Pith. sign in

REVIEW 4 major objections 6 minor 60 references

Integrating Complexity and Biological Realism: High-Performance Spiking Neural Networks for Breast Cancer Detection

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Spiking neural networks with a Lempel-Ziv complexity readout reach 98.25% accuracy on breast cancer classification at up to 100 times lower computational cost.

desk verdict The LB neuron as defined always fires, so the paper's central LB accuracies are unexplained; the learning-rule comparison is useful but the manuscript is not publishable as written. read the letter →

arxiv 2506.06265 v1 pith:OLCIHIVJ submitted 2025-06-06 cs.NE eess.IVq-bio.NC

Zofia Rudnicka , Januszcz Szczepanski , Agnieszka Pregowska This is my paper
classification cs.NEeess.IVq-bio.NC
keywords spikingneuralnetworksLempel-ZivcomplexityLevy-BaxterneuronLeakyIntegrate-and-FirebreastcancerclassificationANN-to-SNNconversionmedicaldiagnosticsneuromorphiccomputing
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 is trying to establish that spiking neural networks (SNNs) can be made competitive with conventional deep learning for breast cancer classification if their binary spike outputs are scored with Lempel-Ziv complexity (LZC), a measure of sequence structure. On the Breast Cancer Wisconsin dataset, the authors report a top accuracy of 98.25% from an ANN-to-SNN conversion applied to both Leaky Integrate-and-Fire (LIF) and Levy-Baxter (LB) neuron models, matching backpropagation-trained networks while claiming up to 100 times lower computational cost. LB-based networks exceed 90% accuracy across supervised and unsupervised spike-based learning rules; LIF-based networks reach above 85%. The practical interest is that a biologically plausible, event-driven network with a cheap complexity readout could serve resource-constrained or real-time diagnostic tools.

What carries the argument

The load-bearing mechanism is the Lempel-Ziv complexity score $c_2(x_1^n) = C_2(x_1^n)/(n \log_2 n)$, a measure of the number of distinct substrings in a binary sequence, computed on the network's binary spike output and used as the classification readout; a higher complexity score is treated as richer temporal feature content. The paper pairs this with two neuron models: the standard Leaky Integrate-and-Fire dynamics (membrane potential integration with threshold and reset) and the probabilistic Levy-Baxter neuron, whose stochastic release variables $\phi_i$ and $Q_i$ are meant to create the variable spiking behavior LZC exploits. The ANN-to-SNN conversion, mapping trained weights as $w_{\mathrm{SNN}} = w_{\mathrm{ANN}} / \tau_{\mathrm{syn}}$, is the specific procedure that yields the top accuracy.

What would settle it

Re-implement the LB neuron exactly from Eqs. (3) and (4) on the Wisconsin features after the paper's spike encoding and count the output zeros; because every input is nonnegative, the neuron should emit a spike for every pattern, yielding a degenerate spike train and nullifying LZC-based discrimination. That result would settle that the reported 90%+ LB accuracies depend on a model detail not stated in the paper.

Watch

Extended reading notes

Core claim

The central claim is that combining SNNs with LZC as a readout yields diagnostic accuracy comparable to backpropagation-trained deep networks, with the best result of 98.25% obtained by converting a trained ANN into either an LIF or an LB spiking network via weight scaling $w_{\mathrm{SNN}} = w_{\mathrm{ANN}} / \tau_{\mathrm{syn}}$. The authors attribute the method's success to a synergy: LIF and LB neurons encode inputs as spike trains, LZC measures the structural complexity of those trains, and a more variable spike train gives the complexity measure more discriminative information. This is why the probabilistic Levy-Baxter neuron leads in most spike-based learning regimes (94.74% with Hebbian learning, 92.98% with STDP and SpikeProp, 91.23% with Tempotron), while the regular, deterministic firing of LIF neurons makes it stronger in reward-modulated and active-learning hybrids. The paper presents the result as evidence that the biologically inspired approach can bridge the gap between biological plausibility and computational efficiency.

Load-bearing premise

The load-bearing premise is that the Levy-Baxter neuron as defined can produce genuinely variable spike responses; as written it fires on every input, because $\sigma$, a sum of nonnegative terms, always satisfies $\sigma \geq 0$.

Editorial extensions

If this is right

  • ANN-to-SNN conversion can transfer a pre-trained network's accuracy onto spiking hardware without directly training spike times, so the 98.25% result is available to event-driven or neuromorphic implementations.
  • LB-based spiking networks stay above 90% accuracy under Hebbian, STDP, SpikeProp, and Tempotron learning, making probabilistic neurons a plausible alternative to LIF when temporal spike patterns carry the signal.
  • Using LZC as the readout turns the classifier's output into a single interpretable complexity score per spike train, which can simplify the output layer of an SNN.
  • Because the reported SNN accuracy is comparable to backprop while the compute is claimed to be up to 100 times lower, the approach points toward real-time or resource-constrained diagnostic deployment.

