Graduate School and Research Center in Digital Sciences

Walsh-hadamard variational inference for Bayesian deep learning

Rossi, Simone; Marmin, Sébastien; Filippone, Maurizio

Submitted on ArXiv, 27 May 2019

Over-parameterized models, such as DeepNets and ConvNets, form a class of models that are routinely adopted in a wide variety of applications, and for which Bayesian inference is desirable but extremely challenging. Variational inference offers the tools to tackle this challenge in a scalable way and with some degree of flexibility on the approximation, but for over-parameterized models this is challenging due to the over-regularization property of the variational objective. Inspired by the literature on kernel methods, and in particular on structured approximations of distributions of random matrices, this paper proposes Walsh-Hadamard Variational Inference (WHVI), which uses Walsh-Hadamard-based factorization strategies to reduce the parameterization and accelerate computations, thus avoiding over-regularization issues with the variational objective. Extensive theoretical and empirical analyses demonstrate that WHVI yields considerable speedups and model reductions compared to other techniques to carry out approximate inference for over-parameterized models, and ultimately show how advances in kernel methods can be translated into advances in approximate Bayesian inference.

Arxiv Bibtex

Title:Walsh-hadamard variational inference for Bayesian deep learning
Department:Data Science
Eurecom ref:5901
Copyright: © EURECOM. Personal use of this material is permitted. The definitive version of this paper was published in Submitted on ArXiv, 27 May 2019 and is available at :
Bibtex: @inproceedings{EURECOM+5901, year = {2019}, title = {{W}alsh-hadamard variational inference for {B}ayesian deep learning}, author = {{R}ossi, {S}imone and {M}armin, {S}{\'e}bastien and {F}ilippone, {M}aurizio}, booktitle = {{S}ubmitted on {A}r{X}iv, 27 {M}ay 2019}, address = {}, month = {05}, url = {} }
See also: