Adaptive deep learning under data scarcity

Kwark, Dou Hoon

Permalink

https://hdl.handle.net/2142/129216 Copy

Description

Title

Adaptive deep learning under data scarcity

Author(s)

Kwark, Dou Hoon

Issue Date

2025-04-17

Director of Research (if dissertation) or Advisor (if thesis)

Kindratenko, Volodymyr

Department of Study

Siebel School Comp & Data Sci

Discipline

Computer Science

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Deep Learning
• Data Scarcity
• Diffusion Model
• Segmentation

Language

eng

Abstract

Deep learning has catalyzed transformative breakthroughs in computer vision and related fields, but these advances often rely on large-scale datasets that are neither readily accessible nor cost-effective in many real-world contexts. In many practical domains—such as medical imaging and geological mapping—collecting large-scale, expertly annotated datasets is prohibitively expensive. This thesis investigates a range of strategies designed to alleviate the pervasive challenge of data scarcity. Through comprehensive studies on both discriminative and generative tasks—including segmentation, super-resolution, modality translation, and inpainting—we demonstrate novel frameworks that preserve strong predictive performance despite limited training data. Our central goal is to show that deep learning under constrained data can still deliver robust results, provided the modeling pipelines are carefully adapted to the problem at hand. First, we propose a hierarchical diffusion-based approach that synthesizes pseudo-healthy medical images with enhanced 3D consistency but moderate computational overhead. Second, we present a fusion strategy that integrates multiple 2D diffusion models into a lightweight 3D representation, improving volumetric realism when data points are limited. Finally, we explore a multi-encoder pipeline that leverages color-space transformations to better segment complex maps, demonstrating its utility in settings like geological digitization. Taken together, these contributions illustrate that addressing data scarcity does not require sacrificing performance. Rather, it calls for more nuanced model design—incorporating domain-specific insights, ensemble architectures, and complementary data transformations. Our experiments show consistent improvements across diverse tasks, highlighting the promise of deep learning solutions that are nimble enough to excel in resource-constrained environments.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129216

Copyright and License Information

Copyright 2025 Dou Hoon Kwark

Owning Collections