Reading between the lines

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

  • If Section 3's LB equations are read literally, $\sigma$ is a sum of nonnegative terms $\phi_i Q_i x_i$, so the condition $\sigma \geq 0$ is always met and the neuron fires on every input; reproducing the reported LB accuracies therefore requires an unstated modification to the threshold, sign convention, or input encoding.
  • The paper does not report the energy or wall-clock measurement behind the stated computational-cost reduction, and the abstract's 'up to 100 times' differs from the conclusion's 'up to 50 times', so the efficiency claim is best read as a hypothesis about spike sparsity rather than a benchmarked result.
  • The experiments use a single 569-sample dataset with one 80/20 split, so whether the LZC readout transfers to larger medical imaging datasets such as mammography, ultrasound, or histopathology is untested.
  • Because the highest accuracy comes from ANN-to-SNN conversion rather than from spike-based training, the claimed biological realism is strongest for inference, not for learning; a testable extension would be training the converted SNN directly with spike-based rules and checking whether the accuracy gap narrows further.
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

4 major / 6 minor

Summary. The manuscript proposes a spiking neural network (SNN) framework combined with Lempel-Ziv Complexity (LZC) for breast cancer classification on the Wisconsin Diagnostic dataset. It compares Leaky Integrate-and-Fire (LIF) and Levy-Baxter (LB) neuron models under unsupervised, supervised, and hybrid learning rules, reporting accuracies in Table 2, with the highest accuracy of 98.25% attributed to ANN-to-SNN conversion. The paper claims that LB-based models generally outperform LIF-based models and that the approach is computationally efficient and biologically plausible.

Significance. If the results were reproducible and the models correctly specified, the paper would offer a useful benchmark of diverse SNN learning rules on a standard medical dataset. The systematic comparison across Hebbian, STDP, SpikeProp, Tempotron, backpropagation, and hybrid rules is a relevant contribution. However, the significance is severely limited by a load-bearing flaw in the LB neuron definition, the absence of error bars or statistical testing, and unsupported claims about LZC benefit and computational cost. The paper does not release code or detailed experimental settings, so the reported accuracies cannot be independently verified.

major comments (4)
  1. [Section 3, Eqs. (3)-(4)] The Levy-Baxter neuron as defined is degenerate: the total synaptic excitation sigma is a sum of nonnegative terms (xi in {0,1}, phi_i Bernoulli with success probability in [0,1], and Qi uniform on [0,1]), so sigma >= 0 always holds. Equation (4) therefore always sets z = 1, and the branch z = 0 for sigma < 0 can never occur. Every LB neuron emits a constant all-ones spike train for all inputs, so LZC values are identical across samples and the LB accuracies in Table 2 (91-98%) cannot follow from the stated model. No threshold, sign convention, or input-encoding modification is described anywhere in the manuscript. This is an internal mathematical contradiction at the core of the paper's central claim that LB-based models outperform LIF-based models.
  2. [Table 2 and Section 7] Each reported accuracy is a single percentage with no error bars, no number of repeated runs, and no statistical significance tests. Differences of 1-3 percentage points between models (e.g., 94.74% vs. 92.98%) are presented as meaningful advantages, but without variance estimates these differences are not interpretable. In addition, the text states that 'the reported results correspond to the most efficient configuration in each case,' which, combined with the absence of a held-out validation procedure, suggests potential overfitting to the test set through hyperparameter selection.
  3. [Abstract, Section 7, and Conclusions] The computational cost claim is inconsistent and unsupported: the abstract states 'up to 100 times lower computational cost,' while the conclusions state 'up to 50 times lower in some cases.' No methodology for measuring computational cost is given, and no runtime or energy measurements are reported. Furthermore, the 98.25% ANN-to-SNN result is obtained by transferring weights from a trained ANN via Eq. (13), so the accuracy is inherited from the ANN rather than achieved by SNN training; this does not support the claim that the SNN framework itself matches deep learning performance.
  4. [Section 7, LZC claim] The paper claims that 'integrating LZC into the SNN framework significantly improved classification accuracy and reduced computational cost,' but no ablation study with and without LZC is presented in this manuscript. The only supporting evidence cited is reference [60], a self-citation to an unpublished work. This is a load-bearing claim for the proposed method, and it needs direct experimental support within the paper, including a comparison of classification accuracy and computational cost with and without LZC.
minor comments (6)
  1. [Table 2] Table 2 is difficult to read as typeset: the learning algorithm names and category labels are run together without clear separation, and the column formatting is inconsistent.
  2. [Section 5, Eqs. (7) and (14)] Equations (7) and (14) contain misplaced commas in the weight update expressions, which should be products of the listed terms.
  3. [Table 1] Table 1 lists 'tau_m^+' for the membrane time constant, but the text uses tau_m without the superscript; notation should be consistent.
  4. [Section 4] The text states that the network processes binary sequences of fixed length L=30, but the dataset has 30 numerical features; the conversion from continuous features to binary sequences is not described.
  5. [References] Reference [60] is cited as the sole support for the LZC improvement but is a self-citation to an unpublished or non-archived work; it should be clearly labeled or replaced with peer-reviewed evidence.
  6. [Data Availability] The Data Availability Statement says 'Not applicable,' even though the Breast Cancer Wisconsin dataset is publicly available; the statement should describe dataset access and any code availability.

