Powerful Learning via Cortical Microcircuits
By: Rui Ponte Costa, Computational Neuroscience and Machine Learning, Department of Computer Science, University of Bristol
Venue | Senior Common Room, Level 2 (2D17), Priory Road Complex
Date | Thursday 14 March 2019
Time | 13:00
Cortical circuits exhibit intricate excitatory and inhibitory motifs, whose computational functions remain poorly understood. I will start out by introducing our work on how state-of-the-art recurrent neural networks used in machine learning, such as gated-recurrent neural networks may be implemented by cortical microcircuits. In addition, our new results suggest that such biologically plausible recurrent networks exhibit better learning of long-term dependencies. However, learning in such networks relies on solving the credit assignment problem using the classical (implausible) backpropagation algorithm. I will finish my talk discussing our recent work on a biologically plausible solution to the credit assignment problem using well-known properties of cortical circuits, which approximates the backpropagation algorithm. Overall, our work demonstrates how cortical microcircuits may enable powerful learning in the brain.
All Welcome | Tea, coffee and cakes will be available after the seminar.