Back to search

University of British Columbia

Out-of-distribution detection using inter-level features of deep neural networks

Abstract

dc:description

The work presented in this thesis addresses the problem of Out-of-Distribution (OOD) detection in deep learning-based classifiers. The emphasis is to investigate real-world settings, where both the In-distribution (ID) and OOD samples are hard to distinguish due to their high semantic similarities. Despite the significant performance of classification networks, they fail to correctly handle OOD samples. During training, the closed-world assumption is heavily relied upon. It is assumed that models are only forwarded samples that share similar distributions to the training data. In dynamic environments, the assumption is strict and leads to severe degradation in performance. While uncertainty quantification methods are elegant solutions to represent predictive confidence, the presented analysis of said methods illustrates multiple challenges in OOD detection. Motivated by the shortcomings of uncertainty-based detectors, a novel mechanism is proposed to efficiently detect OOD samples. The mechanism leverages features extracted from intermediate layers of a classifier and examines their activations. These activations are distinctive for ID samples and can be utilized to distinguish dissimilar data, even if they are classified to the same label. To verify the performance, the proposed method was first amalgamated with an uncertainty-based classifier and tested to detect OOD samples selected to be similar to the ID samples. The method significantly outperforms the uncertainty threshold-based method for OOD detection. Furthermore, the approach was implemented with common classifiers and has shown improved performance on different common OOD datasets.

Degree

thesis:*
Name thesis:degree_name
Doctor of Philosophy - PhD
Level thesis:degree_level
doctoral
Discipline thesis:degree_discipline
Mechanical Engineering
Grantor dc:publisher
University of British Columbia
Year dc:date
2023

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Fayyad, Jamil
Contributors dc:contributor
  • Najjaran, Homayoun
  • Siddiqua, Sumi

Rights

dc:rights
Statement dc:rights
  • Attribution-NonCommercial-NoDerivatives 4.0 International
Language dc:language
eng

Identifiers

dc:identifier.*
Handle dc:identifier
http://hdl.handle.net/2429/84781

Chain of custody

source
Harvested from
University of British Columbia
Base URL
circle.library.ubc.ca/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Fayyad, Jamil. Out-of-distribution detection using inter-level features of deep neural networks. doctoral thesis, University of British Columbia, 2023. http://hdl.handle.net/2429/84781