Semester of Graduation

Spring 2026

Degree

Master of Engineering (ME)

Department

Computer Science

Document Type

Thesis

Abstract

Neuromorphic computing seeks to reproduce core computational principles of biological neural systems to enable low-power, stateful, event-driven intelligence. By exploiting sparse spike-based computation and local state, neuromorphic platforms reduce the data movement and continuous switching that dominate energy use in conventional architectures, enabling milliwatt-scale, always-on edge deployment. Yet most modern computing remains optimized for dense, synchronous floating-point workloads, and spiking neural networks have not matched the empirical breadth of deep learning across real-world domains.

This thesis investigates how spiking neural networks can be engineered to meet established performance standards in high-impact applications while operating under strict edge and hardware constraints. Four neuromorphic systems are developed and evaluated across domains of significant societal and technological relevance: NeuroRad for nuclear radiation source detection, NeuroEEG for seizure detection, NeuroGuard for uranium hexafluoride gas enrichment identification, and NeuroIDS for network intrusion detection. Together, these systems span healthcare, nuclear safeguards, critical infrastructure monitoring, and edge cybersecurity.

All systems share a neuromorphic foundation built on evolutionary optimization using the EONS framework, but each required targeted adaptations across the computational pipeline. For NeuroRad and NeuroEEG, ensembles were essential to satisfy stringent detection reliability and false alarm constraints, demonstrating that diversity within evolved populations can improve robustness. NeuroGuard required physics-informed spectral binning and feature sensitivity tuning to capture subtle UF6 gas enrichment signatures under pressure-dependent attenuation. NeuroIDS required a novel spike encoding scheme and asynchronous processing workflow to accommodate irregular, event-driven network traffic.

Across all domains, compact recurrent spiking neural networks evolved under hardware-aware constraints achieve performance comparable to or exceeding conventional machine learning baselines while maintaining milliwatt-scale inference capability on FPGA-based neuromorphic platforms. In parallel, this work presents a survey of spiking neural network applications, identifying prevailing architectural trends, training limitations, and open challenges that motivate these adaptations. Collectively, these contributions demonstrate that competitive neuromorphic performance depends on co-design across encoding, objectives, aggregation, and deployment constraints, and they support spiking neural networks as a practical direction for ultra-low-power real-world intelligence.

Date

4-2-2026

Committee Chair

James Ghawaly

LSU Acknowledgement

1

LSU Accessibility Acknowledgment

1

Download

Available for download on Thursday, March 25, 2027

Share

COinS