Circularity Check

2 steps flagged · score 6.0 of 10

The LZC advantage is supported only by a same-author citation, and the headline 98.25% accuracy is copied from a fitted ANN by Eq. (13); the LB neuron equations also preclude the reported sample-dependent results.

  1. self citation load bearing [Section 7, Results and Discussion, final paragraph after Table 2]
    "Moreover, integrating Lempel-Ziv Complexity (LZC) into the SNN framework significantly improved classification accuracy and reduced computational cost, especially when paired with dynamic spiking models. LZC’s ability to quantify sequence complexity helped to extract richer features from the temporally encoded spike patterns, enhancing the discriminative power of the network. The same observations were confirmed by [60]."

    The paper's central contribution is the SNN+LZC combination, yet no ablation (SNN with LZC vs. without LZC) is reported in this manuscript. The only support for the claim that LZC improves accuracy is reference [60], which lists the same three authors (Rudnicka, Szczepanski, Pregowska) and is not machine-checked, code-reproduced, or independently verified here. The load-bearing assertion that the novel component works therefore reduces to a same-author citation rather than to evidence in this paper.

  2. fitted input called prediction [Section 5, Eq. (13) and Section 7, Table 2]
    "The mapping is typically defined as: wSNN = wANN / τsyn ... The highest classification accuracy of 98.25% was achieved using an Artificial-to-Spiking Neural Network conversion, applied to both the Leaky Integrate-and-Fire and Levy-Baxter neuron models."

    Eq. (13) defines the SNN weights as a rescaling of a conventionally trained ANN's weights. The 98.25% figure in Table 2 is therefore the fitted ANN's decision boundary transplanted into the spiking architecture; it is not produced by the SNN's dynamics or by LZC feature extraction. Presenting this converted model as the top result of the proposed hybrid approach treats the fitted ANN's accuracy as an independent SNN prediction, even though the accuracy is inherited by construction through the weight-copying rule.

full rationale

The accuracy numbers in Table 2 are empirical measurements and not circular by themselves. The circularity is in two load-bearing interpretive moves. First, the claimed benefit of LZC—the paper's novel ingredient—is supported only by a same-author citation ([60]) because no with/without-LZC ablation is run; the mechanism's value is thus asserted rather than demonstrated in this work. Second, the headline 98.25% result comes from ANN-to-SNN conversion, where Eq. (13) copies the trained ANN weights into the SNN, so the SNN's accuracy is inherited from a fitted model by construction and is not a validation of the spiking or complexity machinery. Two additional non-circular problems affect the conclusions and should be weighed separately: the LB neuron of Eqs. (3)-(4) always fires (σ≥0 for all inputs since xi, φi, Qi are nonnegative), so the sample-dependent LB accuracies in Table 2 are inconsistent with the stated model; and hyperparameters were selected as the 'most efficient configuration' per algorithm without a reported held-out tuning procedure, raising a multiple-comparisons risk. These are correctness concerns rather than circularity, but they reinforce that the paper's strongest claims outrun its derivations.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The paper's results depend on tuned hyperparameters, several domain assumptions, and an unstated fix to the LB threshold rule. The LZC benefit is supported only by self-citation, and the ANN-to-SNN conversion claim rests on an unverified weight-scaling assumption.

