Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster B in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Bayesian linear encoding to evaluate models on neuroimaging data

Juan Jesús Torre Tresols¹, Sezan Oral¹, Heiko H. Schütt¹; ¹University of Luxemburg

Presenter: Juan Jesús Torre Tresols

Evaluating the similarity of deep neural networks to human brain data is a crucial method to test mechanistic models of cognition. Although several methods to perform these evaluations exist, they present a number of computational and practical disadvantages. We evaluate a Bayesian linear encoding model that allows for the computation of brain similarity without explicitly fitting the feature weights by operating on the kernel matrix. To assess the reliability of our method, we calculate the correlation of model rankings between subjects, for our method and other state of the art methods. Preliminary results show a slightly higher correlation across subjects for our method (τ=0.91), than for a frequentist encoding model (τ=0.88). Our proposed method constitutes an interesting addition to existing tools for brain similarity, achieving slightly better results while being much faster to compute and avoiding common steps such as dimensionality reduction and hyper-parameter setting through cross-validation.

Topic Area: Methods, Tools, Theory & Neural Coding