pith. sign in

arxiv: 2605.00030 · v1 · submitted 2026-04-23 · 💻 cs.AR · cs.SY· eess.SY

Shooting Neutrons at Neurons: Radiation Testing of a Spiking Neural Network on Flash-Based FPGAs

Wim Nijsink , Bruno Endres Forlin , Amirreza Yousefzadeh , Marco Ottavi This is my paper

Pith reviewed 2026-05-09 21:01 UTC · model grok-4.3

classification 💻 cs.AR cs.SYeess.SY
keywords radiation testingspiking neural networksFPGASEUsynaptic plasticityneuromorphic computingfault toleranceonline learning
0
0 comments X

The pith

Enabling online synaptic plasticity in a spiking neural network processor allows it to tolerate more radiation-induced bit flips before classification fails.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The authors map an open-source spiking neural network processor called ODIN onto a flash-based FPGA and subject it to a high-energy neutron beam while it performs MNIST digit classification. They track both the network's accuracy and the state of its synaptic memory to extract single-event upset cross-sections and build a calibrated fault model. Comparing a static inference-only mode against one with Spike-Dependent Synaptic Plasticity enabled shows that the learning rule extends the time until accuracy collapses and permits partial recovery from accumulated faults. The plasticity mechanism adds only modest hardware cost while providing this resilience. The work targets neuromorphic hardware for space and avionics, where radiation tolerance and graceful degradation matter.

Core claim

Irradiating the ODIN processor with SDSP under continuous MNIST inference demonstrates that turning on spike-dependent synaptic plasticity significantly lengthens the interval before application-level failure occurs and enables partial restoration of synaptic weights from bit flips, at modest extra hardware cost, compared with inference-only operation.

What carries the argument

Spike-Dependent Synaptic Plasticity (SDSP), the on-chip learning rule that continuously updates synaptic weights according to spike timing during task execution.

If this is right

  • Neuromorphic processors with on-chip plasticity can operate longer under radiation before accuracy is lost.
  • Continuous learning supplies a form of self-repair against accumulated single-event upsets.
  • The approach incurs only modest resource overhead on flash-based FPGAs.
  • Beam-derived SEU cross-sections and fault models can guide simulation-based reliability analysis for similar SNN designs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same plasticity mechanism might mitigate other transient faults, such as those from voltage droops or temperature extremes, in edge devices.
  • Combining SDSP with conventional error-correcting codes could further raise the radiation tolerance threshold.
  • Long-duration missions might reveal whether repeated recovery cycles eventually exhaust synaptic dynamic range.

Load-bearing premise

The neutron-beam test conditions and the fault model derived from them represent the radiation spectrum and cumulative effects that the hardware would actually encounter in space or avionics.

What would settle it

A real satellite or avionics flight experiment that shows failure times or recovery behavior inconsistent with the model's predictions from the neutron data would falsify the representativeness of the test.

Figures

Figures reproduced from arXiv: 2605.00030 by Amirreza Yousefzadeh, Bruno Endres Forlin, Marco Ottavi, Wim Nijsink.

Figure 1
Figure 1. Figure 1: Experimental setup. Modules in green have automatic TMR applied; red modules are part of the test fixture. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 4
Figure 4. Figure 4: Accuracy under radiation for 60000 with learning [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Accuracy under radiation over time for 6000 images. Top: no learning, [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy over time for 10000 images with simulated radiation. One [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Accuracy over time for 10000 images with simulated radiation. One [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
read the original abstract

Neuromorphic, or spiking, processors are increasingly being considered for use in harsh, radiation-prone environments such as space and avionics, where energy efficiency and graceful degradation are essential. In this study, we propose and experimentally validate a radiation-testing methodology specifically designed for neuromorphic processors that employ on-chip synaptic plasticity. We map the open-source ODIN SNN processor with Spike-Dependent Synaptic Plasticity (SDSP) onto the FPGA and expose it to a high-energy neutron beam while continuously monitoring MNIST classification accuracy and recording the synaptic state. From these measurements, we extract SEU cross-sections for ODIN's synaptic memory and develop a calibrated fault model to inform a complementary fault-injection campaign. By comparing inference-only and online-learning configurations, we demonstrate that enabling SDSP can significantly extend the time to application-level failure and enable partial recovery from accumulated bit flips, with modest hardware overhead.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents an experimental radiation-testing methodology for the open-source ODIN spiking neural network processor with Spike-Dependent Synaptic Plasticity (SDSP) mapped to a flash-based FPGA. The system is exposed to a high-energy neutron beam while monitoring MNIST classification accuracy and synaptic states; SEU cross-sections are extracted to build a calibrated fault model that is then used in a complementary fault-injection campaign. By comparing inference-only and online-learning configurations, the authors claim that enabling SDSP significantly extends time to application-level failure and permits partial recovery from accumulated bit flips, at modest hardware overhead.

