Pith. sign in

REVIEW 5 major objections 4 minor 90 references

The Neural Division of Labor: Biologically-Inspired Modular Architectures for Robust Neuromorphic Computing

T0 review · 5 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read A spiking network whose class pathways are physically isolated experts can match dense accuracy at a fraction of the parameters and synaptic operations, and resists catastrophic forgetting by freezing old experts.

desk verdict The push-pull expert idea is plausible, but the central training scheme never touches the output layer, so the reported accuracies are not reproducible from the text. read the letter →

arxiv 2608.08317 v1 pith:5ERVFBED submitted 2026-08-08 cs.LG

classification cs.LG MSC 68T0792B20
keywords spikingneuralnetworksmodulararchitecturepush-pulllossone-vs-allexpertscatastrophicforgettingenergy-efficientneuromorphiccomputingANN-to-SNNconversioninterpretability
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 sets out to prove that a neural network does not need globally entangled weights to classify images. It reports a Decomposable Spiking Neural Network (D-SNN) in which each output class owns a structurally isolated 'expert' pathway, and a biomimetic push-pull loss forces activity into the correct expert while suppressing all others. The network is trained as a continuous ANN and then ported unchanged to a spiking network, with all reported accuracies measured in the spiking domain. The central claim is that this modular design matches fully dense baselines on MNIST, Fashion-MNIST, and CIFAR-10/100 while using roughly half the hidden parameters of the paper's own dense baselines and cutting firing rates and synaptic operations by one to three orders of magnitude. If true, it gives a practical route to auditable, energy-efficient neuromorphic classifiers without expensive spatio-temporal backpropagation.

What carries the argument

The load-bearing object is the block-diagonal expert decomposition used together with the Push-Pull loss of Eq. (7): $L_{PP} = \sum_m [\log(S^m_{\mathrm{false}} + S^m_{\mathrm{true}}) - \log(S^m_{\mathrm{true}})]$, where $S^m_{\mathrm{true}} = \sum_{\alpha \in B_{k_m}} z_\alpha$ and $S^m_{\mathrm{false}} = \sum_{\alpha \notin B_{k_m}} z_\alpha$ are sums of hidden post-activations inside and outside the correct expert block. This objective acts only on hidden-layer activity, not on output logits, so the training signal itself enforces sparse, class-specific routing. The second ingredient is structural: weight matrices between expert blocks are zeroed by construction, which eliminates cross-talk, lowers the fan-out from $O(M^2)$ to $O(M)$, and lets the network be verified as $K$ independent One-vs-All classifiers.

What would settle it

Run the training protocol exactly as stated, initialize the output-layer weights randomly, and verify that no secondary loss touches them; the SNN readout (output neuron with the most spikes) would then be driven by untrained weights and cannot reproduce the reported 98% MNIST accuracy. Equivalently, computing the gradient of $L_{PP}$ with respect to the output weight matrix shows it is identically zero under Eq. (7), so any reported accuracy requires an output-weight initialization or auxiliary objective that the paper does not disclose.

Watch

Extended reading notes

Core claim

The discovery the authors report is that routing a push-pull loss through a block-diagonal expert architecture produces a spiking network whose classification accuracy is statistically indistinguishable from a dense network, while the hidden activity is confined almost entirely to the expert of the correct class. The loss is $L_{PP} = \sum_m [\log(S^m_{\mathrm{false}} + S^m_{\mathrm{true}}) - \log(S^m_{\mathrm{true}})]$, built from hidden-layer activities: $S^m_{\mathrm{true}}$ sums the surrogate post-activations in the correct expert and $S^m_{\mathrm{false}}$ sums them everywhere else. Optimizing this quantity forces wrong experts to stay metabolically quiet and the correct expert to dominate, so the decision can be read off from which physical pathway fired. In SNN verification this yields 98.15% on MNIST and 91.68% on Fashion-MNIST (matching dense baselines), 82.7-83.3% on CIFAR-10 with only 5 neurons per expert, and a synaptic-operations reduction of roughly 37- to 160-fold on the two small datasets. The same frozen-expert mechanism keeps old-class accuracy at 80-90% in split-MNIST continual learning, where the dense cross-entropy baseline falls to 20-30%.

Load-bearing premise

The reported accuracies depend on the output-layer weights being set in a way the paper never specifies, because the push-pull loss in Eq. (7) is a function only of hidden-layer activations and therefore supplies no gradient to those weights.

Editorial extensions

If this is right

  • If the central claim holds, dense global coupling is unnecessary for these classification tasks: a per-class OvA expert layout reaches the same SNN-verified accuracy with a fraction of the synaptic operations.
  • Because training happens entirely in the continuous ANN and is ported unchanged to the SNN, the approach avoids BPTT and its memory overhead, cutting training time by roughly 10-50x on consumer hardware.
  • The structural $O(M)$ fan-out instead of $O(M^2)$ means the efficiency gain over dense networks should grow with the number of classes, as illustrated by the CIFAR-100 scaling tests up to $M=100$.
  • Freezing the weights of previously learned experts gives a direct, regularization-free defense against catastrophic forgetting in class-incremental learning, holding old-class accuracy near 80-90% in split MNIST.
  • Because only the correct expert fires during inference, every classification has an explicit physical origin, making the network auditable in a way dense networks are not.

