Li, H. and Abhayaratne, C. orcid.org/0000-0002-2799-7395 (2026) SARE: Structure-aware reliability evaluation and processing for event-based neural systems. In: 2026 International Joint Conference on Neural Networks (IJCNN). 2026 International Joint Conference on Neural Networks (IJCNN), 22-26 Jun 2026, Maastricht, Netherlands. Institute of Electrical and Electronics Engineers, pp. 1-8. ISBN: 9798319531360. ISSN: 2161-4393. EISSN: 2161-4407.
Abstract
Event-based neuromorphic sensors produce sparse and asynchronous streams whose reliability depends not only on global event statistics, but also on the preservation of under-lying spatio-temporal structure. In practice, noise or localized signal loss or aggressive pre-processing can disrupt structural continuity while leaving global statistics largely unchanged, yielding events that appear plausible yet are structurally invalid. Such mismatches can lead to abrupt performance collapse or silent failure in neural systems. In this paper, we propose a structure-preserving local–global graph processing framework for reliability diagnosis in event-based neural systems. Rather than serving as a denoising or enhancement module, it is designed as a diagnostic intervention operator that applies controlled structure-aware graph transformations to probe model sensitivity under progressive structural degradation. The framework jointly captures local neighborhood coherence and global geometric organization, enabling targeted structural interventions without directly optimizing task accuracy. We further introduce a structure-centric evaluation protocol based on geometry-driven objects, multi-scale motion settings and standardized degradation mechanisms that explicitly decouple structural corruption from global event statistics. Experiments show that structural integrity degrades much earlier than global statistical indicators and reliably predicts abrupt failure in both spiking and non-spiking neural models. Compared with purely local or purely global strategies, the proposed local–global formulation is more robust under severe structural degradation and exposes failure modes overlooked by conventional preprocessing and task-driven evaluation. The dataset and codes associated with this work are publicly available at https://github.com/Haiyu12716/SARE and https://doi.org/10.15131/shef.data.31159726.
Metadata
| Item Type: | Proceedings Paper |
|---|---|
| Authors/Creators: |
|
| Copyright, Publisher and Additional Information: | © 2026 The Authors. Except as otherwise noted, this author-accepted version of a paper published in 2026 International Joint Conference on Neural Networks (IJCNN) is made available via the University of Sheffield Research Publications and Copyright Policy under the terms of the Creative Commons Attribution 4.0 International License (CC-BY 4.0), which permits unrestricted use, distribution and reproduction in any medium, provided the original work is properly cited. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ |
| Keywords: | Degradation; Modeling; Statistics; Event detection; Noise reduction; Printing; Streams; Accuracy; Robustness; Cameras |
| Dates: |
|
| Institution: | The University of Sheffield |
| Academic Units: | The University of Sheffield > Faculty of Engineering (Sheffield) > School of Electrical and Electronic Engineering |
| Date Deposited: | 22 Oct 2026 08:28 |
| Last Modified: | 22 Oct 2026 08:28 |
| Status: | Published |
| Publisher: | Institute of Electrical and Electronics Engineers |
| Refereed: | Yes |
| Identification Number: | 10.1109/ijcnn69854.2026.11728562 |
| Open Archives Initiative ID (OAI ID): | oai:eprints.whiterose.ac.uk:247034 |
Download
Filename: 2026_IJCNN_HL_SARE_Final.pdf
Licence: CC-BY 4.0

CORE (COnnecting REpositories)
CORE (COnnecting REpositories)