Significance. If the quantitative results hold, the work supplies concrete experimental data on SEU effects in neuromorphic hardware and demonstrates a practical benefit of on-chip plasticity for graceful degradation under radiation. The direct comparison of configurations and the derivation of a fault model from beam measurements are strengths that could inform design of radiation-tolerant spiking processors for space or avionics.

major comments (2)
  1. [fault-model calibration section (likely §4)] The central claim that SDSP extends time-to-failure and enables recovery rests on the calibrated fault model derived from the neutron-beam data. The manuscript does not provide the exact calibration procedure, error bars on the extracted SEU cross-sections, or exclusion criteria for the measured events, preventing independent verification of the quantitative predictions used in the injection campaign.
  2. [discussion and conclusions (likely §6)] The generalizability of the SDSP benefit to space and avionics environments assumes the neutron-induced upset signatures are representative. Neutron spallation produces primarily indirect ionization, whereas space environments include direct ionization from protons and heavy ions across a wide LET range; the paper does not include a concrete cross-check (e.g., comparison of upset patterns or additional proton-beam data) to test whether the observed SDSP advantage persists under those conditions.
minor comments (2)
  1. [results and abstract] The abstract and results sections would benefit from explicit statement of the number of experimental runs, statistical tests used to support 'significantly extend,' and the precise definition of 'partial recovery' (e.g., accuracy threshold and recovery time window).
  2. [implementation and overhead discussion] Hardware-overhead figures for the SDSP implementation are mentioned but lack a breakdown (LUTs, BRAM, power) relative to the inference-only baseline; a small table would improve clarity.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback and positive assessment of the significance of our work. We address each major comment below with point-by-point responses and indicate where revisions have been made to the manuscript.

read point-by-point responses

  1. Referee: [fault-model calibration section (likely §4)] The central claim that SDSP extends time-to-failure and enables recovery rests on the calibrated fault model derived from the neutron-beam data. The manuscript does not provide the exact calibration procedure, error bars on the extracted SEU cross-sections, or exclusion criteria for the measured events, preventing independent verification of the quantitative predictions used in the injection campaign.

    Authors: We agree that the original manuscript provided insufficient detail on the fault-model calibration to support independent verification. In the revised manuscript we have expanded the relevant section to include the complete calibration procedure: SEU cross-sections are extracted via σ = N_events / (fluence × sensitive_bits), with 95% confidence intervals computed from Poisson statistics on the observed event counts; exclusion criteria discard runs showing background rates above 3σ or temporal clustering consistent with multiple-bit upsets. These additions allow full reproduction of the model parameters used for the fault-injection campaign. revision: yes

  2. Referee: [discussion and conclusions (likely §6)] The generalizability of the SDSP benefit to space and avionics environments assumes the neutron-induced upset signatures are representative. Neutron spallation produces primarily indirect ionization, whereas space environments include direct ionization from protons and heavy ions across a wide LET range; the paper does not include a concrete cross-check (e.g., comparison of upset patterns or additional proton-beam data) to test whether the observed SDSP advantage persists under those conditions.

    Authors: We acknowledge that neutron spallation primarily induces indirect ionization while space radiation includes direct ionization from protons and heavy ions. The SDSP recovery mechanism acts on the resulting bit-flip patterns in synaptic weights rather than on the ionization physics itself; therefore the observed benefit is expected to be largely independent of the radiation source provided the upset statistics are comparable. Nevertheless, we lack proton or heavy-ion beam data that would constitute a direct cross-check. The revised discussion now explicitly states this limitation and notes that the reported advantage is demonstrated under the tested neutron environment. revision: partial

standing simulated objections not resolved
  • We do not possess proton-beam or heavy-ion irradiation data that would allow a direct empirical cross-check of whether the SDSP advantage persists under direct-ionization conditions typical of space.

