Explicit Regularisation in Gaussian Noise Injections
Abstract
We study the regularisation induced in neural networks by Gaussian noise injections (GNIs). Though such injections have been extensively studied when applied to data, there have been few studies on understanding the regularising effect they induce when applied to network activations. Here we derive the explicit regulariser of GNIs, obtained by marginalising out the injected noise, and show that it penalises functions with high-frequency components in the Fourier domain; particularly in layers closer to a neural network's output. We show analytically and empirically that such regularisation produces calibrated classifiers with large classification margins.
Cite
Text
Camuto et al. "Explicit Regularisation in Gaussian Noise Injections." Neural Information Processing Systems, 2020.Markdown
[Camuto et al. "Explicit Regularisation in Gaussian Noise Injections." Neural Information Processing Systems, 2020.](https://mlanthology.org/neurips/2020/camuto2020neurips-explicit/)BibTeX
@inproceedings{camuto2020neurips-explicit,
title = {{Explicit Regularisation in Gaussian Noise Injections}},
author = {Camuto, Alexander and Willetts, Matthew and Simsekli, Umut and Roberts, Stephen J. and Holmes, Chris C},
booktitle = {Neural Information Processing Systems},
year = {2020},
url = {https://mlanthology.org/neurips/2020/camuto2020neurips-explicit/}
}