Reading between the lines

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

  • A testable extension the paper does not report: use the total spike count in each expert as a per-class confidence score, since the activity matrices on CIFAR-100 are near-diagonal; this could give cheap rejection of ambiguous or out-of-distribution inputs.
  • The frozen-expert design hints at a physical neuromorphic architecture where experts occupy separate cores or chips and new classes are added by appending a unit, without retraining old ones; the paper motivates but does not build this hardware.
  • Because the SynOps advantage over dense networks is $O(M)$ rather than $O(M^2)$, the gain should widen on many-class problems beyond the 100-class scale tested here, such as large handwriting sets; a scaling measurement on such a dataset would test that prediction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper proposes a Decomposable Spiking Neural Network (D-SNN) in which the hidden layers are partitioned into class-specific experts with block-diagonal connectivity. Training is performed in a continuous ANN using a 'Push-Pull' loss L_PP (Eq. 7) that compares the total hidden-layer activity in the correct expert block with the total activity in all wrong blocks; the learned weights are then ported directly into a Leaky Integrate-and-Fire SNN, and classification is read out over T=100 simulation steps as the output neuron with the most spikes. The authors report results on MNIST, Fashion-MNIST, and CIFAR-10/100, claiming that this modular design matches dense baselines in accuracy while using far fewer parameters, producing much lower firing rates and synaptic operations, and providing protection against catastrophic forgetting when experts are frozen during continual learning.

Significance. If the claims were fully supported, the paper would offer a simple, transparent, and parameter-efficient recipe for modular spiking classifiers, and the three-way comparison among dense CCE, dense hybrid, and independent experts is a useful decomposition that separates the effect of the loss function from the effect of structural isolation. The manuscript also has positive reproducibility signals: standard public datasets, a stated code repository, SNN-based verification of all reported accuracies, and a scaling study up to 100 classes. However, the central claims rest on a training procedure that is incompletely specified, and several headline statements are contradicted by the paper's own tables.

major comments (5)
  1. [§4.6/§4.8, Eq. (7)] The training loss L_PP in Eq. (7) is a function only of the hidden-layer post-activations z_alpha defined in Eq. (6); the output layer never appears in L_PP, no secondary loss for output weights is defined, and no initialization or scaling rule for the output layer is stated. Since Section 4.8 determines classification from the output neuron with the maximum spike count, the described training procedure cannot be expected to produce the reported accuracies unless some omitted readout mechanism exists. This is load-bearing because every accuracy, efficiency, and forgetting result is measured through that output readout. The authors must specify exactly how the output weights are trained or initialized, or add an explicit output-loss term, before the experimental claims can be evaluated.
  2. [§2.3, Table 1, Table 4] The reported MNIST baseline is inconsistent across the paper: Section 2.3 states dense cross-entropy achieves 98.19%, Table 1 reports 98.55% for Dense CCE, and Table 4 reports 98.18% for Dense CCE. A single consistent set of baseline numbers is required, since the central claim of accuracy parity depends on these comparisons.
  3. [Abstract, §2.3, Tables 1–3] The abstract's claim of 'an order of magnitude fewer parameters' relative to fully dense networks is not supported by the paper's own tables. For MNIST, independent experts use about 272K parameters versus about 460K for Dense CCE (a factor of 1.7); for Fashion-MNIST the ratio is about 351K versus 539K (a factor of 1.5); and for CIFAR-10 the independent experts and dense hybrid have essentially identical parameter counts (≈350K versus ≈352K). The text in Section 2.3 correctly describes a 'two-fold reduction' in some places, so the abstract and the phrase 'an order of magnitude' should be corrected to match the actual data.
  4. [§2.8, §4.11] The continual-learning comparison is confounded by an asymmetric freezing protocol: in Phase 2, the independent experts and dense hybrid freeze all weights belonging to old-class experts, while the dense CCE baseline keeps all weights trainable. The paper's own control experiment, reported in Section 2.8, shows that when old experts are left plastic the independent experts' accuracy on legacy tasks drops to about 40%, indicating that structural isolation alone is not sufficient. The claim that physical separation provides 'inherent protection' against catastrophic forgetting should be either retracted or supported by a comparison in which the dense baseline receives an analogous class-dependent freezing or regularization scheme.
  5. [§2.5, §2.7, Table 4] The large SynOps reductions attributed to the method are largely forced by the block-diagonal architecture rather than by learning: for a dense backend the number of backend connections scales as O(M^2), while a block-diagonal expert mask reduces this to O(M) by construction. Similarly, L_PP in Eq. (7) explicitly suppresses activity in non-target experts, so a substantial part of the firing-rate reduction is a direct effect of the objective. The paper should separate architectural arithmetic from learned sparsity when presenting the efficiency results, and should not present the SynOps scaling in Section 2.7 as a discovered phenomenon.
minor comments (4)
  1. [§4.5 vs §4.10] Section 4.5 states that the shared receptor for CIFAR-10/100 comprises 'two' trainable 3x3 convolutional layers, whereas Section 4.10 describes a three-layer convolutional topology with two 5x5 layers and one 3x3 layer; this discrepancy should be resolved.
  2. [Table 4 caption vs §4.9] The caption of Table 4 says MNIST and Fashion-MNIST both used a base filter size of F=16, but Section 4.9 specifies MNIST with F=8 and Fashion-MNIST with F=16; please align the caption with the architecture descriptions.
  3. [Table 4, CIFAR-10 rows] The CIFAR-10 accuracy values in Table 4 (82.74, 83.27, 82.82) are consistent with the text, but Table 3 omits the Dense CCE row entirely; adding it would make the three-way comparison easier to verify.
  4. [References] Several references are incomplete or inconsistently formatted, for example [29] appears as a bare arXiv identifier without authors or title; the reference list should be checked for completeness.

