Estimating Information Flow in Deep Neural Networks
Name
goldfeld19a.pdf
Description
Published version
Size
7.47 MB
Format
Adobe PDF
Checksum (MD5)
72d9c9875456aa9b600ebb611bf54757
Author(s)
Goldfeld, Ziv
van den Berg, Ewout
Greenewald, Kristjan
Melnyk, Igor
Nguyen, Nam
Kinsgbury, Brian
Polyanskiy, Yury
Date Issued
2019
Journal
36th International Conference on Machine Learning, ICML 2019
Citation
Goldfeld, Ziv, van den Berg, Ewout, Greenewald, Kristjan, Melnyk, Igor, Nguyen, Nam et al. 2019. "Estimating Information Flow in Deep Neural Networks." 36th International Conference on Machine Learning, ICML 2019, 2019-June.
Version
Final published version
Abstract
Copyright © 2019 ASME We study the estimation of the mutual information I(X;Tℓ) between the input X to a deep neural network (DNN) and the output vector Tℓ of its ℓth hidden layer (an "internal representation"). Focusing on feedforward networks with fixed weights and noisy internal representations, we develop a rigorous framework for accurate estimation of I(X; Tℓ). By relating I(X; Tℓ) to information transmission over additive white Gaussian noise channels, we reveal that compression, i.e. reduction in I(X;Tℓ) over the course of training, is driven by progressive geometric clustering of the representations of samples from the same class. Experimental results verify this connection. Finally, we shift focus to purely deterministic DNNs, where I(X; Tℓ) is provably vacuous, and show that nevertheless, these models also cluster inputs belonging to the same class. The binning-based approximation of I(X; Tℓ) employed in past works to measure compression is identified as a measure of clustering, thus clarifying that these experiments were in fact tracking the same clustering phenomenon. Leveraging the clustering perspective, we provide new evidence that compression and generalization may not be causally related and discuss potential future research ideas.
MIT Department
MIT-IBM Watson AI Lab
Terms of Use
Article is made available in accordance with the publisher's policy and may be subject to US copyright law. Please refer to the publisher's site for terms of use.
Persistent DSpace Link
DOI of Published Version
http://proceedings.mlr.press/v97/goldfeld19a.html
