Uncovering how the brain learns

Summary

Date: 
January 23, 2018 - 12:00pm
Location: 
Northwest Building, Room B103
About the Speaker
Name: 
Cengiz Pehlevan
Speaker Title: 
Research Scientist
Speaker Affiliation: 
Flatiron Institute

Modern experimental tools have given rise to large datasets of brain structure and activity. How do we infer from these datasets what the brain computes? For that, we need new theories that bridge computation and its biological realization. I will present a new theory of unsupervised learning, called similarity matching, that achieves this goal. The theory starts by posing computational goals of learning as mathematical optimization problems. Then, from these problems, I systematically derive algorithms and neural circuit implementations of these algorithms, linking computation to biological realization.
While this approach has been used before, it failed in providing biologically plausible, local plasticity rules. I will argue that the solution is a change of view in what neural population activity represents. Conventionally, population activity is thought to represent the stimulus. Instead, I propose that the similarity of population activity matches the similarity of the stimuli under certain constraints. I will show that this similarity matching principle naturally leads to algorithms that can be implemented with local plasticity. I will give examples of such algorithms and their biologically plausible implementations.
I will show that the theory achieves the following: 1) It provides new optimization formulations of common unsupervised learning tasks. 2) It answers how efficient self-organization for learning happens with local synaptic plasticity. 3) It provides new circuit motifs and mechanisms that could be in use in the brain. 4) It makes experimental predictions about computations performed in specific circuits and provides computational interpretations of salient features of such circuits.
Finally, I will discuss ongoing and future work on extensions of these ideas and their applications to the problem of object recognition. I will also briefly present some of my other work, including work on reinforcement learning of motor skills.