Explicit Regularisation in Gaussian Noise Injections Article Swipe

MAKE IMPACT

Alexander Camuto , Matthew Willetts , Umut Şimşekli , Stephen Roberts , Chris Holmes · YOU? · · 2020 · Open Access · · DOI: https://doi.org/10.48550/arxiv.2007.07368

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.

Type

preprint

Language

en

Landing Page

http://arxiv.org/abs/2007.07368

PDF

https://arxiv.org/pdf/2007.07368

OA Status

green

Cited By

35

References

48

Related Works

10

OpenAlex ID

https://openalex.org/W3042925324

Raw OpenAlex JSON