Circularity Check

0 steps flagged · score 0.0 of 10

No load-bearing circularity: the accuracy result is an external empirical claim, and the efficiency/forgetting benefits are transparently engineered design consequences rather than predictions smuggled from inputs.

full rationale

The paper's central claim—that D-SNN matches dense accuracy with far fewer parameters—is an externally benchmarked empirical result, not a quantity forced by the method's own equations. L_PP in Eq. (7) is defined over hidden-layer activities, while accuracy is measured by output-neuron spike counts (Sec. 4.8), so the accuracy numbers are not equal to the loss by construction. The efficiency results are openly derived from the architecture: Sec. 2.7 gives an analytical counting argument for why block-diagonal connectivity yields O(M) versus O(M^2) backend scaling, and Table 4's SynOps metric is precisely a count over the nonzero block-diagonal paths. Reporting that a mask reduces operations is a design consequence, not a circular prediction. The low firing rates are likewise a direct effect of the explicitly stated push-pull objective, and the paper says the loss 'forces' sparse localized activation; this is transparent objective design, not a hidden fit. The catastrophic-forgetting protection is explicitly implemented by freezing old-expert weights (Sec. 4.11), and the paper even includes an ablation showing accuracy regresses to ~40% when old experts remain plastic. Thus the forgetting result is conditional on a stated protocol, not an inference that presupposes its conclusion. The self-citations (e.g., refs. 44, 67, 68, 82, all involving author Bazhenov) are used for biological analogy and sleep-replay motivation, not as load-bearing mathematical premises; the loss and architecture are defined in this paper. The only serious defect is the unspecified output-layer readout: Eq. (7) never trains the output weights, so the reported SNN accuracies are not fully reproducible from the described method. That is a correctness/reproducibility gap, not a circular derivation, because nothing in the stated equations forces the max-spike output to equal the training label. No step reduces, by definition or by self-citation, to its own input.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central quantitative claims rest on hand-chosen neuron, surrogate, and architectural parameters, and on an unspecified treatment of the output layer. The efficiency and forgetting results are consequences of the block-diagonal mask and the freezing protocol, so they add little independent evidence beyond the accuracy comparisons. No new physical entities are introduced.

free parameters (7)
  • Membrane decay beta = 0.5 in SNN verification (Sec. 4.8); 0.6 in continual learning (Sec. 4.11)
    Hand-chosen LIF time constant used for ANN-to-SNN porting; the paper uses different values in different experiments without explaining the change.
  • Surrogate steepness epsilon = 0.2
    Hand-chosen sharpness of the sigmoid surrogate in Eq. (5); affects gradient and spike transfer, with no sensitivity analysis.
  • Firing threshold V_th = 1.0
    Hand-chosen LIF threshold; combined with unscaled weights it determines whether SNN firing rates match ANN activations.
  • Synaptic efficacy kappa = 0.7
    Hand-chosen gain factor in Eq. (3)-(4); no calibration procedure is described.
  • SNN simulation horizon T = 100 timesteps
    Hand-chosen; the paper notes accuracy could be achieved with shorter windows, so results depend on this choice.
  • Per-expert hidden widths N1, N2 = MNIST/FMNIST 64/32; CIFAR-10 5/5; CIFAR-100 floor(500/M)
    Architecture sizes are chosen per dataset and directly control both accuracy and the reported efficiency ratios.
  • Rehearsal noise fraction = 0.15
    Hand-chosen for the synthetic memory templates in continual learning; no ablation of this parameter is provided.
assumptions (4)
  • domain assumption Surrogate ANN activations transfer to LIF firing rates without calibration.
    Methods 4.5-4.8 train a sigmoid ANN and port weights directly to an SNN with chosen beta, kappa, and V_th; no calibration or error correction for the mapping is described, yet all reported accuracies are SNN-computed.
  • ad hoc to paper Output-layer weights are trained or initialized by some mechanism not present in Eq. (7).
    The push-pull loss depends only on hidden-layer activity z_alpha, while classification is read from output spike counts in Section 4.8; the paper does not state how the output weights obtain gradients.
  • domain assumption Compressed rehearsal templates preserve old-class information.
    Continual learning uses averaged images plus 15% Gaussian noise as the only memory of old classes in Section 4.11; this assumes those templates are sufficient to keep frozen experts accurate.
  • domain assumption Class structure is known before training.
    Independent experts assign one fixed pathway per class; the bookkeeping of experts and frozen weights presumes a pre-specified class list.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Neural Division of Labor: Biologically-Inspired Modular Architectures for Robust Neuromorphic Computing." pith.science (2026). https://pith.science/paper/5ERVFBED

@misc{pith2026260808317,
  author       = {Pith},
  title        = {Pith review of: The Neural Division of Labor: Biologically-Inspired Modular Architectures for Robust Neuromorphic Computing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5ERVFBED}},
  note         = {Machine review of arXiv:2608.08317}
}
read the original abstract

Biological neural systems achieve high efficiency and robustness through compartmentalized architectures. In contrast, modern artificial neural networks rely on globally entangled structures, which obscure decision logic and suffer from catastrophic forgetting. Here, we report a Decomposable Spiking Neural Network (D-SNN) that eliminates global synaptic entanglement by structurally isolating classification pathways into independent experts. Optimized via a bio-inspired push-pull loss function, the D-SNN achieves competitive accuracies on MNIST, Fashion-MNIST, and CIFAR-10/100 benchmarks. This modular approach matches the performance of fully dense networks while utilizing an order of magnitude fewer parameters. In addition, our networks operate with up to several orders of magnitude lower firing rates and fewer synaptic operations. Furthermore, physically severing connections between experts provides inherent protection against catastrophic forgetting during sequential learning. Crucially, these isolated pathways generate auditable neural signals, increasing decision transparency. This biomimetic, verifiable architecture establishes an efficient foundation for deploying deterministic neuromorphic intelligence in resource-constrained edge environments.

