Comparing Representations Learned by Neural Networks - Deconfounded Similarity Measures

Explore advanced techniques for comparing representations learned by neural networks in this insightful 42-minute talk by Pekka Marttinen from the Finnish Center for Artificial Intelligence FCAI. Delve into the limitations of current comparison measures and discover a novel approach to address the confounding effect of input data structure. Learn how this improved method enhances the identification of functionally similar neural networks, provides valuable insights for transfer learning, and better reflects out-of-distribution accuracy. Examine findings from a recent NeurIPS 2022 conference paper on deconfounded representation similarity. Gain expertise from Pekka Marttinen, an associate professor in machine learning at Aalto University, whose research spans Bayesian machine learning, causality, deep learning, and applications in biology and healthcare.