Simpson, Ivor J A, Vicente, Sara and Campbell, Neil D F (2022) Learning structured gaussians to approximate deep ensembles. IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR), New Orleans, Louisiana, USA, 19-24 June 2022. Published in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 1-9. IEEE/CVF ISSN 2575-7075 ISBN 9781665469463

PDF - Accepted Version Download (8MB)

Abstract

This paper proposes using a sparse-structured multivariate Gaussian to provide a closed-form approximator for the output of probabilistic ensemble models used for dense image prediction tasks. This is achieved through a convolutional neural network that predicts the mean and covariance of the distribution, where the inverse covariance is parameterised by a sparsely structured Cholesky matrix. Similarly to distillation approaches, our single network is trained to maximise the probability of samples from pretrained probabilistic models, in this work we use a fixed ensemble of networks. Once trained, our compact representation can be used to efficiently draw spatially correlated samples from the approximated output distribution. Importantly, this approach captures the uncertainty and structured correlations in the predictions explicitly in a formal distribution, rather than implicitly through sampling alone. This allows direct introspection of the model, enabling visualisation of the learned structure. Moreover, this formulation provides two further benefits: estimation of a sample probability, and the introduction of arbitrary spatial conditioning at test time. We demonstrate the merits of our approach on monocular depth estimation and show that the advantages of our approach are obtained with comparable quantitative performance.

Item Type:	Conference Proceedings
Schools and Departments:	School of Engineering and Informatics > Informatics
SWORD Depositor:	Mx Elements Account
Depositing User:	Mx Elements Account
Date Deposited:	28 Jul 2022 11:55
Last Modified:	01 Nov 2022 14:46
URI:	http://sro.sussex.ac.uk/id/eprint/107043

View download statistics for this item

📧 Request an update