CMPUT 624 - Machine Learning and the Brain

Overview

When we train machine learning models to perform certain tasks, what do they learn? Does the algorithm they learn have any resemblance to the algorithm the human brian uses to do the same task? There is some evidence that the algorithm encoded in machine learning models is very different from what the brain does (e.g. adversarial examples trick CNNs but not people). But, there is also evidence that some representations learned by ML models resemble the representations we can measure in the brain via brain imaging.

This class explores a variety of ways of comparing ML models to brain activity through computer vision, language models, and reinforcement learning. We will read a collection of papers: those that set the stage for the field, and more recent work. Students will present papers to the class and participate in paper discussions. Students will work on a project (alone or in groups of 2-3), write a project report, and present that project to the class. Some students have used these project reports as the basis of papers and/or theses.

The class assumes you are a strong programmer. Some prior exposure to Machine Learning would be advantageous, though we will refresh many of the basics in class and as we go through the papers. This class does not assume any neuroscience knowledge.

Objectives

• Understand some of the basics of machine learning and neuroscience
• Understand what a representation is, and how we can compare representational spaces
• Learn how to read and present a paper, and how to lead an engaging discussion about that paper
• Get hands-on experience in machine learning and neuroscience as part of a group project
• Write a paper about the project, coming as close to a submittable paper as possible by the end of the semester

Course Work

• Paper presentation
• Project
• In-class participation