REVIEW 3 major objections 5 minor 28 references
Neuromorphic Computing for Embodied Intelligence in Autonomous Systems: Current Trends, Challenges, and Future Directions
T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read A unified workflow for secure, efficient, reliable spiking networks is proposed as the foundation for embodied intelligence in autonomous systems.
desk verdict A competent but overclaimed SNN survey whose five-stage workflow is never composed or validated, and whose manuscript carries a block of unrelated IEEE content. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the five-stage cross-layer workflow: data collection and pre-processing, secure SNN training, reliable SNN design, hardware- and software-level optimizations, and real-time adaptation. Each stage contributes a named mechanism—spike trains from event discretization, spike-aware adversarial training, BFV encryption, voltage-threshold and time-window tuning, pruning and quantization, hardware-aware neural architecture search, noise filtering, and continual learning—and the workflow's claim is that these mechanisms compose into a single deployable system. Without this workflow, the paper would be a list of isolated results; with it, the results become a proposed design path for embodied intelligence.
What would settle it
Build one SNN-based autonomous perception system that runs all five workflow stages—adversarial training, BFV-encrypted inference, pruning and quantization, continual learning, and fault-injection testing—on a single event-based dataset such as DVS-Gesture or NCARS, and compare its accuracy, energy, and latency against each stage run alone; if the stages conflict, say 4-bit quantization breaks encrypted inference or continual learning erases adversarial defenses, the workflow's central promise fails.
Extended reading notes
Core claim
The paper's central claim is that the building blocks for embodied intelligence in autonomous systems already exist and can be composed into one workflow. In its own terms, the claim is that "these stages result in a secure, efficient, and reliable SNN system that supports embodied intelligence." The argument is that each stage has a demonstrated mechanism: event discretization and encoding turns sensor streams into spike trains; adversarial training with spike-aware regularization and BFV homomorphic encryption make training and inference secure; voltage-threshold and time-window tuning plus inherent SNN fault tolerance provide reliability; pruning, quantization, and hardware-aware neural architecture search provide efficiency; noise filters and continual learning provide adaptability. The cited evidence includes a spiking AlexNet matching a conventional AlexNet under BFV encryption, SNNs retaining accuracy better than the compared DNNs under fault injection, and an event-based car SNN that keeps high accuracy at 10-bit precision.
Load-bearing premise
The load-bearing premise is that the five workflow stages described in Section III and Figure 2—adversarial training, BFV encryption, fault-tolerant design, pruning and quantization, and continual learning—can be combined in a single spiking network without their constraints conflicting, even though the paper presents each technique in isolation and gives no end-to-end demonstration.
Editorial extensions
If this is right
- Engineers can use the workflow as a checklist for building an SNN-based autonomous agent, choosing one technique per stage instead of assembling methods ad hoc.
- Event-based perception in vehicles and drones could run at lower energy and latency than frame-based deep networks, because each optimization stage is aimed at reducing spike traffic and memory footprint.
- Secure deployment becomes a concrete target: the cited results suggest an SNN can be adversarially trained, encrypted, and still reach accuracy comparable to an unencrypted conventional network.
- Low-precision SNNs should be able to achieve large memory savings without major accuracy loss, enabling deployment on neuromorphic chips with small on-chip memory.
- Continual learning at low precision gives autonomous agents a path to updating behavior online without catastrophic forgetting, which is what embodied intelligence requires.
Reading between the lines
- A direct test is an end-to-end benchmark that applies all five stages to one event-based dataset and reports combined accuracy, energy, and latency; the paper itself does not run this test.
- If encrypted, low-precision SNN inference holds up, privacy-preserving edge AI becomes a natural target, since homomorphic encryption and aggressive quantization both stress numeric ranges and the cited SNN results suggest spiking models tolerate both.
- The workflow treats security and adaptability as separate stages, leaving open the harder question of maintaining adversarial robustness while the network continually learns from new, possibly poisoned events.
- Fault-injection comparisons could be run on the whole workflow rather than a single model, revealing which combination of encryption and compression is most fault-sensitive.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This manuscript is a short survey/position paper on neuromorphic computing for autonomous systems. It proposes a five-stage 'cross-layer workflow' — data collection and event encoding; secure SNN training via adversarial training and BFV homomorphic encryption; reliable SNN design via hyperparameter tuning and fault tolerance; hardware/software optimization via pruning, quantization, and hardware-aware NAS; and real-time adaptation via noise filtering and continual learning — and illustrates each stage with plots adapted from the authors' prior publications. The paper closes with a list of challenges and open questions. The central claim is that composing these stages yields a 'secure, efficient, and reliable SNN system' that supports embodied intelligence.