free parameters (8)
  • membrane threshold theta = range [0.1, 0.5]; exact best values not reported
    Systematically varied; reported results use the most efficient configuration (Section 4).
  • membrane decay delta = range [0.01, 0.1]; exact best values not reported
    Systematically varied with threshold and learning rate (Section 4).
  • learning rate eta = range [1e-4, 1e-1]; exact best values not reported
    Tuned per learning algorithm with Optuna.
  • regularization coefficient lambda = not reported
    L2 penalty in Eq. (5), tuned with Optuna to encourage sparsity.
  • STDP amplitudes and time constants A+, A-, tau+, tau- = not reported
    Define STDP weight change in Eq. (8).
  • hidden layer size n = selected from {2, 8, 16, 30}
    Architecture choice per condition, not justified.
  • synaptic time constant tau_syn = not reported
    Used in ANN-to-SNN conversion Eq. (13); affects weight scaling.
  • PCA components and SMOTE parameters = not reported
    Preprocessing choices that change input representation.
assumptions (5)
  • standard math Normalized Lempel-Ziv complexity c2(x) asymptotically tends to 1 for random sequences and 0 for deterministic sequences
    Standard result from Ziv and Lempel, used in Eq. (6).
  • domain assumption The Wisconsin Diagnostic dataset's 30 features and binary labels are a valid proxy for breast cancer diagnostic imaging
    The paper generalizes conclusions to medical imaging and clinical use from this tabular benchmark.
  • ad hoc to paper Applying LZC to spike outputs improves classification accuracy and interpretability
    No mechanism or ablation is given; the claim is supported by self-citation [60].
  • ad hoc to paper The Levy-Baxter threshold rule in Eq. (4) produces meaningful spike activity
    As written, the condition sigma >= 0 is always true, so neurons always fire; results require an unstated threshold or sign convention.
  • ad hoc to paper ANN-to-SNN conversion via wSNN = wANN / tau_syn preserves accuracy
    Eq. (13) is used without threshold analysis or simulation details.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Integrating Complexity and Biological Realism: High-Performance Spiking Neural Networks for Breast Cancer Detection." pith.science (2026). https://pith.science/paper/OLCIHIVJ

@misc{pith2026250606265,
  author       = {Pith},
  title        = {Pith review of: Integrating Complexity and Biological Realism: High-Performance Spiking Neural Networks for Breast Cancer Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OLCIHIVJ}},
  note         = {Machine review of arXiv:2506.06265}
}
read the original abstract

Spiking Neural Networks (SNNs) event-driven nature enables efficient encoding of spatial and temporal features, making them suitable for dynamic time-dependent data processing. Despite their biological relevance, SNNs have seen limited application in medical image recognition due to difficulties in matching the performance of conventional deep learning models. To address this, we propose a novel breast cancer classification approach that combines SNNs with Lempel-Ziv Complexity (LZC) a computationally efficient measure of sequence complexity. LZC enhances the interpretability and accuracy of spike-based models by capturing structural patterns in neural activity. Our study explores both biophysical Leaky Integrate-and-Fire (LIF) and probabilistic Levy-Baxter (LB) neuron models under supervised, unsupervised, and hybrid learning regimes. Experiments were conducted on the Breast Cancer Wisconsin dataset using numerical features derived from medical imaging. LB-based models consistently exceeded 90.00% accuracy, while LIF-based models reached over 85.00%. The highest accuracy of 98.25% was achieved using an ANN-to-SNN conversion method applied to both neuron models comparable to traditional deep learning with back-propagation, but at up to 100 times lower computational cost. This hybrid approach merges deep learning performance with the efficiency and plausibility of SNNs, yielding top results at lower computational cost. We hypothesize that the synergy between temporal-coding, spike-sparsity, and LZC-driven complexity analysis enables more-efficient feature extraction. Our findings demonstrate that SNNs combined with LZC offer promising, biologically plausible alternative to conventional neural networks in medical diagnostics, particularly for resource-constrained or real-time systems.

Figures

Figures reproduced from arXiv: 2506.06265 by the authors.

Figure 1
Figure 1. Schematic of the LIF and Levy–Baxter neuron models. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Spiking neural network scheme inspired by biological neuron connectivity. Input [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Schema of the operating principle of the Hebbian learning algorithm. Neurons [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Schema of the operating principle of the Hebbian learning algorithm combined [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Schema of the operating principle of the Spike-Timing-Dependent Plasticity [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Schema of the operating principle of the backpropagation learning algorithm. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Scheme of the operating principle of the SpikeProp learning rule. Input neurons [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 9
Figure 9. Figure 9: The mapping is typically defined as: wSNN = wANN τsyn , (13) where τsyn denotes the synaptic time constant. Activation patterns from ANNs are preserved in SNNs by frequency coding, where the firing rates approximate continuous outputs. However, network accuracy is sens…
Figure 8
Figure 8. Figure 8: Schematic illustration of the operating principle of the Tempotron learning rule. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Schematic illustration of the operating principle of theANN-to-SNN conversion. [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Schematic illustration of the operating principle of the reward-based SNN [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Schematic of the BAL. Input x is processed by the network fSNN , producing a prediction yˆ. The synaptic update ∆wis driven by a plasticity rule modulated by the synaptic uncertainty U(wi) and the expected mutual information I(Spost; Spre) between pre- and post-synapt…
Figure 12
Figure 12. Figure 12: Construction of the Breast Cancer Wisconsin (Diagnostic) dataset. [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 50 canonical work pages

  1. [60]

    Bio-Inspired Classification: Combining Information Theory and Spiking Neural Networks - Influ- ence of the Learning Rules