Circularity Check

0 steps flagged

No circularity detected in the experimental claims

full rationale

The paper's methodology is grounded in direct experimental data from neutron beam exposure of the ODIN SNN processor on FPGA. SEU cross-sections are extracted from measured bit flips and accuracy changes, and a fault model is calibrated from this data for additional injection studies. However, the key result—that SDSP enables extended time to failure and partial recovery—is demonstrated through comparative observations in the inference-only versus online-learning setups during the actual radiation testing. No mathematical derivations or predictions reduce to fitted inputs by construction, and there are no load-bearing self-citations that form a circular chain. The work is self-contained as an empirical validation study.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on experimental measurements and a calibrated fault model derived from them; no free parameters, axioms, or invented entities are introduced beyond standard hardware-testing assumptions.

pith-pipeline@v0.9.0 · 5470 in / 1069 out tokens · 33575 ms · 2026-05-09T21:01:12.493881+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

11 extracted references · 11 canonical work pages

  1. [1]

    From ground to orbit: A robust and efficient test methodology for risc-v soft-cores,

    B. Forlin, K. B ¨ohmer, C. Cazzaniga, P. Rech, G. Furano, N. Alachiotis, and M. Ottavi, “From ground to orbit: A robust and efficient test methodology for risc-v soft-cores,”IEEE Transactions on Device and Materials Reliability, vol. 25, no. 1, pp. 27–36, 2025

    work page 2025

  2. [2]

    Yield and reliability challenges at 7nm and below,

    A. J. Strojwas, K. Doong, and D. Ciplickas, “Yield and reliability challenges at 7nm and below,” in2019 Electron Devices Technology and Manufacturing Conference (EDTM). IEEE, 2019, pp. 179–181

    work page 2019

  3. [3]

    Data center silent data errors: Implications to artificial intelligence workloads & mitigations,

    B. Bittel, M. Shamsa, B. Inkley, A. Gur, D. Lerner, and M. Adams, “Data center silent data errors: Implications to artificial intelligence workloads & mitigations,” in2024 IEEE International Reliability Physics Sympo- sium (IRPS). IEEE, 2024, pp. 1–5

    work page 2024

  4. [4]

    A survey examining neuromorphic architecture in space and challenges from radiation,

    J. Naoukin, M. Isik, and K. Tiwari, “A survey examining neuromorphic architecture in space and challenges from radiation,”arXiv preprint arXiv:2311.15006, 2023

    work page arXiv 2023

  5. [5]

    Radiation tolerance and mitigation for neuromorphic processors,

    J. Schumann, “Radiation tolerance and mitigation for neuromorphic processors,” 2022

    work page 2022

  6. [6]

    On the trustworthiness of spiking neural networks and neuromorphic systems,

    T. Spyrou, H.-G. Stratigopoulos, I. Alouani, S. Hamdioui, and A. Ge- bregiorgis, “On the trustworthiness of spiking neural networks and neuromorphic systems,” in2025 IEEE European Test Symposium (ETS). IEEE, 2025, pp. 1–10

    work page 2025

  7. [7]

    Device reliability affecting coding schemes in neuromorphic circuits,

    S. M. Picardo, J. B. Shaik, S. Singhal, and N. Goel, “Device reliability affecting coding schemes in neuromorphic circuits,” in2022 IEEE Region 10 Symposium (TENSYMP). IEEE, 2022, pp. 1–6

    work page 2022

  8. [8]

    Dynamic reliability management in neuromorphic computing,

    S. Song, J. Hanamshet, A. Balaji, A. Das, J. L. Krichmar, N. D. Dutt, N. Kandasamy, and F. Catthoor, “Dynamic reliability management in neuromorphic computing,”ACM Journal on Emerging Technologies in Computing Systems (JETC), vol. 17, no. 4, pp. 1–27, 2021

    work page 2021

  9. [9]

    Radiation test performance of the intel loihi neuromorphic processor,

    R. Scrofano, S. C. Davis, and J. L. Taggart, “Radiation test performance of the intel loihi neuromorphic processor,” in2024 IEEE Space Com- puting Conference (SCC). IEEE, 2024, pp. 86–92. 6

    work page 2024

  10. [10]

    A 0.086-mm 2 12.7- pj/sop 64k-synapse 256-neuron online-learning digital spiking neuro- morphic processor in 28-nm cmos,

    C. Frenkel, M. Lefebvre, J.-D. Legat, and D. Bol, “A 0.086-mm 2 12.7- pj/sop 64k-synapse 256-neuron online-learning digital spiking neuro- morphic processor in 28-nm cmos,”IEEE transactions on biomedical circuits and systems, vol. 13, no. 1, pp. 145–158, 2018

    work page 2018

  11. [11]

    Memory wall is not gone: A critical outlook on memory architecture in digital neuromorphic computing,

    A. Yousefzadeh, S. Sohail, and A. L. Varbanescu, “Memory wall is not gone: A critical outlook on memory architecture in digital neuromorphic computing,” in2025 IEEE Computer Society Annual Symposium on VLSI (ISVLSI), vol. 1. IEEE, 2025, pp. 1–4

    work page 2025