Implicit Bias of Gradient Descent on Linear Convolutional Networks

Part of Advances in Neural Information Processing Systems 31 (NeurIPS 2018)

Authors

Suriya Gunasekar, Jason D. Lee, Daniel Soudry, Nati Srebro

Abstract

We show that gradient descent on full-width linear convolutional networks of depth $L$ converges to a linear predictor related to the $\ell_{2/L}$ bridge penalty in the frequency domain. This is in contrast to linearly fully connected networks, where gradient descent converges to the hard margin linear SVM solution, regardless of depth.