Figures

Figures reproduced from arXiv: 2608.08317 by the authors.

Figure 1
Figure 1. Decomposable sensory architecture enabling parallel class verification in the D-SNN. In analogy of decentralized and specialized nervous systems found in insects, our Decomposable Spiking Neural Network (D-SNN) is designed with explicit structural modularity. Left panel: A raw visual input signal, illustrated by an MNIST digit ’3’. Center panel: The signal is first processed by a tiered layers of convolutional layer… view at source ↗
Figure 2
Figure 2. Absence of structural modularity in SNN optimizated by standard global [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Emergence of functional and structural modularity driven by biomimetic opti [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Comparative visualization of spiking activity localized within class-specific ex [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: To quantify this, we plot the efficiency multipliers, calculated as the ratio of the [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 5
Figure 5. Figure 5: Left panel: Plot of classification accuracy as a function of the number of target [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Orthogonalization of spiking activity and signal fidelity at maximum classifi [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Disparity between knowledge retention and overall SNN accuracy. Top panels: Confusion matrices for continual learning (M = 4). Dense CCE (left) shows drastic accuracy loss and over-attribution to newly learned classes (5–9), whereas In￾dependent Experts (right) remain …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

90 extracted references · 68 canonical work pages

  1. [1]

    Imagenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. InAdvances in Neural Information Processing Systems, volume 25, 2012

  2. [2]

    Deep learning.Nature, 521(7553):436–444, 2015

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning.Nature, 521(7553):436–444, 2015

  3. [3]

    High-performance medicine: the convergence of human and artificial intelligence.Nature Medicine, 25(1):44–56, 2019

    Eric J Topol. High-performance medicine: the convergence of human and artificial intelligence.Nature Medicine, 25(1):44–56, 2019

  4. [4]

    Military applications of artificial intel- ligence: ethical concerns in an uncertain world

    Forrest E Morgan, Benjamin Boudreaux, Andrew J Lohn, Mark Ashby, Christian Curriden, Kelly Klima, and Derek Grossman. Military applications of artificial intel- ligence: ethical concerns in an uncertain world. Technical report, RAND Corporation, 2020

  5. [5]

    An introduction to deep learning for the physical layer.IEEE Transactions on Cognitive Communications and Networking, 3(4):563– 575, 2017

    Timothy O’Shea and Jakob Hoydis. An introduction to deep learning for the physical layer.IEEE Transactions on Cognitive Communications and Networking, 3(4):563– 575, 2017

  6. [6]

    Language mod- els are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, et al. Language mod- els are few-shot learners. InAdvances in Neural Information Processing Systems, volume 33, pages 1877–1901, 2020

  7. [7]

    GPT-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    OpenAI. GPT-4 technical report.arXiv preprint arXiv:2303.08774, 2023

  8. [8]

    Scaling laws for neural lan- guage models.arXiv preprint arXiv:2001.08361, 2020

    Jared Kaplan, Sam McCandlish, Tom Henighan, et al. Scaling laws for neural lan- guage models.arXiv preprint arXiv:2001.08361, 2020

Show all 90 references
  1. [9]

    AI and compute.https://openai.com/index/ai-and-compute/, 2018

    OpenAI. AI and compute.https://openai.com/index/ai-and-compute/, 2018. Accessed: 2024

  2. [10]

    P. Dhar. The carbon impact of artificial intelligence.Nature Machine Intelligence, 2:423, 2020. 32

  3. [11]

    Desislavov, F

    R. Desislavov, F. Mart ´ ınez-Plumed, and J. Hern´ andez-Orallo. Trends in AI inference energy consumption: Beyond the performance-vs-parameter laws of deep learning. Sustainable Computing: Informatics and Systems, 38:100857, 2023

  4. [12]

    Data centres and data transmis- sion networks.https://www.iea.org/energy-system/buildings/ data-centres-and-data-transmission-networks#overview, 2023

    International Energy Agency. Data centres and data transmis- sion networks.https://www.iea.org/energy-system/buildings/ data-centres-and-data-transmission-networks#overview, 2023. Accessed: 2024

  5. [13]

    Shehabi, S

    A. Shehabi, S. J. Smith, A. Hubbard, A. Newkirk, N. Lei, M. A. B. Siddik, B. Holecek, J. Koomey, E. Masanet, and D. Sartor. 2024 United States Data Center Energy Usage Report. Technical report, Energy Analysis & Environmental Impacts Division, 2024

  6. [14]

    Stop explaining black box machine learning models for high stakes de- cisions and use interpretable models instead.Nature machine intelligence, 1(5):206– 215, 2019

    Cynthia Rudin. Stop explaining black box machine learning models for high stakes de- cisions and use interpretable models instead.Nature machine intelligence, 1(5):206– 215, 2019

  7. [15]

    Concept bottleneck models

    Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. InInternational Conference on Machine Learning, pages 5338–5348. PMLR, 2020

