SVCCA: Singular Vector Canonical Correlation Analysis for Deep Learning Dynamics and Interpretability
This provides a practical tool for researchers and practitioners in machine learning to better understand and optimize neural network representations, though it is incremental as it builds on existing correlation analysis methods.
The authors tackled the problem of comparing representations in deep learning by introducing SVCCA, a tool that is invariant to affine transforms and computationally efficient, enabling analysis of layer dimensionality, learning dynamics, and class-specific information, leading to suggestions for training regimes that reduce computation and overfitting.
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/