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

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Evaluating Bayesian Representational Similarity Measures between Neural Networks

Sezan Oral¹, Heiko H. Schütt¹; ¹University of Luxemburg

Presenter: Sezan Oral

Reliable comparison of representations remains an unresolved challenge in interpreting deep neural networks. Here, we propose and evaluate BayesCompare, a Bayesian framework that uses linear readout for measuring representational similarities of neural networks. We evaluate our method, comparing it to existing methods using 5 instances each of six ImageNet-trained neural network architectures. Our methods perform best at matching layers across instances, match the order of layers well and yield more stable results across random image choice. These results establish BayesCompare as an effective way of measuring representational similarity while being easy and fast to compute.

Topic Area: Methods, Tools, Theory & Neural Coding