From Connection Loss to Collision: Closed-Loop Fault Assessment of a Fly-Connectome SNN
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: I'm Ines.
Marcus: And I'm Marcus, here with Yuki.
Ines: We're looking at the paper: "From Connection Loss to Collision".
Marcus: Here's the gist of the paper. The public release of the complete adult fruit-fly brain connectome enables brain-scale spiking-network studies, motivating closed-loop assessment of connection loss Across 576 test episodes in paired campaigns, the baseline mapping at 50% connection-row deletion yields 11/48 proportional and 0/48 relay collisions by 1 s, despite paired P/R trigger/burst schedules The prespecified secondary relay-minus-proportional net-loss difference is −22.9 percentage points (95% seed-cluster CI
−33.3, −12.5: ) These results quantify finite-horizon collisions and the gains and limits of record checks Network Architecture and Experimental Setup The experiment utilizes the FlyWire v783-derived model, which comprises 138,639 LIF neurons and 15,091,983 signed, contact-weighted connection rows The network dynamics are governed by specific parameters including a reset condition at-45 mV and input starting at 100 ms with a visual period δv = 0.020 s Sensory inputs are processed through LC4 and LPLC2 cells, with command features provided by eight descending cells The system involves a synthetic pursuit–escape plant where the pursuer starts 0.25 m away at ±30◦ and targets the agent’s segment-start position Fault Assessment Metrics The primary endpoints for assessment are defined as paired loss L = YH(1 − YF), gain G = (1−YH)YF, and activity exposure Secondary metrics include the relay-minus-proportional contrast and warnings, which are prespecified to be secondary The net loss D = L−G is derived from these primary endpoints Specific ratios such as Rstop and Rcommon are used to measure cumulative spikes S and T∗ = min(TF, TH) Conditional Performance Analysis The first-contact loss depends on the declared horizon, as shown in Fig.
Ines: 3(a), where RF0 has nine additional P contacts in the final 100 ms compared to other conditions The prespecified secondary R-minus-P net-loss contrast was −22.9 pp for RF-0 (Fig.
Marcus: 3(b)) However, A/B yielded smaller contrasts of −4.2 pp, establishing neither equivalence nor an isolated RF effect Post-hoc fault-minus-healthy trigger-time differences averaged 80.0/57.5/57.5 ms in RF-0/A/B across bearing pairs Audit and Consistency Checks The audit gain quantifies the impact of checking mechanisms, where for 87 corruptions, A/B/C/D detected 36/54/75/81 cases respectively The author-designed catalogue detects 81/87 corruptions versus 36/87 for trusted headers, accepting all 20 distinct valid inputs (six misses) The frozen command-rate warning detects only 2/13 baseline collisions across all tested deletion levels Evidence concerns one graph, an assumed RF family, a synthetic plant, unequal healthy reserve and sparse failures Functional row loss establishes no biological tolerance or hardware safety benefit The study quantifies conditional task preservation and record-consistency gains for evaluating how future device-level nonidealities propagate through neuromorphic controllers to closed-loop behavior The results show that walk-only contact by 0.8 s reflects plant speeds, indicating finite-horizon preservation rather than sustained avoidance The study concludes that trace consistency cannot authenticate neural execution References
1: P.
Ines: K.
Title and authors: Ines: So we’re looking at the paper "From Connection Loss to Collision: Closed-Loop Fault Assessment of a Fly-Connectome SNN," and basically, it uses that whole brain map to test if losing connections actually causes real collisions when the system is running autonomously.
Marcus: Right, it takes this massive connectome structure and runs it through synthetic pursuit escape scenarios to see exactly how much record consistency you lose when connections start failing.
Yuki: From a population genetics view, I wonder if seeing this kind of structural failure detection helps us understand how robust these neural circuits are across different species or even developmental stages.
Ines: They found that even when they deleted fifty percent of the connection rows, they still saw eleven out of forty-eight proportional and zero out of forty-eight relay collisions within one second.