  8. [16]

    Interactive concept bottleneck models

    Kushal Chauhan, Rishabh Tiwari, Jan Freyberg, Pradeep Shenoy, and Krishna- murthy Dvijotham. Interactive concept bottleneck models. InProceedings of the aaai conference on artificial intelligence, volume 37, pages 5948–5955, 2023

  9. [17]

    Towards monosemanticity: Decomposing language models with dictionary learning.Trans- former Circuits Thread, 2023

    Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, et al. Towards monosemanticity: Decomposing language models with dictionary learning.Trans- former Circuits Thread, 2023

  10. [18]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. InInternational conference on machine learning, pages 3319–3328. PMLR, 2017

  11. [19]

    Towards artificial general intelligence with hybrid Tianjic chip architecture.Nature, 572(7767):106–111, 2019

    Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenyu Chen, Wei He, et al. Towards artificial general intelligence with hybrid Tianjic chip architecture.Nature, 572(7767):106–111, 2019

  12. [20]

    TrueNorth: Accelerating from zero to 64 million neurons in 10 years.Computer, 52(5):20–29, 2019

    Michael V DeBole, Brian Taba, Arnon Amir, Filipp Akopyan, Alexander Andreopou- los, William P Risk, Jeff Kusnitz, Carlos Ortega Otero, Tapan K Nayak, Rathinaku- mar Appuswamy, et al. TrueNorth: Accelerating from zero to 64 million neurons in 10 years.Computer, 52(5):20–29, 2019

  13. [21]

    Loihi: a neuromorphic manycore processor with on-chip learning.IEEE Micro, 38(1):82–99, 2018

    Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Buffani, Bruno Guilioni, Manyun Liao, Jing Wang, et al. Loihi: a neuromorphic manycore processor with on-chip learning.IEEE Micro, 38(1):82–99, 2018

  14. [22]

    Incorporating learnable membrane time constant to enhance learn- ing of spiking neural networks

    Wei Fang, Zhaofei Yu, Yanqi Chen, Timoth´ ee Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learn- ing of spiking neural networks. In2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2641–2651, 2021. 33

  15. [23]

    Brain-inspired learning on neuromorphic substrates.Proceedings of the IEEE, 109(5):935–950, 2021

    Friedemann Zenke and Emre O Neftci. Brain-inspired learning on neuromorphic substrates.Proceedings of the IEEE, 109(5):935–950, 2021

  16. [24]

    Spike-driven transformer V2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips

    Man Yao, JiaKui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer V2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. InThe Twelfth International Conference on Learning Re...

  17. [25]

    Spiking deep convolutional neural networks for energy-efficient object recognition.International Journal of Computer Vision, 113(1):54–66, 2015

    Yongqiang Cao, Yang Chen, and Deepak Khosla. Spiking deep convolutional neural networks for energy-efficient object recognition.International Journal of Computer Vision, 113(1):54–66, 2015

  18. [26]

    Towards spike-based ma- chine intelligence with neuromorphic computing.Nature, 575(7784):607–617, 2019

    Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. Towards spike-based ma- chine intelligence with neuromorphic computing.Nature, 575(7784):607–617, 2019

  19. [27]

    Benchmarking energy consumption and latency for neuro- morphic computing in condensed matter and particle physics.APL Machine Learn- ing, 1(1), 2023

    Dominique J K¨ osters et al. Benchmarking energy consumption and latency for neuro- morphic computing in condensed matter and particle physics.APL Machine Learn- ing, 1(1), 2023

  20. [28]

    Neuromor- phic principles for efficient large language models on Intel Loihi 2

    Steven Abreu, Sumit Bam Shrestha, Rui-Jie Zhu, and Jason Eshraghian. Neuromor- phic principles for efficient large language models on Intel Loihi 2. InFirst Workshop on Scalable Optimization for Efficient and Adaptive Foundation Models, 2025

  21. [29]

    Energy-efficient neuromorphic computing for edge AI: A framework with adaptive spiking neural networks and hardware-aware optimization

    Olaf Yunus Laitinen Imanov, Derya Umut Kulali, Taner Yilmaz, Duygu Erisken, and Rana Irem Turhan. Energy-efficient neuromorphic computing for edge AI: A framework with adaptive spiking neural networks and hardware-aware optimization. arXiv preprint arXiv:2602.02439, 2026

  22. [30]

    Backpropagation through time: what it does and how to do it

    Paul J Werbos. Backpropagation through time: what it does and how to do it. Proceedings of the IEEE, 78(10):1550–1560, 1990

  23. [31]

    Efficient training of spiking neural networks with temporally-truncated local back- propagation through time.Frontiers in neuroscience, 17:1047008, 2023

    Wenzhe Guo, Mohammed E Fouda, Ahmed M Eltawil, and Khaled Nabil Salama. Efficient training of spiking neural networks with temporally-truncated local back- propagation through time.Frontiers in neuroscience, 17:1047008, 2023

  24. [32]

    Neftci, Hesham Mostafa, and Friedemann Zenke

    Emre O. Neftci, Hesham Mostafa, and Friedemann Zenke. Surrogate gradient learn- ing in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks.IEEE Signal Processing Magazine, 36(6):51–63, 2019

  25. [33]

    The remarkable robustness of surrogate gra- dient learning for instilling complex function in spiking neural networks.Neural Computation, 33(4):899–925, 2021