Significance. If the workflow were fully demonstrated, it would provide a valuable organizer for a fragmented literature and would be appropriate for a robustness-oriented venue like IOLTS. The paper has useful strengths: it is concise, concretely links event-based sensing (DVS), SNN algorithms, and neuromorphic hardware (TrueNorth, Loihi, etc.), and each stage is illustrated with a quantitative figure from a peer-reviewed source. However, the paper's main value as a 'unified workflow' is currently not supported: the stages are validated only in isolation on different datasets and models, mostly from the authors' own group, and the composition is asserted rather than tested. The general claims about SNN fault tolerance and encrypted inference also go beyond what the displayed data establish.
major comments (3)
- [Sec. III, Fig. 2] The central claim of the paper is the workflow in Figure 2, and the text states 'These stages result in a secure, efficient, and reliable SNN system.' This composition is not demonstrated anywhere in the manuscript. In particular, stage 2 (BFV homomorphic encryption, Sec. III.B) and stage 4 (pruning/quantization, Sec. III.D) and stage 5 (continual learning, Sec. III.E) are presented as independent techniques evaluated on different datasets (FashionMNIST, NCARS, MNIST, DVS-Gesture, CIFAR-10) and different model architectures. BFV is a leveled homomorphic scheme: the set of operations must be fixed at encryption time and the noise budget grows with multiplications, so pruning, quantization, and weight updates performed after encryption would require decryption or re-encryption, which defeats the stated privacy goal. The manuscript provides no end-to-end experiment, no formal analysis of stage interfaces, and no discussion of whether the plaintext modulus or noise budget is compatible across the stages. Without such support, the 'unified workflow' is a diagram and not a demonstrated design path.
- [Sec. III.C, Fig. 5] The claim that 'SNNs outperform ResNet18 and VGG11 in maintaining accuracy under both output and weight faults' is not supported by the displayed comparison. Figure 5 compares fault injection results for ResNet18 and VGG11 on CIFAR-10 with results for a 4-layer SNN on DVS-Gesture; the models, datasets, and possibly task difficulties differ across panels. This is not a controlled comparison, and no error bars or repeated-run statistics are provided, so the difference could reflect dataset or model architecture rather than a general superiority of SNNs. Because the 'reliability' pillar of the proposed workflow rests on this fault-tolerance claim, the claim should be either restricted to the specific setting or supported by a matched comparison on the same dataset with identical fault-injection protocol.
- [Sec. III.B, Fig. 3] The encryption-related claim that 'Spiking-AlexNet maintains competitive accuracy under encryption' and 'shows improved resilience in low-modulus regimes' is presented without numerical values in the text and without error bars in Figure 3. Since the figure is a single adapted curve from reference [17], and since AlexNet is a non-spiking baseline rather than a spiking model with matched capacity and training protocol, the reader cannot assess whether the difference is significant. Please provide the test accuracies, the number of runs, and the exact hyperparameters used for both models.
minor comments (5)
- [Sec. III.A] Section III.A contains a sentence fragment: '...followed by a pre-processing stage that includes event discretization and encoding. Yielding temporally structured event sequences suitable for spiking neural processing.' The second clause should be joined to the first sentence.
- [Fig. 2 caption] The figure caption of Figure 2 describes step 2 as 'secure SNN training' followed by 'reliable SNN design' in step 3, while the body text of Section III presents reliable design before secure training; the inconsistency should be resolved.
- [References] The manuscript's bibliography appears to include two different reference lists; for example, the Introduction's [1] and [2] are deep-learning surveys, but the final reference list's [1] and [2] are CarSNN and LaneSNNs papers. The in-text citations should be matched to a single consistent bibliography.
- [Fig. 3 and Fig. 5] Figure 3 and Figure 5 would benefit from error bars or confidence bands, and from a statement of the number of independent runs, since the claims are comparative.
- [Fig. 5 caption] In Figure 5's caption and text, the identity of 'our SNN model' should be specified (e.g., the 4-layer SNN from enpheeph) so that the reader can locate the architecture details.
Circularity Check
No circularity: the paper is a survey whose workflow diagram organizes independently cited component results, and none of the workflow stages is derived from or equivalent to the paper's own inputs.
full rationale
The paper does not claim a quantitative derivation or a first-principles prediction. Its central contribution is the organizing statement: "This paper presents a unified workflow that integrates recent advances across neuromorphic algorithms, hardware platforms, and event-based sensors, with cross-layer optimization techniques aimed at improving energy efficiency, security, reliability, and adaptability." That workflow (Figure 2 and Section III) is a survey-level assembly of separately published results, not a derivation in which one stage is defined in terms of another. Each cited component, such as enpheeph fault injection [19], SNN4Agents [20], FastSpiker [21], and lpSpikeCon [24], is backed by its own experiments on MNIST, NCARS, DVS-Gesture, or CIFAR-10, so the self-citations are real evidence rather than circular justification. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work to force a choice, and no ansatz is smuggled in via citation; the figures adapted from earlier papers are empirical comparisons, not definitional identities. The assertion that "These stages result in a secure, efficient, and reliable SNN system" is not demonstrated end-to-end and the composability of BFV encryption with pruning, quantization, and continual learning is not established, but this is an unvalidated compositional claim or a completeness gap, not a circular reduction. Because the paper contains no derivation chain whose output is equivalent to its input by construction, the circularity score is 0.
