Online Archive of University of Virginia Scholarship
Tailored Deep Learning for Enhanced Performance in Resource-Constrained Settings128 views
Author
Shrivastava, Aman, Computer Science - School of Engineering and Applied Science, University of Virginia
Advisors
Fletcher, Tom, EN-Elec & Comp Engr Dept, University of Virginia
Abstract
Tailored deep learning methods enable consistent and critical progress toward enhancing reasoning capabilities, particularly in data- and resource-constrained settings like healthcare. Improvements in knowledge distillation using information maximization enables cheaper optimization and domain-transfer for large vision models. Information-efficient contrastive learning for aligning images with textual data leads to better performance on small datasets. Structure-preserving generative adversarial networks help minimize visual variations in medical images. Conditional diffusion can be used to develop end-to-end models for synthesizing inherently-annotated histology tissue samples with pixel-perfect nuclei localization.
Shrivastava, Aman. Tailored Deep Learning for Enhanced Performance in Resource-Constrained Settings. University of Virginia, Computer Science - School of Engineering and Applied Science, PHD (Doctor of Philosophy), 2024-11-27, https://doi.org/10.18130/yfg9-qx29.
