Pith. sign in

REVIEW 1 major objections 2 minor 32 references

CLANE: Continual Learning of Actions on Neuromorphic Hardware from Event Cameras

T0 review · 1 major / 2 minor · reviewed 2026-06-29 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read CLANE runs continual action recognition end-to-end on neuromorphic hardware from event cameras, reaching 70.4 percent accuracy with over 100 times lower energy use than a GPU baseline.

desk verdict CLANE puts a full continual-learning pipeline for event-camera action recognition onto Loihi 2 and reports solid efficiency numbers, but the GPU baseline comparison lacks enough implementation detail to judge the claimed gains. read the letter →

arxiv 2605.28387 v1 pith:A6IYCHLH submitted 2026-05-27 cs.LG cs.AIcs.NE

Elvin Hajizada , Michael Neumeier , Edward Paxon Frady , Yulia Sandamirskaya , Axel von Arnim , Bing Li , Eyke Hüllermeier This is my paper · ORCID
classification cs.LGcs.AIcs.NE
keywords continuallearningneuromorphichardwareeventcamerasactionrecognitionon-devicespikingneuralnetworksLoihi2energyefficiency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces CLANE as an end-to-end pipeline that learns new human actions continuously on neuromorphic chips without forgetting earlier ones. It processes sparse data from event cameras using a spiking CNN for features and a specialized learning head on Loihi 2, with added layers for handling action sequences and normalization. This setup targets privacy and low-latency needs in AR/VR and robotics by keeping all sensing, processing, and adaptation on the device. A reader would care because it shows how event-based sensing and neuromorphic hardware can support real-world continual learning tasks at far lower power than conventional edge GPUs.

What carries the argument

CLP-SNN learning head extended by Temporal Aggregation Layer and fixed-point Normalization Layer on Loihi 2, which carries the continual on-chip adaptation from event-camera streams.

What would settle it

Running the identical CNN+GRU+CLP algorithm on the same edge GPU hardware and measuring its energy and latency under the exact same input streams and continual learning schedule as CLANE would confirm or refute the 100x energy and 16x latency claims.

Watch

Extended reading notes

Core claim

CLANE is the first end-to-end continual on-device learning pipeline for event-based action recognition using neuromorphic hardware. It combines a spiking 2D CNN for spatiotemporal feature extraction with CLP-SNN as its on-chip learning head, extended to action clips via a Temporal Aggregation Layer and a fixed-point Normalization Layer, both novel Loihi 2 modules. On the THU E-ACT-50 dataset of 50 classes captured under real-world conditions, the system achieves 70.4 percent accuracy in a continual learning task while delivering more than 100 times energy reduction and 16 times lower latency over a sequential CNN+GRU+CLP edge GPU baseline, validated through iso-algorithm cross-platform bench

Load-bearing premise

The iso-algorithm cross-platform benchmarking equates the neuromorphic implementation to the GPU baseline without unstated differences in optimization, data handling, or hardware-specific tuning that could change the reported energy and latency gains.

Editorial extensions

If this is right

  • Event cameras paired with neuromorphic chips can support privacy-preserving on-device adaptation for AR/VR and robotics without sending data to the cloud.
  • Continual learning of actions becomes feasible at the edge with dramatically lower power draw than sequential GPU pipelines.
  • The approach scales to 50-class real-world datasets while maintaining accuracy close to offline methods.
  • Fixed-point normalization and temporal aggregation modules can be added to other spiking networks on Loihi 2 to handle sequential data.
  • The pipeline demonstrates that spiking feature extractors and on-chip learners can be combined for end-to-end action recognition.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Similar on-chip continual learning heads could be applied to other sparse sensory streams such as audio or tactile data.
  • The energy savings open the possibility of deploying adaptive action recognition on battery-powered wearable devices.
  • Extending the temporal aggregation layer to longer sequences might improve performance on more complex multi-step activities.
  • The architecture suggests a path for merging event-based vision with other neuromorphic sensors in a single chip.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 2 minor

Summary. The manuscript presents CLANE as the first end-to-end continual on-device learning pipeline for event-based action recognition on neuromorphic hardware. It deploys a spiking 2D CNN for spatiotemporal feature extraction paired with a CLP-SNN learning head on Intel Loihi 2, extended via novel Temporal Aggregation Layer and fixed-point Normalization Layer modules. On the THU E-ACT-50 dataset, CLANE reports 70.4% accuracy under continual learning while claiming >100x energy reduction and 16x lower latency versus a sequential CNN+GRU+CLP edge-GPU baseline, with the efficiency gains validated through iso-algorithm cross-platform benchmarking at three evaluation levels.

