InfoResearchPeer-reviewed
Enhancing Robustness in Deep Convolutional Neural Networks Through Multiresolution Learning
- Published
- Record updated
Summary
This study proposes multiresolution learning, which rebuilds training data at several resolutions and trains a deep neural network sequentially from coarsest to finest. The authors report significant gains in noise and adversarial robustness on 1-D signal and 2-D image prediction tasks, including with limited training data. They state that robustness improves without sacrificing standard accuracy, at the same computational cost as standard training.