Abstract

Over the past decade, deep neural networks have enormously advanced machine perception, especially object classification, object detection, and multimodal scene understanding. But, a major limitation of these systems is that they assume a closed-world setting, i.e., the train and the test distribution match exactly. As a result, any input belonging to a category that the system has never seen during training will not be recognized as unknown. However, many real-world applications often need this capability. For example, self-driving cars operate in a dynamic world where the data can change over time due to changes in season, geographic location, sensor types, etc. Handling such changes requires building models with open-world learning capabilities. In open-world learning, the system needs to detect novel examples which are not seen during training and update the system with new knowledge, without retraining from scratch. In this dissertation, we address gaps in the open-world learning literature and develop methods that enable efficient multimodal open-world learning in deep neural networks.

Library of Congress Subject Headings

Neural networks (Computer science); Machine learning

Publication Date

6-28-2022

Document Type

Dissertation

Student Type

Graduate

Degree Name

Imaging Science (Ph.D.)

Department, Program, or Center

Chester F. Carlson Center for Imaging Science (COS)

Advisor

Christopher Kanan

Advisor/Committee Member

Ferat Sahin

Advisor/Committee Member

Nathan Cahill

Campus

RIT – Main Campus

Plan Codes

IMGS-PHD

Download

Share

COinS