Significance. If the results and benchmarking hold, the work would advance on-device continual learning for event-camera-based action recognition in privacy-sensitive, low-power AR/VR and robotics settings by demonstrating a complete neuromorphic deployment. The concrete hardware implementation and introduction of Loihi 2-specific modules for temporal aggregation and normalization constitute tangible engineering contributions that bridge spiking networks with continual learning.

major comments (1)
  1. [Abstract] Abstract: the claims of 'more than 100x energy reduction and 16x lower latency' rest on 'iso-algorithm cross-platform benchmarking across three evaluation levels,' yet the manuscript provides no description of the CNN+GRU+CLP edge GPU baseline implementation, optimization level, data batching, preprocessing pipeline, or hardware-specific tuning. Without these details it is impossible to confirm that the comparison equates equivalent algorithmic work, which directly undermines the central efficiency claims.
minor comments (2)
  1. [Abstract] Abstract: the reported 70.4% accuracy is given as a single scalar with no error bars, standard deviation, number of runs, or description of the continual learning protocol (class order, memory strategy, or rehearsal).
  2. [Abstract] Abstract: the phrase 'three evaluation levels' is used without defining what the levels comprise or how they relate to the iso-algorithm comparison.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the need for greater transparency in our efficiency benchmarking. We address the single major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claims of 'more than 100x energy reduction and 16x lower latency' rest on 'iso-algorithm cross-platform benchmarking across three evaluation levels,' yet the manuscript provides no description of the CNN+GRU+CLP edge GPU baseline implementation, optimization level, data batching, preprocessing pipeline, or hardware-specific tuning. Without these details it is impossible to confirm that the comparison equates equivalent algorithmic work, which directly undermines the central efficiency claims.

    Authors: We agree that the manuscript currently lacks the requested implementation details for the CNN+GRU+CLP edge-GPU baseline, which is necessary to fully substantiate the iso-algorithm claims. In the revised manuscript we will add a new subsection (tentatively 4.3) that explicitly describes: (i) the exact CNN and GRU architectures and layer dimensions used in the baseline, (ii) the optimization framework and precision settings (including any quantization or TensorRT-style tuning), (iii) batch sizes, data-loading pipeline, and event-to-frame preprocessing steps to ensure input equivalence with the spiking path, and (iv) the three evaluation levels and how algorithmic work is matched across platforms. These additions will be cross-referenced from the abstract and results sections. The core accuracy and energy numbers themselves remain unchanged. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: purely empirical system measurements

full rationale

The paper describes an end-to-end neuromorphic system implementation (spiking CNN + CLP-SNN + novel Loihi 2 layers) and reports direct hardware measurements of accuracy, energy, and latency on THU E-ACT-50. No equations, derivations, fitted parameters, or first-principles predictions are present that could reduce to their own inputs by construction. The iso-algorithm benchmarking claim is an empirical validation statement, not a mathematical reduction. Self-citations, if any, are not load-bearing for any claimed derivation. This is a standard self-contained empirical systems paper.

Assumptions & free parameters 0 free parameters · 0 assumptions · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; the work is an engineering implementation whose internal modeling choices remain unspecified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLANE: Continual Learning of Actions on Neuromorphic Hardware from Event Cameras." pith.science (2026). https://pith.science/paper/A6IYCHLH

@misc{pith2026260528387,
  author       = {Pith},
  title        = {Pith review of: CLANE: Continual Learning of Actions on Neuromorphic Hardware from Event Cameras},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A6IYCHLH}},
  note         = {Machine review of arXiv:2605.28387}
}
read the original abstract

Recognizing and continuously learning novel human actions without forgetting prior classes is a requirement for emerging AR/VR and robotics applications. For these applications, both on-device processing and learning are essential for privacy and low-latency adaptation. Event cameras address the efficiency of visual sensing with sparse, asynchronous output that is naturally compatible with neuromorphic processing. Yet no prior system has deployed a continual on-device learning pipeline for event-based action recognition using neuromorphic hardware. We present CLANE, Continual Learning of Actions on Neuromorphic Hardware from Event Cameras, deployed end-to-end on Intel Loihi 2. CLANE combines a spiking 2D CNN for spatiotemporal feature extraction with CLP-SNN as its on-chip learning head, extended to action clips via a Temporal Aggregation Layer and a fixed-point Normalization Layer, both novel Loihi 2 modules. On THU E-ACT-50, a 50-class dataset captured under real-world conditions, CLANE achieves 70.4% accuracy in a continual learning task while delivering more than 100x energy reduction and 16x lower latency over a sequential CNN+GRU+CLP edge GPU baseline, validated through iso-algorithm cross-platform benchmarking across three evaluation levels.

Figures

Figures reproduced from arXiv: 2605.28387 by the authors.

