Bayesian Neural Networks (BNNs) have recently received increasing attention
for their ability to provide well-calibrated posterior uncertainties. However,
model selection---even choosing the number of nodes---remains an open question.
In this work, we apply a horseshoe prior over node pre-activations of a
Bayesian neural network, which effectively turns off nodes that do not help
explain the data. We demonstrate that our prior prevents the BNN from
under-fitting even when the number of nodes required is grossly over-estimated.
Moreover, this model selection over the number of nodes doesn't come at the
expense of predictive or computational performance; in fact, we learn smaller
networks with comparable predictive performance to current approaches.
Description
[1705.10388] Model Selection in Bayesian Neural Networks via Horseshoe Priors
%0 Journal Article
%1 ghosh2017model
%A Ghosh, Soumya
%A Doshi-Velez, Finale
%D 2017
%K approximate bayesian readings
%T Model Selection in Bayesian Neural Networks via Horseshoe Priors
%U http://arxiv.org/abs/1705.10388
%X Bayesian Neural Networks (BNNs) have recently received increasing attention
for their ability to provide well-calibrated posterior uncertainties. However,
model selection---even choosing the number of nodes---remains an open question.
In this work, we apply a horseshoe prior over node pre-activations of a
Bayesian neural network, which effectively turns off nodes that do not help
explain the data. We demonstrate that our prior prevents the BNN from
under-fitting even when the number of nodes required is grossly over-estimated.
Moreover, this model selection over the number of nodes doesn't come at the
expense of predictive or computational performance; in fact, we learn smaller
networks with comparable predictive performance to current approaches.
@article{ghosh2017model,
abstract = {Bayesian Neural Networks (BNNs) have recently received increasing attention
for their ability to provide well-calibrated posterior uncertainties. However,
model selection---even choosing the number of nodes---remains an open question.
In this work, we apply a horseshoe prior over node pre-activations of a
Bayesian neural network, which effectively turns off nodes that do not help
explain the data. We demonstrate that our prior prevents the BNN from
under-fitting even when the number of nodes required is grossly over-estimated.
Moreover, this model selection over the number of nodes doesn't come at the
expense of predictive or computational performance; in fact, we learn smaller
networks with comparable predictive performance to current approaches.},
added-at = {2020-01-23T01:41:44.000+0100},
author = {Ghosh, Soumya and Doshi-Velez, Finale},
biburl = {https://www.bibsonomy.org/bibtex/2a24d558a46d6f08307d20ecbbbff8904/kirk86},
description = {[1705.10388] Model Selection in Bayesian Neural Networks via Horseshoe Priors},
interhash = {7c8ac3c6c8f7c04ed380ece5c49ca631},
intrahash = {a24d558a46d6f08307d20ecbbbff8904},
keywords = {approximate bayesian readings},
note = {cite arxiv:1705.10388},
timestamp = {2020-01-23T01:41:44.000+0100},
title = {Model Selection in Bayesian Neural Networks via Horseshoe Priors},
url = {http://arxiv.org/abs/1705.10388},
year = 2017
}