    Rudnicka Z, Szczepanski J, Pregowska A. Bio-Inspired Classification: Combining Information Theory and Spiking Neural Networks - Influ- ence of the Learning Rules. 2025. 25

    work page 2025

  2. [1]

    Breast Cancer Statistics

    American Cancer Society. Breast Cancer Statistics. https://www.cancer.org/cancer/types/breast-cancer/about/how- common-is-breast-cancer.html. Accessed May 5, 2025

    work page 2025

  3. [2]

    Thomassin-Naggara I, Kilburn-Toppin F, Athanasiou A, et al. Misdi- agnosis in breast imaging: a statement paper from European Society Breast Imaging (EUSOBI)—Part 1: The role of common errors in ra- diology in missed breast cancer and implications of misdiagnosis. Eur Radiol. 2024. https://doi.org/10.1007/s00330-024-11128-1. 19

    work page doi:10.1007/s00330-024-11128-1 2024

  4. [3]

    Oncolytic virus-based combi- nation therapy in breast cancer

    Bahreyni A, Mohamud Y, Luo H. Oncolytic virus-based combi- nation therapy in breast cancer. Cancer Lett. 2024;585:216634. https://doi.org/10.1016/j.canlet.2024.216634

    work page arXiv 2024

  5. [4]

    Gerratana L, Davis AA, Foffano L, et al. Integrating machine learning-predicted circulating tumor cells (CTCs) and circulating tu- mor DNA (ctDNA) in metastatic breast cancer: A proof of principle study on endocrine resistance profiling. Cancer Lett. 2025;609:217325. https://doi.org/10.1016/j.canlet.2024.217325

    work page arXiv 2025

  6. [5]

    SAFNet: A deep spa- tial attention network with classifier fusion for breast cancer detection

    Lu SY, Wang SH, Zhang YD. SAFNet: A deep spa- tial attention network with classifier fusion for breast cancer detection. Comput Biol Med. 2022;148:105812. https://doi.org/10.1016/j.compbiomed.2022.105812

    arXiv 2022

  7. [6]

    Breast cancer: patho- genesis and treatments

    Xiong X, Zheng LW, Ding Y, et al. Breast cancer: patho- genesis and treatments. Sig Transduct Target Ther. 2025;10:49. https://doi.org/10.1038/s41392-024-02108-4

    work page doi:10.1038/s41392-024-02108-4 2025

  8. [7]

    Spiking Neural Network

    Ghosh-Dastidar S, Adeli H. Spiking Neural Network. Int J Neural Syst. 2009;19:295–308

    work page 2009

Show all 60 references
  1. [8]

    Neural networks for computer-aided diagnosis in medicine: A review

    Lin D, Vasilakos AV, Tang Y, Yao Y. Neural networks for computer-aided diagnosis in medicine: A review. Neurocomputing. 2016;216:700–708

    2016

  2. [9]

    Applicable artifi- cial intelligence for brain disease: A survey

    Huang C, Wang J, Wang SH, Zhang YD. Applicable artifi- cial intelligence for brain disease: A survey. Neurocomputing. 2022;504C:223–239

    2022

  3. [10]

    Backpropagation- Based Learning Techniques for Deep Spiking Neural Networks: A Sur- vey

    Dampfhoffer M, Mesquida T, Valentian A, Anghel L. Backpropagation- Based Learning Techniques for Deep Spiking Neural Networks: A Sur- vey. IEEE Trans Neural Netw Learn Syst. 2023

    2023

  4. [11]

    Cluster analysis of breast cancer data using Genetic Algorithm and Spiking Neural Networks

    Sangeetha V, Preethi J. Cluster analysis of breast cancer data using Genetic Algorithm and Spiking Neural Networks. In: Proceedings of the 2015 IEEE 9th International Con- ference on Intelligent Systems and Control (ISCO); 2015:1–5. https://doi.org/10.1109/ISCO.2015.7282345. 20

    2015

  5. [12]

    Training Spiking Neural Networks with Metaheuristic Algorithms

    Javanshir A, Nguyen TT, Mahmud MAP, Kouzani AZ. Training Spiking Neural Networks with Metaheuristic Algorithms. Appl Sci. 2023;13:4809. https://doi.org/10.3390/app13084809

    2023 doi

  6. [13]

    Leveraging stochasticity in memris- tive synapses for efficient and reliable neuromorphic systems

    Das H, Patel KP, Febbo RD, et al. Leveraging stochasticity in memris- tive synapses for efficient and reliable neuromorphic systems. npj Un- conventional Comput. 2025;2:3. https://doi.org/10.1038/s44335-024- 00017-x

    2025 doi

  7. [14]

