On the Trajectory of Stochastic Gradient Descent in the Information PlaneDownload PDF

27 Sep 2018 (modified: 21 Dec 2018)ICLR 2019 Conference Blind SubmissionReaders: Everyone
  • Abstract: Studying the evolution of information theoretic quantities during Stochastic Gradient Descent (SGD) learning of Artificial Neural Networks (ANNs) has gained popularity in recent years. Nevertheless, these type of experiments require estimating mutual information and entropy which becomes intractable for moderately large problems. In this work we propose a framework for understanding SGD learning in the information plane which consists of observing entropy and conditional entropy of the output labels of ANN. Through experimental results and theoretical justifications it is shown that, under some assumptions, the SGD learning trajectories appear to be similar for different ANN architectures. First, the SGD learning is modeled as a Hidden Markov Process (HMP) whose entropy tends to increase to the maximum. Then, it is shown that the SGD learning trajectory appears to move close to the shortest path between the initial and final joint distributions in the space of probability measures equipped with the total variation metric. Furthermore, it is shown that the trajectory of learning in the information plane can provide an alternative for observing the learning process, with potentially richer information about the learning than the trajectories in training and test error.
  • Keywords: Stochastic gradient descent, Deep neural networks, Entropy, Information theory, Markov chains, Hidden Markov process.
  • TL;DR: We look at SGD as a trajectory in the space of probability measures, show its connection to Markov processes, propose a simple Markov model of SGD learning, and experimentally compare it with SGD using information theoretic quantities.
9 Replies