    Friedemann Zenke and Tim P Vogels. The remarkable robustness of surrogate gra- dient learning for instilling complex function in spiking neural networks.Neural Computation, 33(4):899–925, 2021

  26. [34]

    Training spik- ing neural networks using lessons from deep learning.Proceedings of the IEEE, 111(9):1016–1054, 2023

    Jason K Eshraghian, Max Ward, Emre O Neftci, Xinxin Wang, Gregor Lenz, Girish Dwivedi, Mohammed Bennamoun, Doo Seok Jeong, and Wei D Lu. Training spik- ing neural networks using lessons from deep learning.Proceedings of the IEEE, 111(9):1016–1054, 2023

  27. [35]

    Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks

    Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks. InInternational Conference on Learning Representations, 2022. 34

  28. [36]

    Deep learning in spiking neural networks.Neural Networks, 111:47–63, 2019

    Amirhossein Tavanaei, Masoud Ghodrati, Saeed Reza Kheradpisheh, Timoth´ ee Masquelier, and Anthony Maida. Deep learning in spiking neural networks.Neural Networks, 111:47–63, 2019

  29. [37]

    Spatio-temporal back- propagation for training high-performance spiking neural networks

    Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal back- propagation for training high-performance spiking neural networks. InFrontiers in Neuroscience, volume 12, page 331, 2018

  30. [38]

    Feature at- tribution explanations for spiking neural networks

    Elisa Nguyen, Meike Nauta, Gwenn Englebienne, and Christin Seifert. Feature at- tribution explanations for spiking neural networks. In2023 IEEE 5th International Conference on Cognitive Machine Intelligence (CogMI), pages 59–68, 2023

  31. [39]

    Gradient-based feature- attribution explainability methods for spiking neural networks.Frontiers in Neu- roscience, 17:1153999, 2023

    Ammar Bitar, Rafael Rosales, and Michael Paulitsch. Gradient-based feature- attribution explainability methods for spiking neural networks.Frontiers in Neu- roscience, 17:1153999, 2023

  32. [40]

    Binary spiking neural net- works as causal models

    Aditya Kar, Emiliano Lorini, and Timoth´ ee Masquelier. Binary spiking neural net- works as causal models. InThe Thirteenth International Conference on Learning Representations, 2025

  33. [41]

    Deep learning of explainable EEG patterns as dynamic spatiotemporal clusters and rules in a brain-inspired spiking neural network.Sensors, 21(14):4900, 2021

    Maryam Doborjeh, Zohreh Doborjeh, Nikola Kasabov, Molood Barati, and Grace Y Wang. Deep learning of explainable EEG patterns as dynamic spatiotemporal clusters and rules in a brain-inspired spiking neural network.Sensors, 21(14):4900, 2021

  34. [42]

    In defense of one-vs-all classification.Journal of machine learning research, 5(Jan):101–141, 2004

    Ryan Rifkin and Aldebaro Klautau. In defense of one-vs-all classification.Journal of machine learning research, 5(Jan):101–141, 2004

  35. [43]

    One-vs-one classification for deep neural networks

    Pornntiwa Pawara, Emmanuel Okafor, Marc Groefsema, Sheng He, Lambert RB Schomaker, and Marco A Wiering. One-vs-one classification for deep neural networks. Pattern Recognition, 108:107528, 2020

  36. [44]

    Plasticity in inhibitory networks improves pattern separation in early olfactory processing.Communications biology, 8(1):590, 2025

    Shruti Joshi, Seth Haney, Zhenyu Wang, Fernando Locatelli, Hong Lei, Yu Cao, Brian Smith, and Maxim Bazhenov. Plasticity in inhibitory networks improves pattern separation in early olfactory processing.Communications biology, 8(1):590, 2025

  37. [45]

    Synaptic activity and the construction of cortical circuits.Science, 274(5290):1133–1138, 1996

    Lawrence C Katz and Carla J Shatz. Synaptic activity and the construction of cortical circuits.Science, 274(5290):1133–1138, 1996

  38. [46]

    Synapse elimination and indelible memory

    Jeff W Lichtman and Howard Colman. Synapse elimination and indelible memory. Neuron, 25(2):269–278, 2000

  39. [47]

    G. Yang, C. S. Lai, J. Cichon, L. Ma, W. Li, and W. B. Gan. Sleep promotes branch-specific formation of dendritic spines after learning.Science, 344:1173–1178, 2014

  40. [48]

    Y. Zhou, C. S. W. Lai, Y. Bai, W. Li, R. Zhao, G. Yang, M. G. Frank, and W. B. Gan. REM sleep promotes experience-dependent dendritic spine elimination in the mouse cortex.Nature Communications, 11:4819, 2020

  41. [49]

    Complementary contributions of non-REM and REM sleep to visual learning.Nature neuroscience, 23(9):1150–1156, 2020

    Masako Tamaki, Zhiyan Wang, Tyler Barnes-Diana, DeeAnn Guo, Aaron V Berard, Edward Walsh, Takeo Watanabe, and Yuka Sasaki. Complementary contributions of non-REM and REM sleep to visual learning.Nature neuroscience, 23(9):1150–1156, 2020. 35

  42. [50]

    Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278– 2324, 1998

    Yann LeCun, L´ eon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278– 2324, 1998

  43. [51]

    Fashion-MNIST: a novel image dataset for benchmarking machine learning algorithms.arXiv preprint arXiv:1708.07747, 2017

    Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-MNIST: a novel image dataset for benchmarking machine learning algorithms.arXiv preprint arXiv:1708.07747, 2017