    Breast Cancer Recognition Using Saliency-Based Spik- ing Neural Network

    Fu Q, Dong H. Breast Cancer Recognition Using Saliency-Based Spik- ing Neural Network. Wirel Commun Mob Comput. 2022;2022:8369368. https://doi.org/10.1155/2022/8369368

    2022 doi

  8. [15]

    Effective multispike learning in a spiking neural network with a new temporal feedback backpropaga- tion for breast cancer detection

    Heidarian M, Karimi G, Payandeh M. Effective multispike learning in a spiking neural network with a new temporal feedback backpropaga- tion for breast cancer detection. Expert Syst Appl. 2024;252B:124010. https://doi.org/10.1016/j.eswa.2024.124010

    2024

  9. [16]

    Spiking neural network reinforcement learning method based on temporal coding and STDP

    Sboev A, Vlasov D, Rybka R, Serenko A. Spiking neural network reinforcement learning method based on temporal coding and STDP. Procedia Comput Sci. 2018;145:458–463. https://doi.org/10.1016/j.procs.2018.11.107

    2018 doi

  10. [17]

    Multi-stages attention breast cancer classification based on nonlinear spiking neural P neurons with autapses

    Tang Y, Yang B, Peng H, Luo X. Multi-stages attention breast cancer classification based on nonlinear spiking neural P neurons with autapses. Eng Appl Artif Intell. 2025;142:109869. https://doi.org/10.1016/j.engappai.2024.109869

    2025

  11. [18]

    SWAT: A Spiking Neural Network Training Algorithm for Classification Problems

    Wade JJ, McDaid LJ, Santos JA, Sayers HM. SWAT: A Spiking Neural Network Training Algorithm for Classification Problems. IEEE Trans Neural Netw. 2010;21(11):1817–1830. https://doi.org/10.1109/TNN.2010.2074212

    2010

  12. [19]

    Nonlinear Spiking Neural Systems With Au- tapses for Predicting Chaotic Time Series

    Liu Q, Peng H, Long L, Wang J, Yang Q, Pérez-Jiménez MJ, Orellana-Martín D. Nonlinear Spiking Neural Systems With Au- tapses for Predicting Chaotic Time Series. IEEE Trans Cybern. 2024;54(3):1841–1853. https://doi.org/10.1109/TCYB.2023.3270873

    2024

  13. [20]

    DoB-SNN: A New Neuron Assembly-Inspired Spiking Neural Network for 21 Pattern Classification

    Saranirad V, McGinnity TM, Dora S, Coyle D. DoB-SNN: A New Neuron Assembly-Inspired Spiking Neural Network for 21 Pattern Classification. In: Proceedings of the 2021 Interna- tional Joint Conference on Neural Networks (IJCNN); 2021:1–6. https://doi.org/10.1109/IJCNN52387.2021.9534283

    2021

  14. [21]

    On the complexity of finite sequences

    Ziv J, Lempel A. On the complexity of finite sequences. IEEE Trans Inf Theory. 1976;22:75–81

    1976

  15. [22]

    Theoretical Neuroscience: Computational and Mathematical Modeling of Neural Systems

    Dayan P, Abbott LF. Theoretical Neuroscience: Computational and Mathematical Modeling of Neural Systems. 2000

    2000

  16. [23]

    A leaky integrate-and- fire model with adaptation for the generation of a spike train

    Buonocore A, Caputo L, Pirozzi E, Carfora MF. A leaky integrate-and- fire model with adaptation for the generation of a spike train. Math Biosci Eng. 2015;13(3):483–493

    2015

  17. [24]

    Leaky Integrate and Fire Neuron by Charge-Discharge Dynamics in Floating-Body MOSFET

    Dutta S, Kumar V, Shukla A, et al. Leaky Integrate and Fire Neuron by Charge-Discharge Dynamics in Floating-Body MOSFET. Sci Rep. 2017;7:8257

    2017

  18. [25]

    A generalized leaky integrate-and-fire neuron model with fast implementation method

    Wang Z, Guo L, Adjouadi M. A generalized leaky integrate-and-fire neuron model with fast implementation method. Int J Neural Syst. 2014;24(05):1440004

    2014

  19. [26]

    Glif: A unified gated leaky integrate-and- fire neuron for spiking neural networks

    Yao X, Li F, Mo Z, Cheng J. Glif: A unified gated leaky integrate-and- fire neuron for spiking neural networks. Adv Neural Inf Process Syst. 2022;35:32160–32171

    2022

  20. [27]

    Simulating leaky integrate-and-fire neuron with integers

    Vidybida AK. Simulating leaky integrate-and-fire neuron with integers. Math Comput Simul. 2019;159:154–160

    2019

  21. [28]

    CLIF: Complementary leaky integrate-and-fire Neuron for spiking neural networks

    Huang Y, Lin X, Ren H, et al. CLIF: Complementary leaky integrate-and-fire Neuron for spiking neural networks. arXiv preprint arXiv:2402.04663. 2024