Assumptions & free parameters
assumptions (4)
- domain assumption SNNs process information via discrete events, enabling low-power, low-latency inference.
- ad hoc to paper The five workflow stages can be composed into one pipeline without adverse interactions.
- domain assumption SNNs are inherently more fault-tolerant than DNNs because of sparse and distributed activation.
- domain assumption The adapted figures from cited works faithfully represent those works' results.
Cite this review
Pith. "Pith review of Neuromorphic Computing for Embodied Intelligence in Autonomous Systems: Current Trends, Challenges, and Future Directions." pith.science (2026). https://pith.science/paper/HTUTTHN2
@misc{pith2026250718139,
author = {Pith},
title = {Pith review of: Neuromorphic Computing for Embodied Intelligence in Autonomous Systems: Current Trends, Challenges, and Future Directions},
year = {2026},
howpublished = {\url{https://pith.science/paper/HTUTTHN2}},
note = {Machine review of arXiv:2507.18139}
}
read the original abstract
The growing need for intelligent, adaptive, and energy-efficient autonomous systems across fields such as robotics, mobile agents (e.g., UAVs), and self-driving vehicles is driving interest in neuromorphic computing. By drawing inspiration from biological neural systems, neuromorphic approaches offer promising pathways to enhance the perception, decision-making, and responsiveness of autonomous platforms. This paper surveys recent progress in neuromorphic algorithms, specialized hardware, and cross-layer optimization strategies, with a focus on their deployment in real-world autonomous scenarios. Special attention is given to event-based dynamic vision sensors and their role in enabling fast, efficient perception. The discussion highlights new methods that improve energy efficiency, robustness, adaptability, and reliability through the integration of spiking neural networks into autonomous system architectures. We integrate perspectives from machine learning, robotics, neuroscience, and neuromorphic engineering to offer a comprehensive view of the state of the field. Finally, emerging trends and open challenges are explored, particularly in the areas of real-time decision-making, continual learning, and the development of secure, resilient autonomous systems.
Figures
Figures from the paper (4 more)
Reference graph
Works this paper leans on
-
[17]
A homomorphic encryption framework for privacy- preserving spiking neural networks,
F. Nikfam et al. , “A homomorphic encryption framework for privacy- preserving spiking neural networks,” Information, 2023
work page 2023
-
[1]
A. Viale et al. , “Carsnn: An efficient spiking neural network for event- based autonomous cars on the loihi neuromorphic research processor,” in IJCNN, 2021
work page 2021
-
[2]
Lanesnns: Spiking neural networks for lane detection on the loihi neuromorphic processor,
A. Viale et al. , “Lanesnns: Spiking neural networks for lane detection on the loihi neuromorphic processor,” in IROS, 2022
work page 2022
-
[3]
Embedded neuromorphic using intel’s loihi processor,
A. Marchisio and M. Shafique, “Embedded neuromorphic using intel’s loihi processor,” in Embedded Machine Learning for Cyber-Physical, IoT, and Edge Computing: Software Optimizations and Hardware/Software Codesign, Springer, 2023
work page 2023
-
[4]
R. V . W. Putra et al. , “Embodied neuromorphic artificial intelligence for robotics: Perspectives, challenges, and research development stack,” in ICARCV, 2024
work page 2024
-
[5]
Networks of spiking neurons: the third generation of neural network models,
W. Maass, “Networks of spiking neurons: the third generation of neural network models,” Neural networks , 1997
work page 1997
-
[6]
Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip,
F. Akopyan et al., “Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip,” IEEE TCAD , 2015
work page 2015
-
[7]
Efficient neuromorphic signal processing with loihi 2,
G. Orchard et al., “Efficient neuromorphic signal processing with loihi 2,” in SiPS, 2021
work page 2021
Show all 28 references
-
[8]
The brainscales-2 accelerated neuromorphic system with hybrid plasticity,
C. Pehle et al. , “The brainscales-2 accelerated neuromorphic system with hybrid plasticity,” Frontiers in Neuroscience , 2022
2022
-
[9]
Dynap-se2: a scalable multi-core dynamic neuromorphic asynchronous spiking neural network processor,
O. Richter et al., “Dynap-se2: a scalable multi-core dynamic neuromorphic asynchronous spiking neural network processor,” Neuromorphic Computing and Engineering , 2024
