SVCCA: Singular Vector Canonical Correlation Analysis for Deep Learning Dynamics and Interpretability

TL;DR

SVCCA is a novel technique for efficiently comparing neural network representations, invariant to affine transformations, enabling insights into layer dimensionality, learning dynamics, and class-specific information.

Contribution

Introduces SVCCA, a fast, affine-invariant method for comparing neural network representations, facilitating analysis of training dynamics and network interpretability.

Findings

Networks can be over-parameterized in some layers.

Networks converge from bottom to top during training.

Class-specific information is localized within certain network layers.

Abstract

We propose a new technique, Singular Vector Canonical Correlation Analysis (SVCCA), a tool for quickly comparing two representations in a way that is both invariant to affine transform (allowing comparison between different layers and networks) and fast to compute (allowing more comparisons to be calculated than with previous methods). We deploy this tool to measure the intrinsic dimensionality of layers, showing in some cases needless over-parameterization; to probe learning dynamics throughout training, finding that networks converge to final representations from the bottom up; to show where class-specific information in networks is formed; and to suggest new training regimes that simultaneously save computation and overfit less. Code: https://github.com/google/svcca/