Abstract

This thesis examines the impact of data augmentation techniques on model performance within a distributed learning framework, focusing on enhancing feature diversity and improving representation for under-represented classes. Data augmentation, commonly used to address data imbalance, significantly influences the feature space learned by deep learning models, with varied effects in distributed settings where data is split across nodes. Our study reveals that inconsistencies in feature learning across nodes reduce the benefits of local augmentation in capturing complex patterns, leading to suboptimal model performance. To address this, we propose a coherent augmentation approach that embeds consistent transformations in the central server, resulting in improved class distinction, and balanced learning across classes. Our findings underscore the potential of embedding augmentation to optimize distributed models, suggesting paths for refining augmentation parameters and enhancing feature space visualization in decentralized environments, with future applications across diverse data structures and distributed learning paradigms.

Date of publication

Fall 12-11-2024

Document Type

Thesis

Language

english

Persistent identifier

http://hdl.handle.net/10950/4790

Committee members

Sagnik Dakshit, Yi Li, Arun Kulkarni

Degree

Master of Science in Computer Science

Thesis Final Approval Prabahar.pdf (100 kB)
Aprroval from Grad school for acceptec thesis.

Share

COinS