Abstract: Comparing representations learned by different neural networks is required to understand for example differences between model architectures, usefulness of transfer learning, or robustness of the models. In this talk, I will discuss how current measures for comparing representations between NNs may not well reflect their functional similarity due to the confounding effect of structure of the data in the input space. I will present a simple fix to this problem and show how it improves the resolution to identify functionally similar neural networks, leads to improved insights in transfer learning, and better reflects out-of-distribution accuracy. The talk is based on an article presented recently at the NeurIPS 2022 conference [1].[1] Tianyu Cui, Yogesh Kumar, Pekka Marttinen, and Samuel Kaski (2022). Deconfounded Representation Similarity for Comparison of Neural Networks. Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS 2022). Pre-print: https://arxiv.org/abs/2202.00095

Speaker: Pekka Marttinen is an associate professor in machine learning in the department of computer science at Aalto university, where he leads the Machine Learning for Health (Aalto-ML4H) research group. His research interests include for example Bayesian machine learning, causality, deep learning, and applications in biology and healthcare, and he has published over 70 articles in this topics.

Affiliation: Aalto University

Place of Seminar:  Otaniemi, T5 (in person) & Kumpula exactum C323 (streaming)