FriSem
Imran Thobani, Philosophy PhD student with Dan Yamins and Rosa Cao, Stanford University
Title: Inter-animal transforms as a guide to model-brain alignment
Abstract: The appropriate methods for aligning neural network models to the brain remain controversial. Ideally, a good alignment method should be powerful enough to enable accurate predictions of neural responses under a mapping from model units to neurons, while also being specific enough to distinguish a particular kind of response pattern (e.g. typical patterns of a given brain area) from other kinds of response patterns. It has generally been assumed that the goals of predictivity and specificity are in tension with each other, with methods that severely restrict the possible relationships between model and target being better for specificity, and more flexible methods yielding higher predictivity. We show that this apparent tension does not in fact exist. Fundamentally, this is because specificity requires not only distinguishing response patterns from different brain areas (i.e. separation), but also recognizing response patterns from the same brain area as being similar across subjects (i.e. identification). Taking this into account, we find that relatively flexible methods, like linear regression, can exhibit greater specificity compared to stricter methods, while also enabling better predictions. Motivated by the idea that the correct balance between strict and loose is manifested by the empirical relationships between subjects in a population, we introduce an alignment method that incorporates known aspects of the biological circuit, further improving predictivity without reducing specificity.