Neuromorphic Ensemble Workflow
Date Issued
August 2026
Author(s)
White, Chris Anthony
The University of Tennessee
Advisor(s)
James Plank
Additional Advisor(s)
Garret Rose
Catherine Schuman
Abstract
Ensemble classification has long demonstrated accuracy advantages over single-model approaches, yet applying ensemble methods to Spiking Neural Networks~(SNNs) has traditionally required software-level vote aggregation external to the network, breaking the neuromorphic computation pipeline. We present the Neuromorphic Ensemble Workflow~(NEW), a novel architecture that realizes multiclass ensemble classification entirely within the spiking domain. The NEW decomposes multiclass classification into a set of independently trained binary SNNs, one per class, whose outputs are integrated through a cascade of custom-designed spiking subnetworks for spike-to-time conversion and time comparison, culminating in a time-to-first-spike decoded ensemble output. The result is a single, self-contained SNN that accepts one input presentation and produces a multiclass classification in a single forward pass with no external voting logic required. We present a Time-to-Last-Spike~(TTLS) ensemble crafted by the NEW and evaluate it on the MNIST dataset against multiclass SNN baselines trained under identical conditions. The best NEW configuration achieves 0.9042 accuracy compared to 0.3347 for the multiclass baseline, an improvement of over 57 percentage points, while introducing a post-hoc software tiebreaking procedure that further improves accuracy as ensemble size grows.
Degree
Master of Science
Major
Computer Science
File(s)![Thumbnail Image]()
Name
WHITE-PRIMARY-2026.pdf
Size
1.1 MB
Format
Adobe PDF
Checksum (MD5)
61c8bc3eed3a68d6bea9fa135a6f0160