  44. [52]

    Unsupervised learning of digit recognition us- ing spike-timing-dependent plasticity

    Peter U Diehl and Matthew Cook. Unsupervised learning of digit recognition us- ing spike-timing-dependent plasticity. InFrontiers in Computational Neuroscience, volume 9, page 99, 2015

  45. [53]

    STDP-based spiking deep convolutional neural networks for object recog- nition.Neural Networks, 99:56–67, 2018

    Saeed Reza Kheradpisheh, Mohammad Ganjtabesh, Simon J Thorpe, and Timoth´ ee Masquelier. STDP-based spiking deep convolutional neural networks for object recog- nition.Neural Networks, 99:56–67, 2018

  46. [54]

    Training deep spiking neural networks using backpropagation.Frontiers in Neuroscience, 10:508, 2016

    Jun Haeng Lee, Tobi Delbruck, and Michael Pfeiffer. Training deep spiking neural networks using backpropagation.Frontiers in Neuroscience, 10:508, 2016

  47. [55]

    A biologically plausible super- vised learning method for spiking neural networks using the symmetric STDP rule

    Yunzhe Hao, Xuhui Huang, Meng Dong, and Bo Xu. A biologically plausible super- vised learning method for spiking neural networks using the symmetric STDP rule. Neural Networks, 121:387–395, 2020

  48. [56]

    LISNN: Improving spiking neural networks with lateral interactions for robust object recognition

    Xiang Cheng, Yunzhe Hao, Jiaming Xu, and Bo Xu. LISNN: Improving spiking neural networks with lateral interactions for robust object recognition. InIJCAI, pages 1519–1525. Yokohama, 2020

  49. [57]

    Diet-snn: Direct input encoding with leak- age and threshold optimization in deep spiking neural networks.arXiv preprint arXiv:2008.03658, 2020

    Nitin Rathi and Kaushik Roy. Diet-snn: Direct input encoding with leak- age and threshold optimization in deep spiking neural networks.arXiv preprint arXiv:2008.03658, 2020

  50. [58]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, Toronto, Ontario, 2009

  51. [59]

    Spiking deep networks with LIF neurons

    Eric Hunsberger and Chris Eliasmith. Spiking deep networks with LIF neurons. arXiv preprint arXiv:1510.08829, 2015

  52. [60]

    Direct training for spiking neural networks: Faster, larger, better

    Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. InProceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 1311–1318, 2019

  53. [61]

    Enabling deep spiking neural networks with hybrid conversion and spike timing de- pendent backpropagation

    Nitin Rathi, Gopalakrishnan Srinivasan, Priyadarshini Panda, and Kaushik Roy. Enabling deep spiking neural networks with hybrid conversion and spike timing de- pendent backpropagation. InInternational Conference on Learning Representations, 2020

  54. [62]

    Activity pruning for efficient spiking neural networks

    Tong Bu, Xinyu Shi, and Zhaofei Yu. Activity pruning for efficient spiking neural networks. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

  55. [63]

    Catastrophic forgetting in connectionist networks.Trends in cognitive sciences, 3(4):128–135, 1999

    Robert M French. Catastrophic forgetting in connectionist networks.Trends in cognitive sciences, 3(4):128–135, 1999. 36

  56. [64]

    Replay in deep learning: Current approaches and missing biological elements.Neural Computation, 33(11):2908–2950, 2021

    Tyler L Hayes, Giri P Krishnan, Maxim Bazhenov, Hava T Siegelmann, Terrence J Sejnowski, and Christopher Kanan. Replay in deep learning: Current approaches and missing biological elements.Neural Computation, 33(11):2908–2950, 2021

  57. [65]

    Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory

    James L McClelland, Bruce L McNaughton, and Randall C O’Reilly. Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychological review, 102(3):419–457, 1995

  58. [66]

    Catastrophic interference in connectionist net- works: The sequential learning problem

    Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist net- works: The sequential learning problem. InPsychology of Learning and Motivation, volume 24, pages 109–165. Elsevier, 1989

  59. [67]

    Biologi- cally inspired sleep algorithm for increased generalization and adversarial robustness in deep neural networks

    Timothy Tadros, Giri Krishnan, Ramyaa Ramyaa, and Maxim Bazhenov. Biologi- cally inspired sleep algorithm for increased generalization and adversarial robustness in deep neural networks. InInternational Conference on Learning Representations, 2020

  60. [68]

    Toward lifelong learning in equilibrium propagation: Sleep-like and awake rehearsal for enhanced stability

    Yoshimasa Kubo, Jean Erik Delanois, and Maxim Bazhenov. Toward lifelong learning in equilibrium propagation: Sleep-like and awake rehearsal for enhanced stability. arXiv preprint arXiv:2508.14081, 2025

  61. [69]

    Learning without forgetting.IEEE transactions on pattern analysis and machine intelligence, 40(12):2935–2947, 2017

    Zhizhong Li and Derek Hoiem. Learning without forgetting.IEEE transactions on pattern analysis and machine intelligence, 40(12):2935–2947, 2017

  62. [70]

    Overcoming catastrophic forgetting in neural networks

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Des- jardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sci...

