eCommons

 

Novelty Detection and Analysis in Convolutional Neural Networks

Other Titles

Author(s)

Abstract

While many computer vision researchers race to architect improved convolutional neural networks (CNNs) and increase task accuracy, relatively little work has been done to understand and quantify the types of errors networks make. Regarding classification networks, certain incorrect responses may be better than others. In an animal classification task, for instance, the network categorizes an input image as one out of a set of trained labels. When such a network gets the label incorrect, there are many wrong labels it can choose. There are, however, varying degrees of incorrectness; a network which classifies a German shepherd as a poodle is a better network than one which classifies the German shepherd as a blue whale. In this work, I explore the responses of CNNs to sets of images in both known classes and novel classes (those which the network was not trained on). I analyze the predicted labels for these sets of images, as well as the predicted label distribution over many images in a given class. This paper also includes a discussion of the hierarchy of predicted labels and true labels, and what a network’s response in terms of higher-level categories reveals about its ability to generalize. Finally, I will discuss how humans and networks each measure the similarity between images, and show novelty's effect on networks' agreement with people.

Journal / Series

Volume & Issue

Description

97 pages

Sponsorship

Date Issued

2020-05

Publisher

Keywords

alexnet; CNN; computer vision; convolutional neural network; imagenet; novelty

Location

Effective Date

Expiration Date

Sector

Employer

Union

Union Local

NAICS

Number of Workers

Committee Chair

Hariharan, Bharath

Committee Co-Chair

Committee Member

Field, David

Degree Discipline

Computer Science

Degree Name

M.S., Computer Science

Degree Level

Master of Science

Related Version

Related DOI

Related To

Related Part

Based on Related Item

Has Other Format(s)

Part of Related Item

Related To

Related Publication(s)

Link(s) to Related Publication(s)

References

Link(s) to Reference(s)

Previously Published As

Government Document

ISBN

ISMN

ISSN

Other Identifiers

Rights

Attribution 4.0 International

Types

dissertation or thesis

Accessibility Feature

Accessibility Hazard

Accessibility Summary

Link(s) to Catalog Record