Marcus: That’s a big number for collision rate, especially with the paired trigger schedules in place. It suggests there's a lot of redundancy or resilience built into this fly system that we didn't expect.
Yuki: But then they got into the secondary metric, which is the relay minus proportional net-loss difference—they found that was negative twenty-two point nine percentage points with a confidence interval between negative thirty three point three and negative twelve point five.
Ines: That secondary number tells you something specific about the way the loss is distributed across different connection types, not just a simple total count of broken lines.
Marcus: It helps quantify those finite horizon collisions and shows how good or bad those record checks are in predicting what's going to happen before a collision happens.
Yuki: The study also looked at how far ahead you declare your horizon; they found that RF zero had nine additional proportional contacts in the final one hundred milliseconds compared to other conditions.
Ines: So it shows that declaring a longer time window actually helps preserve more connections in the last part of the task, which is interesting for understanding temporal dynamics.
Marcus: They also compared different scenarios like RF zero against A and B, and the secondary contrast was much smaller—negative four point two percentage points for A and B.
Yuki: That means declaring a longer horizon doesn't give you that big advantage across all conditions; the effect is not universal.
Ines: Then they looked at trigger time differences between fault minus healthy conditions, which averaged eighty point zero over fifty seven point five over fifty seven point five milliseconds across those bearing pairs.
Marcus: That average shows how much the timing of a fault versus a healthy event differs when you're dealing with these various RF configurations.
Yuki: They also had an audit check on the detection mechanisms themselves; for eighty seven corruptions, the author designed catalogue detected eighty one cases versus thirty six for just using trusted headers.
Ines: That shows that having a custom catalogue is way better at finding faults than relying on standard checks alone.
Marcus: So, overall, this work shows how checking things in a closed loop can tell you about the limits of prediction and consistency when you try to evaluate how hardware nonidealities affect these neural controllers.
The paper's summary: Ines: Now they’re talking about how to make these fault assessments better than just looking at the final numbers they got from the test runs. They're suggesting ways to use those paired campaigns and stage-resolved measurements for law discordant contacts under shared schedules.
Marcus: Right, they’re suggesting using those paired campaigns and stage-resolved measurements to get a better look at law discordant contacts under shared schedules.
Yuki: From a population genetics standpoint, I want to know if using these paired, stage-resolved task measurements helps us get a better picture of the underlying biological process across development.
Ines: They think this method lets you measure things like uncertain RF transfer and frozen prefix warning failures at the same time.
Marcus: It’s about getting a more detailed look at the conditional task preservation and record consistency audit gains, which is way more precise than just looking at one final collision number.
Yuki: That makes sense; if you can resolve those differences across stages, you can map out how the system handles failure in real time, which connects to how species evolve their resilience.
Ines: They also mentioned an author designed fly delayed tank catalogue to detect corruptions by comparing checks plus model specific replay against trusted headers.
Marcus: It sounds like they’re proposing a way for the system to distinguish between different fault types, which is a key improvement because it tells you what kind of failure you're seeing.
Yuki: That ability to classify the fault source is important because it moves us beyond just saying something broke and starts giving us structural information about how the hardware failed.
Ines: They found that this new system can distinguish between eighty one corruptions versus thirty six for trusted headers, which is a solid improvement over just looking at the raw detection rates.
Marcus: That’s a big win for verification because it shows you have a more robust way to filter out false alarms or misinterpret what the neural network is actually doing.
Yuki: It suggests that we can get deeper insights into how these systems respond to different stresses, which supports the idea that structure matters when we study these complex biological machines.
Ines: But they also acknowledged some limitations; they're still working with one graph and an assumed RF family, and the synthetic plant doesn't perfectly match a real biology.
Marcus: They flagged those issues too, which is important because it keeps us grounded; the results are strong but we can’t assume this holds up in every biological context without more validation.
Yuki: So what they’re saying is that the method has potential for deeper structural analysis, but we still need to build a bridge from these synthetic models back to real-world biological systems.
