PAC-Bayesian Margin Bounds for Convolutional Neural Networks - Technical Report

Recently the generalisation error of deep neural networks has been analysed through the PAC-Bayesian framework, for the case of fully connected layers. We adapt this approach to the convolutional setting.