Uncertainty Estimation in Deep Neural Networks for Image Classification

Drost, Folke (2020) Uncertainty Estimation in Deep Neural Networks for Image Classification. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text mAI_2020_DrostFG.pdf Download (3MB) | Preview
	Text Toestemming.pdf Restricted to Registered users only Download (94kB)

Abstract

For this thesis, we investigated four methods for adding uncertainty estimations to the output of a deep neural network used for image classification: Two stochastic regularisation techniques, Dropout and Batch Normalisation, an Ensemble and a novel method named Error Output. All methods, except Error Output, perform multiple different predictions for a single example to obtain an uncertainty estimate. Error Output trains additional outputs to estimate its error. We trained and evaluated the performance of the four methods on two separate datasets. We used an additional dataset, on which the networks were not trained, to evaluate the uncertainty estimations. We found that the training time significantly increased for Ensemble, but not for the other methods. The inference time increased significantly for all methods except Error Output. Our results show that the uncertainty estimation on its own did not improve our ability to detect wrong predictions. We did, however, find that Dropout, Batch Normalisation and Ensemble, when increased inference time and memory requirements allow it, provide useful methods to lower the uncertainty of predictions. Overall, the addition of an uncertainty estimate proves useful for detecting untrained classes and the mean prediction improves prediction quality in general.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Wiering, M.A.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	20 Aug 2020 15:55
Last Modified:	20 Aug 2020 15:55
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/23135

Actions (login required)

View Item