The paper's improvements: Ines: So we’re wrapping up on "From Connection Loss to Collision: Closed-Loop Fault Assessment of a Fly-Connectome SNN" by summarizing how this study uses closed loops to assess connection loss in a fly brain model.
Marcus: Exactly, it boils down to using the whole connectome map to test if losing connections actually causes real collisions when the system is running autonomously.
Yuki: From my side, it’s interesting because we’re seeing how population genetics might be able to use these kinds of structural assessments to predict resilience in other complex systems.
Ines: They showed that even with a fifty percent connection row deletion, they still saw eleven out of forty-eight proportional and zero out of forty-eight relay collisions within one second.
Marcus: That’s the main finding for me; it suggests there's significant redundancy in this fly system that keeps it from immediately failing when connections are lost.
Yuki: But then they got into the secondary metric, which was negative twenty-two point nine percentage points, with a confidence interval between negative thirty three point three and negative twelve point five.
Ines: That secondary number gives you a specific way to measure the loss difference across different connection types, not just a simple total count of broken lines.
Marcus: It helps quantify those finite horizon collisions and shows the limits of what you can reliably predict before a collision happens.
Marcus: That average shows how much the timing of a fault versus a healthy event differs when you're dealing with these various RF configurations. And finally, they had an audit check on the detection mechanisms themselves; for eighty seven corruptions, the author designed catalogue detected eighty one cases versus thirty six for just using trusted headers.
Marcus: So, while it’s useful for modeling how things propagate through these systems, it doesn't automatically mean the system is safe from every single kind of failure mode.
Ines: That means the system can only reliably predict a collision within a window determined by how fast the fly is moving toward its target.
Marcus: So for someone just listening to the show, it’s about quantifying those record consistency gains while being aware of where those checks actually stop working.
Yuki: It’s a valuable tool, but we have to remember the caveats about the synthetic nature of the plant and the specific assumptions they made about healthy reserves.
Ines: Right. It shows what's possible with this kind of brain scale study, but it also highlights where we need more real-world data or better modeling to make these assessments fully reliable.
Conclusion: Ines: So we're wrapping up on "From Connection Loss to Collision: Closed-Loop Fault Assessment of a Fly-Connectome SNN," which really shows how testing a system in a closed loop lets you assess connection loss in real time using that whole brain map.
Marcus: Exactly, it boils down to taking this connectome and running it through synthetic pursuit escape scenarios to see how much record consistency you lose when things break down.
Yuki: From my side, I’m still thinking about how this kind of structural failure detection could help us understand the broader resilience of neural circuits across different species.
Yuki: But then they got into that secondary metric, which is the relay minus proportional net-loss difference—they found that was negative twenty-two point nine percentage points with a confidence interval between negative thirty three point three and negative twelve point five.
Marcus: It helps quantify those finite horizon collisions and how good or bad those record checks are, showing the limits of what you can reliably predict before a collision happens.
Yuki: So for someone just listening to the show, it suggests that checking things in a closed loop can give us information about how far into the future we can reliably predict a collision before it happens.
Ines: Exactly. They also looked at trigger time differences between fault-minus-healthy conditions, and those averaged eighty point zero over fifty seven point five over fifty seven point five milliseconds across those bearing pairs for RF zero, A, and B.
Yuki: And finally, they had an audit check on the detection mechanisms themselves; for eighty seven corruptions, the author designed catalogue detected eighty one cases versus thirty six for just using trusted headers.
Marcus: But they also noted that a frozen command rate warning only picked up two out of thirteen baseline collisions across all those deletion levels, which is a bit disappointing for the warning system itself.
Ines: It’s a good tool for understanding conditional task preservation, but they also flagged limitations like using a synthetic plant and assuming unequal healthy reserves.
Marcus: So that’s where we are on this one, looking at the results from "From Connection Loss to Collision: Closed-Loop Fault Assessment of a Fly-Connectome SNN."
