Title: Building foundation models for neuroscience

Date: Tuesday, August 20th, 2024

Time: 2:30 pm EST

Location: Coda C1115 Druid Hills & Zoom https://gatech.zoom.us/j/97816662750

Mehdi Azabou

www.mehai.dev

Machine Learning PhD Student

School of Electrical and Computer Engineering
Georgia Institute of Technology

Committee

1 Dr. Eva L. Dyer, Advisor, School of Biomedical Engineering, Georgia Tech

2 Dr. Chethan Pandarinath, School of Biomedical Engineering, Georgia Tech

3 Dr. Anqi Wu,  School of Computational Science and Engineering, Georgia Tech

4 Dr. Hannah Choi,  School of Mathematics, Georgia Tech

5 Dr. Blake Richards, School of Computer Science and Montreal Neurological Institute, McGill University

Abstract

The brain’s complexity enables its remarkable functions, but this very complexity makes it hard to understand. Current methodologies for recording brain activity often provide narrow views of the brain's function, limited by the constraints of current recording technology and the structured nature of the standard neuroscience experiment. This fragmentation of datasets has hampered the development of robust and comprehensive computational models of brain function that generalize across diverse conditions, tasks, and individuals. Our work is motivated by the need for a large-scale foundation model in neuroscience--one that can go beyond the limitations of single-dataset approaches and offer a fuller, more comprehensive picture of brain function. In this thesis, we propose novel methodologies and frameworks aimed at addressing the challenges of building such a model. We discuss three main contributions. The first contribution is towards building scalable and unified approaches for training on diverse neural datasets. The second contribution aims to develop self-supervised methods for understanding dynamics of behavior at multiple timescales. The third contribution is to develop methods for building invariances in neural data to further our understanding of the brain.