Figure 1
Figure 1. Our CLANE pipeline with Spiking CNN, Temporal Aggregation, Normaliza￾tion, and CLP-SNN including sparse binary or graded spike activity maps and vectors. Schematic visualization of pretraining in A and continual learning in B. work. This alignment extends naturally to learning that is event-driven, sparse, and local. Yet despite on-chip learning mechanisms being available on chips like Loihi 2, deploying this full s… view at source ↗
Figure 2
Figure 2. Continual learning accuracy on THUE-ACT-50 (12 hold-out classes). (a) Cumu￾lative accuracy after each class-incremental step (10-shot). (b) Final accuracy versus shots per class. CLP-Loihi runs on Loihi 2; all others on Jetson Orin Nano. of the 12 hold-out classes and demonstrate that naive fine-tuning catastroph￾ically forgets as new classes are learned, while CLP and other CL methods retain knowledge of previous c… view at source ↗
Figure 3
Figure 3. Comparison of different feature extraction and learning method combinations [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 5 canonical work pages

  1. [1]

    In: Proceedings of the IEEE CVPR

    Amir, A., et al.: A low power, fully event-based gesture recognition system. In: Proceedings of the IEEE CVPR. pp. 7243–7252 (2017)

    2017

  2. [2]

    de Blegiers, T., Dave, I.R., Yousaf, A., Shah, M.: EventTransAct: A Video Transformer-Based Framework for Event-Camera Based Action Recognition. IROS pp. 1–7 (2023) 12 E. Hajizada et al

    2023

  3. [3]

    Davies, N

    Davies,M.,etal.:Loihi:ANeuromorphicManycoreProcessorwithOn-ChipLearn- ing. IEEE Micro38, 82–99 (2018). https://doi.org/10.1109/MM.2018.112130359

    work page doi:10.1109/mm.2018.112130359 2018

  4. [4]

    IEEE TPAMI44(7) (2021)

    De Lange, M., Aljundi, R., Masana, M., Parisot, S., et al.: A continual learning survey: Defying forgetting in classification tasks. IEEE TPAMI44(7) (2021)

    2021

  5. [5]

    Fang, W., et al.: Incorporating Learnable Membrane Time Constant to Enhance Learning of Spiking Neural Networks. ICCV pp. 2641–2651 (2021)

    2021

  6. [6]

    Frontiers in neural circuits9, 85 (2016)

    Frémaux, N., Gerstner, W.: Neuromodulated spike-timing-dependent plasticity, and theory of three-factor learning rules. Frontiers in neural circuits9, 85 (2016)

    2016

  7. [7]

    Trends in cogni- tive sciences3(4), 128–135 (1999)

    French, R.M.: Catastrophic forgetting in connectionist networks. Trends in cogni- tive sciences3(4), 128–135 (1999)

    1999

  8. [8]

    IEEE TPAMI44(1), 154–180 (2020)

    Gallego, G., Delbrück, T., Orchard, G., Bartolozzi, C., et al.: Event-based vision: A survey. IEEE TPAMI44(1), 154–180 (2020)

    2020

Show all 32 references
  1. [9]

    IEEE TPAMI45(12), 14081–14097 (2023)

    Gao, Y., et al.: Action recognition and benchmark using event cameras. IEEE TPAMI45(12), 14081–14097 (2023)

    2023

  2. [10]

    In: IROS

    Hajizada, E., Swaminathan, B., Sandamirskaya, Y.: Continual Learning for Au- tonomous Robots: A Prototype-based Approach. In: IROS. IEEE (2024)

    2024

  3. [11]

    arXiv preprint arXiv:2511.01553 (2025)

    Hajizada, E., et al.: Real-time Continual Learning on Intel Loihi 2. arXiv preprint arXiv:2511.01553 (2025)

    2025 arXiv

  4. [12]

    arXiv preprint arXiv:2308.04749 (2023)

    Han, B., Zhao, F., Zeng, Y., Pan, W., Shen, G.: Enhancing efficient continual learning with dynamic structure development of spiking neural networks. arXiv preprint arXiv:2308.04749 (2023)

    2023

  5. [13]

    ECCV (2019)

    Hayes, T.L., Kafle, K., Shrestha, R., Acharya, M., Kanan, C.: REMIND Your Neural Network to Prevent Catastrophic Forgetting. ECCV (2019)

    2019

  6. [14]

    In: CVPR Workshop (June 2020)

    Hayes, T.L., Kanan, C.: Lifelong Machine Learning With Deep Streaming Linear Discriminant Analysis. In: CVPR Workshop (June 2020)

    2020

  7. [15]

    arXiv preprint arXiv:2203.10681 (2022)

    Hayes, T.L., Kanan, C.: Online continual learning for embedded devices. arXiv preprint arXiv:2203.10681 (2022)

    2022

  8. [16]