Waseda University
q-bio.NC, cs.NE
Submitted: 2026-10-31
Updated: 2026-10-31
Comments: 5 pages, 4 figures
Code: https://github.com/Scophield/drosophila-fault-injection
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 88/100
The gist: The public release of the complete adult fruit-fly brain connectome enables brain-scale spiking-network studies…
Key concepts
- Connectome
- This is a detailed map of all the connections between neurons in the fly's brain. It includes information about every possible link between 138,639 neurons, which allows researchers to study how removing specific connections affects overall brain function.
- Closed-Loop Fault Assessment
- This involves testing a system that uses its own performance data to check for faults in real-time. In this study, it means using the brain's behavior during tasks to see if connection errors are causing measurable problems and how well the system can detect them.
- Finite-Horizon Collisions
- These are specific events where a fault causes a problem that only lasts for a limited time, determined by the task's duration. The study shows that these collisions happen and helps define the limits of how much performance can be maintained when connections are lost.
- Record Consistency Audit
- This is a method used to verify if the system's output remains consistent across different testing sessions. By comparing results from paired campaigns, researchers can audit whether the system is correctly handling errors and maintaining reliability over time.
Terminology
Summary
The public release of the complete adult fruit-fly brain connectome enables brain-scale spiking-network studies, motivating closed-loop assessment of connection loss Across 576 test episodes in paired campaigns, the baseline mapping at 50% connection-row deletion yields 11/48 proportional and 0/48 relay collisions by 1 s, despite paired P/R trigger/burst schedules The prespecified secondary relay-minus-proportional net-loss difference is −22. p. 09 percentage points (95% seed-cluster CI [−33. p. 03, −12.5]) These results quantify finite-horizon collisions and the gains and limits of record checks
Network Architecture and Experimental Setup
The experiment utilizes the FlyWire v783-derived model, which comprises 138,639 LIF neurons and 15,091,983 signed, contact-weighted connection rows The network dynamics are governed by specific parameters including a reset condition at-45 mV and input starting at 100 ms with a visual period δv = 0.020 s Sensory inputs are processed through LC4 and LPLC2 cells, with command features provided by eight descending cells The system involves a synthetic pursuit–escape plant where the pursuer starts 0. p. 025 m away at ±30◦ and targets the agent’s segment-start position
Fault Assessment Metrics
The primary endpoints for assessment are defined as paired loss L = YH(1 − YF), gain G = (1−YH)YF, and activity exposure Secondary metrics include the relay-minus-proportional contrast and warnings, which are prespecified to be secondary The net loss D = L−G is derived from these primary endpoints Specific ratios such as Rstop and Rcommon are used to measure cumulative spikes S and T∗ = min(TF, TH)
Conditional Performance Analysis
The first-contact loss depends on the declared horizon, as shown in Fig. 3(a), where RF0 has nine additional P contacts in the final 100 ms compared to other conditions The prespecified secondary R-minus-P net-loss contrast was −22. p. 29 pp for RF-0 (Fig. p. 0 3(b)) However, A/B yielded smaller contrasts of −4. p. 32 pp, establishing neither equivalence nor an isolated RF effect Post-hoc fault-minus-healthy trigger-time differences averaged 80. p. 30/57.5/57.5 ms in RF-0/A/B across bearing pairs
Audit and Consistency Checks
The audit gain quantifies the impact of checking mechanisms, where for 87 corruptions, A/B/C/D detected 36/54/75/81 cases respectively The author-designed catalogue detects 81/87 corruptions versus 36/87 for trusted headers, accepting all 20 distinct valid inputs (six misses) The frozen command-rate warning detects only 2/13 baseline collisions across all tested deletion levels
Conclusion and Limitations