  63. [71]

    Continual learning through synap- tic intelligence

    Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synap- tic intelligence. InProceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3987–3995. JMLR. org, 2017

  64. [72]

    Continual learning of context- dependent processing in neural networks.Nature Machine Intelligence, 1(8):364–372, 2019

    Guanxiong Zeng, Yang Chen, Bo Cui, and Shan Yu. Continual learning of context- dependent processing in neural networks.Nature Machine Intelligence, 1(8):364–372, 2019

  65. [73]

    Intrinsic and circuit prop- erties favor coincidence detection for decoding oscillatory input.Journal of Neuro- science, 24(26):6037–6047, 2004

    Javier Perez-Orive, Maxim Bazhenov, and Gilles Laurent. Intrinsic and circuit prop- erties favor coincidence detection for decoding oscillatory input.Journal of Neuro- science, 24(26):6037–6047, 2004

  66. [74]

    R. A. Dielenberg and I. S. McGregor. Defensive behavior in rats towards predatory odors: a review.Neuroscience and Biobehavioral Reviews, 25:597–609, 2001

  67. [75]

    R. T. Card´ e. Navigation along windborne plumes of pheromone and resource-linked odors.Annual Review of Entomology, 66:317–336, 2021

  68. [76]

    Tumkaya, S

    T. Tumkaya, S. Burhanudin, A. Khalilnezhad, J. Stewart, H. Choi, and A. Claridge- Chang. Most primary olfactory neurons have individually neutral effects on behavior. eLife, 11:e71238, 2022. 37

  69. [77]

    Distributed plasticity for olfactory learning and memory in the honey bee brain

    Brian H Smith, Ram´ on Huerta, Maxim Bazhenov, and Irina Sinakevitch. Distributed plasticity for olfactory learning and memory in the honey bee brain. InHoneybee Neurobiology and Behavior: A Tribute to Randolf Menzel, pages 393–408. Springer, 2011

  70. [78]

    Novelty detection in early olfactory processing of the honey bee, Apis mellifera.Plos one, 17(3):e0265009, 2022

    Hong Lei, Seth Haney, Christopher M Jernigan, Xiaojiao Guo, Chelsea N Cook, Maxim Bazhenov, and Brian H Smith. Novelty detection in early olfactory processing of the honey bee, Apis mellifera.Plos one, 17(3):e0265009, 2022

  71. [79]

    Pedigo, Christopher L

    Michael Winding, Benjamin D. Pedigo, Christopher L. Barnes, Heather G. Patsolic, Youngser Park, Tom Kazimiers, Akira Fushiki, Ingrid V. Andrade, Avinash Khan- delwal, Javier Valdes-Aleman, Feng Li, Nadine Randel, Elizabeth Barsotti, Ana Cor- reia, Richard D. Fetter, Volker Har...

  72. [80]

    The REM sleep-memory consolidation hypothesis.Science, 294(5544):1058–1063, 2001

    Jerome M Siegel. The REM sleep-memory consolidation hypothesis.Science, 294(5544):1058–1063, 2001

  73. [81]

    John Peever and Patrick M. Fuller. The biology of REM sleep.Current Biology, 27:R1237–R1248, 2017

  74. [82]

    Sleep- like unsupervised replay reduces catastrophic forgetting in artificial neural networks

    Timothy Tadros, Giri P Krishnan, Ramyaa Ramyaa, and Maxim Bazhenov. Sleep- like unsupervised replay reduces catastrophic forgetting in artificial neural networks. Nature Communications, 13(1):7742, 2022

  75. [83]

    Improving robustness of convolutional networks through sleep-like replay

    Jean Erik Delanois, Aditya Ahuja, Giri P Krishnan, Timothy Tadros, Julian McAuley, and Maxim Bazhenov. Improving robustness of convolutional networks through sleep-like replay. In2023 22nd IEEE International Conference on Machine Learning and Applications (ICMLA), pages 257–26...

  76. [84]

    Stanley Heinze, Jeremy Florman, Surainder Asokaraj, Basil el Jundi, and Steven M. Reppert. Anatomical basis of sun compass navigation II: The neuronal composition of the central complex of the monarch butterfly.Journal of Comparative Neurology, 521(2):267–298, 2013

  77. [85]

    An anatomically constrained model for path integration in the bee brain.Current Biology, 27(20):3069–3085, 2017

    Thomas Stone, Barbara Webb, Andrea Adden, Nicolai Ben Weddig, Anna Honkanen, Rachel Templin, William Wcislo, Luca Scimeca, Eric Warrant, and Stanley Heinze. An anatomically constrained model for path integration in the bee brain.Current Biology, 27(20):3069–3085, 2017

  78. [86]

    CASIA online and offline chinese handwriting databases

    Cheng-Lin Liu, Fei Yin, Da-Han Wang, and Qiu-Feng Wang. CASIA online and offline chinese handwriting databases. In2011 International Conference on Document Analysis and Recognition, pages 37–41. IEEE, 2011

  79. [87]

    Adap- tive mixtures of local experts.Neural Computation, 3(1):79–87, 1991

    Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adap- tive mixtures of local experts.Neural Computation, 3(1):79–87, 1991

  80. [88]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. InInternational Conference on Learning Representations (ICLR), 2017. 38

  81. [89]

    An empirical study of catastrophic forgetting in large language models during continual fine-tuning

    Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. An empirical study of catastrophic forgetting in large language models during continual fine-tuning. IEEE Transactions on Audio, Speech and Language Processing, 33:3776–3786, 2025

  82. [90]

    A continual learning survey: Defying forgetting in classification tasks.IEEE transactions on pattern analysis and machine intelligence, 44(7):3366–3385, 2021

    Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleˇ s Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks.IEEE transactions on pattern analysis and machine intelligence, 44(7):3366–33...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.