Analyzing noise compensation properties of trained recurrent neural networks

Reliability is a fundamental requirement for computational systems, brains and artificial models alike: a system should respond the same way for repeated presentations of the same stimulus. However, the brain has two features that can threaten its reliability: intrinsic stochasticity and chaos. Stochasticity takes the form of random fluctuations affecting the reliability of components of the system, whereas chaos is an emergent property of the entire system that causes similar inputs or initial conditions to produce totally different outputs. The brain must have mechanisms to compensate for its noisy and unreliable machinery, and the goal of our project is to characterize these mechanisms. To this end, we will first develop tools for quantifying the reliability of models of neural circuits, and we will subsequently extend the capabilities of these tools to analyze data collected from real neuroscience experiments.

Faculty Supervisor:

Guillaume Lajoie

Student:

Partner:

New York University

Discipline:

Life Sciences

Sector:

Education

University:

Université de Montréal

Program: