Bottom Up or Top Down? Dynamics of Deep Representations via Canonical Correlation Analysis

Maithra Raghu, Jason Yosinski, Jascha Sohl-Dickstein

Feb 17, 2017 (modified: Mar 05, 2017) ICLR 2017 workshop submission readers: everyone
  • Abstract: We present a versatile quantitative framework for comparing representations in deep neural networks, based on Canonical Correlation Analysis, and use it to analyze the dynamics of representation learning during the training process of a deep network. We find that layers converge to their final representation from the bottom-up, but that the representations themselves migrate downwards in the net-work over the course of learning.
  • TL;DR: Use CCA to look at representation learning dynamics of neural networks, and find bottom up convergence, top down representation crawling.
  • Keywords: Theory, Deep learning
  • Conflicts: google.com, cs.cornell.edu, uber.com

Loading