are held at 11:30AM in Cullimore Hall, Room 611, unless noted otherwise. For questions about the seminar schedule, please contact Casey Diekman.
Date: October 11, 2016
Speaker: Ashok Litwin-Kumar
Center for Theoretical Neuroscience,
Columbia University
Title: "Optimal Connectivity for Random and Learned Neural Representations"
Abstract:
The number of synaptic connections that an individual neuron in the brain receives varies over several orders of magnitude. Cerebellar granule cells, the most numerous neurons in the human brain, receive on average four inputs, while cerebrocortical neurons receive thousands. What controls the optimal degree of synaptic connectivity for a given neuronal type? I will discuss recent work we have done focusing on how the dimension of a representation of sensory stimuli formed by a population of neurons depends on the number of inputs they receive and what this entails for a classifier that uses this representation to learn associations. The theory I will describe predicts optimal values for the number of inputs received by cerebellar granule cells and Kenyon cells of the Drosophila mushroom body that match those observed experimentally, demonstrating that sparse connectivity (fewer than ten inputs per neuron) can be superior to dense connectivity when synaptic wiring is random. If synaptic strengths are subject to plasticity during learning, however, the optimal connectivity shifts toward dense wiring. If I have time, I will also describe analysis of an electron microscopy reconstruction of the larval Drosophila mushroom body that we have used to verify key predictions of our theory.