    Innocenti, S.U., Becattini, F., Pernici, F., Bimbo, A.: Temporal Binary Represen- tation for Event-Based Action Recognition. ICPR pp. 10426–10432 (2020)

    2020

  9. [17]

    Nature Reviews Neuroscience21(6), 335–346 (2020)

    Lillicrap, T.P., Santoro, A., Marris, L., Akerman, C.J., Hinton, G.: Backpropaga- tion and the brain. Nature Reviews Neuroscience21(6), 335–346 (2020)

    2020

  10. [18]

    IEEE transactions on pattern analysis and machine intelligence35(11), 2624–2637 (2013)

    Mensink, T., Verbeek, J., Perronnin, F., Csurka, G.: Distance-based image classifi- cation: Generalizing to new classes at near-zero cost. IEEE transactions on pattern analysis and machine intelligence35(11), 2624–2637 (2013)

    2013

  11. [19]

    IEEE Access (2025)

    Minhas, M.F., Putra, R.V.W., Awwad, F., Hasan, O., Shafique, M.: Continual learning with neuromorphic computing: Foundations, methods, and emerging ap- plications. IEEE Access (2025)

    2025

  12. [20]

    Minhas, M.F., et al.: Replay4NCL: An Efficient Memory Replay-based Methodol- ogyforNeuromorphicContinualLearninginEmbeddedAISystems.arXivpreprint arXiv:2503.17061 (2025)

    2025

  13. [21]

    FPGA Related (2020), https://www.fpgarelated.com/ showarticle/1347.php

    Morris, M.: Computing fixed-point square roots and their reciprocals using goldschmidt algorithm. FPGA Related (2020), https://www.fpgarelated.com/ showarticle/1347.php

    2020

  14. [22]

    IEEE Signal Processing Magazine36(6), 51–63 (2019)

    Neftci,E.O.,Mostafa,H.,Zenke,F.:SurrogateGradientLearninginSpikingNeural Networks: Bringing the Power of Gradient-Based Optimization to Spiking Neural Networks. IEEE Signal Processing Magazine36(6), 51–63 (2019)

    2019

  15. [23]

    In: ICONS

    Neumeier, M., et al.: EEvAct: Early Event-Based Action Recognition with High- Rate Two-Stream Spiking Neural Networks. In: ICONS. pp. 41–48 (2025)

    2025

  16. [24]

    In: 2021 IEEE Workshop on SiPS

    Orchard, G., et al.: Efficient neuromorphic signal processing with loihi 2. In: 2021 IEEE Workshop on SiPS. pp. 254–259. IEEE (2021) CLANE 13

    2021

  17. [25]

    IEEE JETCAS8(1) (2017)

    Panda, P., Allred, J.M., Ramanathan, S., Roy, K.: Asp: Learning to forget with adaptive synaptic plasticity in spiking neural networks. IEEE JETCAS8(1) (2017)

    2017

  18. [26]

    In: AICAS

    Pes, L., Luiken, R., Corradi, F., Frenkel, C.: Active dendrites enable efficient con- tinual learning in time-to-first-spike neural networks. In: AICAS. IEEE (2024)

    2024

  19. [27]

    In: 2023 IEEE 33rd International Workshop on MLSP

    Proietti,M.,Ragno,A.,Capobianco,R.:Memoryreplayforcontinuallearningwith spiking neural networks. In: 2023 IEEE 33rd International Workshop on MLSP. pp. 1–6. IEEE (2023)

    2023

  20. [28]

    Nature Computational Science2(1), 10–19 (2022)

    Schuman, C.D., et al.: Opportunities for neuromorphic computing algorithms and applications. Nature Computational Science2(1), 10–19 (2022)

    2022

  21. [29]

    In: 2024 IEEE ISMAR

    Shen, J., De Lange, M., Xu, X., et al.: Towards open-world gesture recognition. In: 2024 IEEE ISMAR. pp. 1236–1245. IEEE (2024)

    2024

  22. [30]

    IEEE JETCAS10(4), 512–521 (2020)

    Stewart, K., Orchard, G., Shrestha, S.B., Neftci, E.: Online few-shot gesture learn- ing on a neuromorphic processor. IEEE JETCAS10(4), 512–521 (2020)

    2020

  23. [31]

    Neurocomputing611, 128657 (2025)

    Vicente-Sola, A., et al.: Spiking neural networks for event-based action recognition: A new task to understand their advantage. Neurocomputing611, 128657 (2025)

    2025

  24. [32]

    Brain sciences12(7), 863 (2022)

    Yamazaki, K., Vo-Ho, V.K., Bulsara, D., Le, N.: Spiking neural networks and their applications: A review. Brain sciences12(7), 863 (2022)

    2022

Pith tools

Reviewed June 29, 2026 · model on record in the stance chip above.