Bayesian Neural Networks With Maximum Mean Discrepancy Regularization

Bayesian Neural Networks (BNNs) are trained to optimize an entire distribution over their weights instead of a single set, having significant advantages in terms of, e.g., interpretability, multi-task learning, and calibration. Because of the intractability of the resulting optimization problem, mos...

Full description

Bibliographic Details

Main Authors:	Pomponi, Jary, Scardapane, Simone, Uncini, Aurelio
Format:	Article
Published:	2020
Subjects:	Machine Learning
Online Access:	https://arxiv.org/abs/2003.00952 http://arxiv.org/abs/2003.00952
Tags:	Add Tag No Tags, Be the first to tag this record!