Dissertation
Scalable and interpretable deep learning via convex architectures and bio-inspired strategies
Doctor of Philosophy (Ph.D.), Drexel University
Dec 2024
DOI:
https://doi.org/10.17918/00010819
Abstract
Deep learning has achieved remarkable success across numerous domains, yet persistent challenges in robustness, interpretability, and computational efficiency limit its application in high-stakes environments. This dissertation addresses these challenges by introducing a comprehensive framework that integrates bio-inspired neural architectures, convex optimization principles, and sparse representations to enhance both the theoretical foundations and practical applications of deep learning systems. The work begins by investigating bio-inspired neural models that leverage sparsity as a core principle to enhance computational efficiency and interpretability. Building on this foundation, a convexified neural network is introduced as a surrogate model for process optimization tasks, demonstrating its ability to solve complex problems through a simplified closed-form solution while incorporating domain-specific constraints, such as operational feasibility and physical parameter limits, to ensure robust and optimized outcomes. This foundation enables the development of a novel convex attention mechanism, which dynamically weights input features to improve learning focus and resource allocation, and a convex capsule network, which advances spatial hierarchical learning by capturing complex feature relationships while maintaining robustness and interpretability. Through comparative evaluations in wearable gesture recognition and industrial optimization tasks, these convexified frameworks consistently demonstrate superior performance over traditional neural networks. Key results include faster convergence, enhanced robustness to input variability, and improved interpretability, with measurable benefits such as reduced computation time and increased predictive accuracy. By addressing critical challenges in deep learning, this dissertation highlights the transformative potential of convex optimization-based approaches, offering a robust foundation for scalable, interpretable, and reliable machine learning solutions.
Metrics
45 File views/ downloads
31 Record Views
Details
- Title
- Scalable and interpretable deep learning via convex architectures and bio-inspired strategies
- Creators
- Daniel Ethan Schwartz
- Contributors
- Ali Shokoufandeh (Advisor)
- Awarding Institution
- Drexel University
- Degree Awarded
- Doctor of Philosophy (Ph.D.)
- Publisher
- Drexel University; Philadelphia, Pennsylvania
- Number of pages
- xiii, 141 pages
- Resource Type
- Dissertation
- Language
- English
- Academic Unit
- Computer Science (Computing) (2013-2026); College of Computing and Informatics (2013-2026); Drexel University
- Other Identifier
- 991022020016804721