    2024 arXiv

  22. [29]

    Energy efficient neural codes

    Levy WB, Baxter RA. Energy efficient neural codes. Neural Comput. 1996;8(3):531–543. https://doi.org/10.1162/neco.1996.8.3.531

    1996 doi

  23. [30]

    Optimizing information pro- cessing in brain-inspired neural networks

    Paprocki B, Pregowska A, Szczepanski J. Optimizing information pro- cessing in brain-inspired neural networks. Bull Pol Acad Sci Tech Sci. 2020;68(2). 22

    2020

  24. [31]

    Computational Intelligence and optimization techniques in commu- nications and control

    Klamka J, Szczepański J, Węgrzyn-Wolska KM, Rojek I, Prokopowicz P. Computational Intelligence and optimization techniques in commu- nications and control. Bull Pol Acad Sci Tech Sci. 2020:181–184

    2020

  25. [32]

    Does Adding of Neurons to the Network Layer Lead to Increased Transmission Efficiency? IEEE Access

    Paprocki B, Pregowska A, Szczepanski J. Does Adding of Neurons to the Network Layer Lead to Increased Transmission Efficiency? IEEE Access. 2024;12:42701–42709

    2024

  26. [33]

    Stable Spike-Timing Dependent Plasticity Rule for Multilayer Unsupervised and Su- pervised Learning

    Shrestha A, Ahmed K, Wang Y, Qiu Q. Stable Spike-Timing Dependent Plasticity Rule for Multilayer Unsupervised and Su- pervised Learning. In: Proceedings of the 2017 International Joint Conference on Neural Networks (IJCNN); 2017:1999–2006. https://doi.org/10.1109/IJCNN.2017.7966096

    2017

  27. [34]

    Su- pervised Hebbian Learning

    Alemanno F, Aquaro M, Kanter I, Barra A, Agliari E. Su- pervised Hebbian Learning. Europhys Lett. 2023;141:11001. https://doi.org/10.1209/0295-5075/aca55f

    2023 doi

  28. [35]

    Exploring Neuromorphic Com- puting Based on Spiking Neural Networks: Algorithms to Hardware

    Rathi N, Chakraborty I, Kosta A, et al. Exploring Neuromorphic Com- puting Based on Spiking Neural Networks: Algorithms to Hardware. ACM Comput Surv. 2023;55:243

    2023

  29. [36]

    In: Proceedings of the Image Analysis and Processing—ICIAP 2019; Sept 9–13, 2019; Trento, Italy

    AmatoG,CarraraF,FalchiF,GennaroC,LaganiG.HebbianLearning Meets Deep Convolutional Neural Networks. In: Proceedings of the Image Analysis and Processing—ICIAP 2019; Sept 9–13, 2019; Trento, Italy. p. 9–13

    2019

  30. [37]

    A history of spike- timing-dependent plasticity

    Markram H, Gerstner W, Sjöström PJ. A history of spike- timing-dependent plasticity. Front Synaptic Neurosci. 2011;3:4. https://doi.org/10.3389/fnsyn.2011.00004

    2011 arXiv

  31. [38]

    MerollaPA,ArthurJV,Alvarez-IcazaR,etal.Amillionspiking-neuron integratedcircuitwithascalablecommunicationnetworkandinterface. Science. 2014;345:668–673. https://doi.org/10.1126/science.1254642

    2014 doi

  32. [39]

    Characterization of Gen- eralizability of Spike Timing Dependent Plasticity Trained Spiking Neural Networks

    Chakraborty B, Mukhopadhyay S. Characterization of Gen- eralizability of Spike Timing Dependent Plasticity Trained Spiking Neural Networks. Front Neurosci. 2021;15:695357. https://doi.org/10.3389/fnins.2021.695357. 23

    2021

  33. [40]

    Spiking Neural Net- works and Bio-Inspired Supervised Deep Learning: A Survey

    Lagani G, Falchi F, Gennaro C, Amato G. Spiking Neural Net- works and Bio-Inspired Supervised Deep Learning: A Survey. arXiv. 2023;arXiv:2307.16235

    2023 arXiv

  34. [41]

    Digital Twins: The New Frontier for Personalized Medicine? Appl Sci

    Cellina M, Cè M, Alì M, et al. Digital Twins: The New Frontier for Personalized Medicine? Appl Sci. 2023;13:7940

    2023

  35. [42]

    Artificial Neural Net- work (ANN) to Spiking Neural Network (SNN) Converters Based on Diffusive Memristors

    Midya R, Wang Z, Asapu S, et al. Artificial Neural Net- work (ANN) to Spiking Neural Network (SNN) Converters Based on Diffusive Memristors. Adv Electron Mater. 2019;5:1900060. https://doi.org/10.1002/aelm.201900060