Evidence concerns one graph, an assumed RF family, a synthetic plant, unequal healthy reserve and sparse failures Functional row loss establishes no biological tolerance or hardware safety benefit The study quantifies conditional task preservation and record-consistency gains for evaluating how future device-level nonidealities propagate through neuromorphic controllers to closed-loop behavior The results show that walk-only contact by 0.8 s reflects plant speeds, indicating finite-horizon preservation rather than sustained avoidance The study concludes that trace consistency cannot authenticate neural execution
References
[1] P. K. Shiu, G. R. Sterne, N. Spiller, R. p. 4 Franconville, A. p. 4 M. Sandoval, J. p. 4, “A Drosophila computational brain model reveals sensorimotor processing,” Nature, vol. p. 4, 634, no., 8032, pp., 210–219, 2024
[5] S. Jha, S. S. Banerjee, T. p. 4, et al., “ML-based fault injection for autonomous vehicles: A case for bayesian fault injection,” in 2019 49th Annual IEEE/IFIP International Conference on Dependable Systems and Networks (DSN), 2019, pp. p. 4, 112–124
[7] A. B. Go¨gebakan, E. p. 4, et al., “SpikingJET: Enhancing fault injection for fully and convolutional spiking neural networks,” in 2024 IEEE 30th International Symposium on On-Line Testing and Robust System Design (IOLTS). p. 4 IEEE, 2024, pp. p. 4, 1–7
[8] A. Colucci, A. p. 4, et al., “enpheeph: A fault injection framework for spiking and compressed deep neural networks,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). p. 4 IEEE, Oct., 2022, pp. p. 4, 5155–5162
[9] P. Glasserman and D. p. 4 D. Yao, “Some guidelines and guarantees for common random numbers,” Management Science, vol. p. 4, 38, no., 6, pp., 884–908, 1992
[10] A. Donze, “Breach, a toolbox for verification and parameter synthesis ´ of hybrid systems,” in Computer Aided Verification (CAV). p. 4 Springer, 2010, pp. p. 4, 167–170
[13] G. Chessa, Y. Yang, E., et al., “Cross-layer analysis of permanent faults in asynchronous NoCs for neuromorphic computing,” in Proceedings of the International Conference on Neuromorphic Systems. p. 4 ACM, 2026, pp. p. 4, 31–38
[14] P. Schlegel, Y. p. 4, et al., “Whole-brain annotation and multi-connectome cell typing of Drosophila,” Nature, vol. p. 4, 634, no., 8032, pp., 139–152, 2024
[15] M. Stimberg, R. p. 4 Brette, and D. p. 4 F., “Brian 2, an intuitive and efficient neural simulator,” eLife, vol. p. 4, 8, p., e47314, Aug. p. 4, 2019
[16] S.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Close-loop fault assessment of connection loss can be implemented by testing a connectome-derived controller in synthetic pursuit–escape scenarios to check record consistency across paired campaigns. This allows systems to quantify
finite-horizon collisions and the gains and limits of record checks.
-
The system can utilize paired, stage-resolved task measurements of law-discordant contacts under shared schedules, horizon sensitivity, uncertain RF transfer and frozen-prefix warning failures. This provides a method to measure
conditional task preservation and record-consistency audit gains.
-
Implement an author-designed fly/delayed-tank catalogue to detect corruptions by comparing checks plus model-specific replay against trusted headers. This system can distinguish between different fault types, detecting
81/87 corruptions versus 36/87 for trusted headers.
Sources
- Simulating Alzheimer's amyloid-like pathology on a fly central nervous system connectome
- SpikeFI: A Fault Injection Framework for Spiking Neural Networks
Related papers
- Self-Supervised Learning of Local Predictive Directions as a Neural Computation
- Synchronized Hearts, Desynchronized Brains: A Nonlinear Dynamics Perspective for Function and Failure
- Message interaction dynamics covary with brain volume across mammalian connectomes
- Planning as Dynamics Relaxation: Hippocampal Recurrent Network Realizes Optimal Goal-Directed Navigation
- The Platonic brain bridge hypothesis: brain networks as architectural priors for multimodal large language models
- BrainPFN: Amortized Brain-Behaviour Prediction via Prior-Data Fitted Networks