Timezone: »

Explicit Regularisation in Gaussian Noise Injections
Alexander Camuto · Matthew Willetts · Umut Simsekli · Stephen J Roberts · Chris C Holmes

Wed Dec 09 09:00 AM -- 11:00 AM (PST) @ Poster Session 3 #1100

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.

Author Information

Alexander Camuto (University of Oxford & The Alan Turing Institute)
Matthew Willetts (University of Oxford)
Umut Simsekli (Inria/ENS)
Stephen J Roberts (University of Oxford)
Chris C Holmes (University of Oxford)

More from the Same Authors