ORCID

0009-0005-9177-6633

Keywords

Label Noise Modeling, Robust Deep Learning, Multi-Label Classification, Tensor Factorization, Intrinsically Disordered Proteins, Hybrid Machine Learning Models

Abstract

This thesis investigates robust and hybrid deep learning with emphasis on improving accuracy and generalization in challenging real-world settings, including learning under noisy supervision and applications to complex scientific data.

In the first part, we address instance-dependent label noise by using a tensor factorization–based methodology that models the noise transition process as an instance-dependent decomposition. We also utilize low-rank block-term structures to capture class-specific and instance-specific correlations. Moreover, we include a robust loss function and rank regularization to achieve stable optimization during model training. We validated this with experiments on benchmark datasets, like Fashion-MNIST, CIFAR-10, and CIFAR-10N, demonstrating higher accuracy and robustness compared to existing methods.

The second part focuses on multi-label recognition under noisy supervision. In this situation, the inter-label dependencies often lead to complex label confusion. We introduce a probabilistic multi-label confusion mixture model that includes latent variables to represent class-driven confusions and applies sparsity regularization to encourage parsimonious label predictions. Evaluations on VOC-2007 and VOC-2012 show consistent improvements of up to 5% in mean average precision across varying noise levels.

Finally, we develop an attention-guided hybrid framework for scientific data modeling that enhances robustness, accuracy, and interpretability. Specifically, we focused our study on intrinsically disordered proteins, designing machine learning models for predicting their conformational properties. Our model combines sequence-based representations with physical features. This helps improve prediction accuracy and interpretability.

These contributions provide useful insights and methodologies for robust and hybrid deep learning, particularly across noisy, complex, and real-world data settings. Our findings demonstrate the importance of developing flexible, generalizable approaches for learning in such scenarios and that deep learning can remain accurate and interpretable even when data are limited or noisy. Altogether, this work shows how combining robustness and hybrid designs leads to more reliable deep learning models capable of performing well across diverse domains.

Completion Date

2025

Semester

Fall

Committee Chair

Ibrahim, Shahana

Degree

Master of Science in Computer Engineering (M.S.Cp.E.)

College

College of Engineering and Computer Science

Department

Electrical and Computer Engineering

Format

PDF

Identifier

DP0029844

Document Type

Thesis

Campus Location

Orlando (Main) Campus

Download

Share

COinS