Keynote Lecture 1
Completion requirements
Methods for analyzing and comparing DNN representations
Pairwise comparison of hidden representations produced by DNNs is a hard and daunting task. Besides the very large dimensionality in which these representations live, one has to take into consideration invariances that common similarity metrics (e.g. cosine similarity) lack. We will be presenting a handful of such metrics (e.g. CCA, CKA…) and the ideas behind them.
In the second part of the lecture, we will show the implementation of these techniques in PyTorch, showing also different ways to extract the hidden representations from DNNs.