Beyond accuracy: robustness and generalization properties of biologically plausible learning rules

The overarching goal of my research is to investigate how the brain learns in order to develop brain-inspired learning algorithms that will reduce the computational cost for many machine learning applications. The standard learning algorithm (or learning rule) for training artificial neural networks (ANNs) can be extremely costly in terms of computation and storage, driving a search for more efficient learning algorithms. Meanwhile, the brain excels at learning patterns across multiple timescales efficiently, thereby motivating an influx of biologically motivated learning algorithms. However, the robustness and generalization properties of these learning rules are severely underexamined. Here, we aim to investigate and improve the generalization capabilities of biologically plausible learning algorithms by leveraging recent experimental neuroscience findings as well as theoretical tools from deep learning.

Faculty Supervisor:

Guillaume Lajoie

Student:

Partner:

University of Washington

Discipline:

Computer science

Sector:

Education

University:

Université de Montréal

Program: