Date of Award

Spring 2026

Document Type

Open Access Dissertation

Department

Computer Science and Engineering

First Advisor

Ramtin Zand

Abstract

As machine learning continues to expand across diverse sectors, including healthcare, transportation, environmental monitoring, and consumer electronics, many applications operate under strict constraints on latency, bandwidth, privacy, and power availability. Meeting these demands increasingly requires deploying machine learning at the edge to process data locally rather than relying on remote cloud infrastructure. These limitations highlight the importance of computational efficiency and motivate the exploration of alternative hardware and algorithmic paradigms capable of delivering real-time, accurate inference under tight resource constraints.

Two major paradigms have emerged for edge intelligence: conventional artificial neural networks (ANNs) deployed on highly parallel edge accelerators such as Intel’s Neural Compute Stick 2, Nvidia Jetson, and Google Coral TPUs, and spiking neural networks (SNNs) executed on asynchronous, event-driven neuromorphic hardware such as Intel’s Loihi. While conventional accelerators leverage mature hardware and algorithmic design employing quantization and parallel arithmetic pipelines to achieve high precision and low latency, neuromorphic systems exploit sparse, spike-based temporally dynamic computation inspired by biological neural systems to achieve extreme energy efficiency. This dissertation investigates the performance trade-offs between these paradigms through a hardware-software co-design perspective, progressing from foundational neural coding techniques to the development of heterogeneous neuromorphic and conventional systems.

First, a neural-coding framework is developed for transforming static images into temporally structured spike trains, enabling SNNs to process visual data with reduced spike counts while preserving reconstruction fidelity. Building on this foundation, the dissertation continues by evaluating the effectiveness of neuromorphic computing for practical edge workloads, including assistive technologies like American Sign Language (ASL) recognition. Results demonstrate that neuromorphic systems can achieve orders-of-magnitude reductions in power and energy consumption, extending device runtime for always-on and assistive technologies while maintaining competitive accuracy. Complementary system-level experiments further compare ANN deployments on the Intel Neural Compute Stick 2 with converted SNN implementations on Intel Loihi, revealing an inherent accuracy-latency trade-off in neuromorphic inference, where improved accuracy requires longer presentation durations and can exceed the latency of conventional accelerators.

To leverage the complementary strengths of each paradigm, we introduce a hybrid co-design framework that integrates spiking and non-spiking computation. In this architecture, SNNs extract temporal features, while ANNs provide high-precision in the final decision layers. To emulate such a system deployed on hardware, each respective model is deployed on separate representative edge accelerators, such as Intel's Loihi neuromorphic processor and the NVIDIA Jetson Nano. To facilitate this integration, we propose custom interface hardware, specifically an accumulator circuit, designed to synchronize asynchronous spike streams for synchronous edge processing, thereby creating an efficient heterogeneous system that integrates both neuromorphic and conventional domains. Evaluating this system on event-based gesture recognition workloads demonstrates that hybrid architectures can outperform purely spiking models in both accuracy and latency while consuming significantly less energy than fully ANN-based systems. The results further show that architectures with a small number of spiking convolutional layers acting as temporal feature extractors provide the best balance between efficiency and performance.

Collectively, this dissertation establishes clear guidelines for selecting between neuromorphic, conventional, and hybrid edge-AI architectures. Neuromorphic systems excel in spatiotemporal, always-on sensing environments operating under extreme power constraints, while conventional accelerators are better suited for high-precision, latency-sensitive tasks. Hybrid co-designed systems offer a promising middle ground by combining the temporal efficiency of SNNs with the accuracy and maturity of ANN-based accelerators. These contributions provide a unified framework for the next generation of neuromorphic capabilities by highlighting areas for improvement and how co-design principles can be expanded to create more capable and reliable autonomous systems while alleviating current problems faced by the immense scale and consumption of AI workloads.

Rights

© 2026, Peyton Chandarana

Download

Share

COinS