Abstract

Active learning (AL) is a Machine Learning (ML) strategy that offers a promising solution to challenges posed by the data-intensive tasks by selectively querying the most informative data, reducing the need for extensive labeled datasets. This approach is particularly valuable for fields like cardiac research and scientific simulations, where data labeling becomes expensive and often impractical. Deep Active Learning (DAL) extends this concept to deep learning models, which are typically data hungry, where the focus is more to optimize data selection in discrete search spaces. Conversely, Explorative Active Learning (EAL), is tailored for exploration of continuous search spaces, employing machine learning techniques like Gaussian Processes. A recurring challenge in domain of DAL is difficulty to consistently reproduce the effectiveness of acquisition strategy in diverse experimental settings, a problem often linked to hyperparameters and experimental settings such as data augmentation. However, an often overlooked source of inconsistency is the choice of underlying model. Similarly, despite developments EAL, its applications to critical regions -- such as boundaries or areas of sharp functional change -- remains limited and often requires complex modifications to the existing acquisition strategies. In addition, reusability of trained model in new tasks is limited with strategies that assume that new task would fall under same distribution as the learned task or inclusion of complex techniques as reinforcement learning. Addressing these gaps, this dissertation outlines following key initiatives to try to solve them: (1) a comprehensive investigation of DAL on diverse dataset, network architectures and acquisition strategies to demonstrate the importance of architecture optimization and its impact on effectiveness in DAL. We further extend this to an application on DAL-based surrogate model development in scientific application. (2) The introduction of a novel acquisition strategy that emphasize data acquisition in regions of sharp functional value changes while maintaining minimal change to existing approaches. Additionally, we present a novel active learning framework that bridges DAL and EAL paradigm by combining a neural network-based surrogate model with EAL acquisition strategy, resulting in a reusable model applicable to unseen tasks.

Library of Congress Subject Headings

Deep learning (Machine learning); Computer architecture--Design and construction; Mathematical optimization

Publication Date

2025

Document Type

Dissertation

Student Type

Graduate

Degree Name

Computing and Information Sciences (Ph.D.)

Department, Program, or Center

Computing and Information Sciences Ph.D, Department of

College

Golisano College of Computing and Information Sciences

Advisor

Linwei Wang

Advisor/Committee Member

Richard D. Lange

Advisor/Committee Member

Rui Li

Campus

RIT – Main Campus

Plan Codes

COMPIS-PHD

Download

Share

COinS