    2019 doi

  36. [43]

    CS- QCFS: Bridging the performance gap in ultra-low latency spiking neural networks

    Yang H, Yang S, Zhang L, Dou H, Shen F, Zhao J. CS- QCFS: Bridging the performance gap in ultra-low latency spiking neural networks. Neural Netw. 2025;184:107076. https://doi.org/10.1016/j.neunet.2024.107076

    2025

  37. [44]

    Machine learning-based multiscale framework for mechanical behavior of nano-crystalline structures

    Khoei AR, Seddighian MR, Rezaei Sameti A. Machine learning-based multiscale framework for mechanical behavior of nano-crystalline structures. Int J Mech Sci. 2024;265:108897. https://doi.org/10.1016/j.ijmecsci.2023.108897

    2024

  38. [45]

    Toward Evolving Neural Networks using Bio- Inspired Algorithms

    Conforth M, Meng Y. Toward Evolving Neural Networks using Bio- Inspired Algorithms. In: Proceedings of the International Conference on Artificial Intelligence (IC-AI); 2008. p. 413–419

    2008

  39. [46]

    Front Neurosci

    BianchiS, Muñoz-MartinI,IelminiD.Bio-inspiredtechniquesin afully digital approach for lifelong learning. Front Neurosci. 2020;14:379

    2020

  40. [47]

    Effective Active Learning Method for Spiking Neural Networks

    Xie X, Yu B, Liu G, Zhan Q, Tang H. Effective Active Learning Method for Spiking Neural Networks. IEEE Trans Neural Netw Learn Syst. 2024;35:12373–12382

    2024

  41. [48]

    Learning representations by back- propagating errors

    Rumelhart D, Hinton G, Williams R. Learning representations by back- propagating errors. Nature. 1986;323:533–536

    1986

  42. [49]

    Back-propagation algorithm with variable adaptive momentum

    Hameed AA, Karlik B, Salman MS. Back-propagation algorithm with variable adaptive momentum. Knowl Based Syst. 2016;114:79–87

    2016

  43. [50]

    The kinematics of the spike trains

    Wójcik D. The kinematics of the spike trains. Acta Phys Pol B. 2018;49:2127–2138. 24

    2018

  44. [51]

    Analysis, classification, and coding of multielectrode spike trains with hidden Markov models

    Radons G, Becker JD, Dülfer B, Krüger J. Analysis, classification, and coding of multielectrode spike trains with hidden Markov models. Biol Cybern. 1994;71:359–373

    1994

  45. [52]

    Comprehensive Overview of Backpropagation Algorithm for Digital Image Denoising

    Singh A, Kushwaha S, Alarfaj M, Singh M. Comprehensive Overview of Backpropagation Algorithm for Digital Image Denoising. Electronics. 2022;11:1590

    2022

  46. [53]

    Back propagation artificial neural network for diagnosis of heart disease

    Kaur J, Khehra BS, Singh A. Back propagation artificial neural network for diagnosis of heart disease. J Reliab Intell Environ. 2023;9:57–85

    2023

  47. [54]

    The tempotron: a neuron that learns spike timing-based decisions

    Gutig R, Sompolinsky H. The tempotron: a neuron that learns spike timing-based decisions. Nat Neurosci. 2006;9(3):420–428

    2006

  48. [55]

    A gradient learning rule for the tempotron

    Urbanczik R, Senn W. A gradient learning rule for the tempotron. Neural Comput. 2009;21:340–343

    2009

  49. [56]

    A brain-inspired spiking neural network model with tem- poral encoding and learning

    Yu W, et al. A brain-inspired spiking neural network model with tem- poral encoding and learning. Neurocomputing. 2014;138:3–13

    2014

  50. [57]

    An STDP training algorithm for a spiking neural network with dynamic threshold neurons

    Strain TJ, McDaid LJ, McGinnity TM, Maguire LP, Sayers HM. An STDP training algorithm for a spiking neural network with dynamic threshold neurons. Int J Neural Syst. 2010;20(06):463–480

    2010

  51. [58]

    Solving a classifica- tion task by spiking neural network with STDP based on rate and temporal input encoding

    Sboev A, Serenko A, Rybka R, Vlasov D. Solving a classifica- tion task by spiking neural network with STDP based on rate and temporal input encoding. Math Meth Appl Sci. 2020;43:7802–7814. https://doi.org/10.1002/mma.6241

    2020 doi

  52. [59]

    Devel- opment of a Self-Regulating Evolving Spiking Neural Network for classification problem

    Dora S, Subramanian K, Suresh S, Sundararajan N. Devel- opment of a Self-Regulating Evolving Spiking Neural Network for classification problem. Neurocomputing. 2016;171:1216–1229. https://doi.org/10.1016/j.neucom.2015.07.086

    2016 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.