2024
-
[10]
A 128x128 120 db 15us latency asynchronous temporal contrast vision sensor,
P. Lichtsteiner, C. Posch, and T. Delbruck, “A 128x128 120 db 15us latency asynchronous temporal contrast vision sensor,” IEEE JSSC , 2008
2008
-
[11]
A low power, fully event-based gesture recognition system,
A. Amir et al. , “A low power, fully event-based gesture recognition system,” in CVPR, 2017
2017
-
[12]
Hots: a hierarchy of event-based time-surfaces for pattern recognition,
X. Lagorce et al. , “Hots: a hierarchy of event-based time-surfaces for pattern recognition,” IEEE TPAMI, 2016
2016
-
[13]
Hats: Histograms of averaged time surfaces for robust event-based object classification,
A. Sironi et al. , “Hats: Histograms of averaged time surfaces for robust event-based object classification,” in CVPR, 2018
2018
-
[14]
Towards deep learning models resistant to adversarial attacks,
A. Madry et al. , “Towards deep learning models resistant to adversarial attacks,” in ICLR, 2018
2018
-
[15]
Snn-rat: Robustness-enhanced spiking neural network through regularized adversarial training,
J. Ding et al. , “Snn-rat: Robustness-enhanced spiking neural network through regularized adversarial training,” NeurIPS, 2022
2022
-
[16]
Somewhat practical fully homomorphic encryption,
J. Fan and F. Vercauteren, “Somewhat practical fully homomorphic encryption,” Cryptology ePrint Archive , 2012
2012
-
[18]
Securing deep spiking neural networks against adversarial attacks through inherent structural parameters,
R. El-Allami et al. , “Securing deep spiking neural networks against adversarial attacks through inherent structural parameters,” in DATE, 2021
2021
-
[19]
enpheeph: A fault injection framework for spiking and compressed deep neural networks,
A. Colucci, A. Steininger, and M. Shafique, “enpheeph: A fault injection framework for spiking and compressed deep neural networks,” in IROS, 2022
2022
-
[20]
Snn4agents: a framework for developing energy-efficient embodied spiking neural networks for autonomous agents,
R. V . W. Putra, A. Marchisio, and M. Shafique, “Snn4agents: a framework for developing energy-efficient embodied spiking neural networks for autonomous agents,” Frontiers in Robotics and AI , 2024
2024
-
[21]
Fastspiker: Enabling fast training for spiking neural networks on event-based data through learning rate enhancements for autonomous embedded systems,
I. Bano et al. , “Fastspiker: Enabling fast training for spiking neural networks on event-based data through learning rate enhancements for autonomous embedded systems,” in ICARCV, 2024
2024
-
[22]
A methodology to study the impact of spiking neural network parameters considering event-based automotive data,
I. Bano et al. , “A methodology to study the impact of spiking neural network parameters considering event-based automotive data,” inICARCV, 2024
2024
-
[23]
R-snn: An analysis and design methodology for robustifying spiking neural networks against adversarial attacks through noise filters for dynamic vision sensors,
A. Marchisio et al. , “R-snn: An analysis and design methodology for robustifying spiking neural networks against adversarial attacks through noise filters for dynamic vision sensors,” in IROS, 2021
2021
-
[24]
lpspikecon: Enabling low-precision spiking neural network processing for efficient unsupervised continual learning on autonomous agents,
R. V . W. Putra and M. Shafique, “lpspikecon: Enabling low-precision spiking neural network processing for efficient unsupervised continual learning on autonomous agents,” in IJCNN, 2022
2022
-
[25]
Implementing spiking neural networks on neuromorphic architectures: A review,
P. K. Huynh et al. , “Implementing spiking neural networks on neuromorphic architectures: A review,” arXiv preprint arXiv:2202.08897 , 2022
2022 arXiv
-
[26]
Data poisoning-based backdoor attack framework against supervised learning rules of spiking neural networks,
L. Jin, M. Lin, W. Jiang, and J. Zhan, “Data poisoning-based backdoor attack framework against supervised learning rules of spiking neural networks,” arXiv preprint arXiv:2409.15670 , 2024
2024 arXiv
-
[27]
The road to commercial success for neuromorphic technologies,
D. R. Muir and S. Sheik, “The road to commercial success for neuromorphic technologies,” Nature Communications, 2025
2025
-
[28]
Neuromorphic computing at scale,
D. Kudithipudi et al. , “Neuromorphic computing at scale,” Nature, 2025. 4
2025
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.