Similarity of Neural Network Representations Revisited


Recent work has sought to understand the behavior of neural networks by comparing representations between layers and between different trained models. We examine methods for comparing neural network representations based on canonical correlation analysis (CCA). We show that CCA belongs to a family of statistics for measuring multivariate similarity, but that neither CCA nor any other statistic that is invariant to invertible linear transformation can measure meaningful similarities between representations of higher dimension than the number of data points. We introduce a similarity index that measures the relationship between representational similarity matrices and does not suffer from this limitation. This similarity index is equivalent to centered kernel alignment (CKA) and is also closely connected to CCA. Unlike CCA, CKA can reliably identify correspondences between representations in networks trained from different initializations.

Jan 14, 2020 12:00 PM
Iman Mirzadeh
Research Assistant

I am a PhD student and Graduate Research Assistant at the Washington State University Embedded and Pervasive Systems Laboratory (EPSL) under supervision of Dr. Hassan Ghasemzadeh. I am interested in the real-world challenges of working with machine learning models such as energy constraints and human-in-the-loop interactions with these models. Specifically, I am focusing on Model Optimization (such as model compression), where my goal is to build more efficient models or use the existing models more efficiently. Before joining EPSL, I was an ML Engineer at Sokhan AI, where we provided accurate and scalable Natural Language Processing (NLP) and Computer Vision (CV) services to businesses.