Pith. sign in

REVIEW 4 major objections 4 minor 16 references

Phase State Space Models: Parallel, Surrogate-Free Training of Spiking Networks

T0 review · 4 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read The paper proposes the Phase State Space Model (φ-SSM), which recasts resonate-and-fire spiking networks so continuous, discrete, and convolutional execution views compute identical potentials, enabling parallel, surrogate-free training.

desk verdict A genuinely interesting phase-domain formulation of R&F SSMs, but the central three-view equivalence is unproven as written due to an off-by-one convolution index, and the empirical section is too thin to verify anything. read the letter →

arxiv 2608.07754 v1 pith:VVMZJH7U submitted 2026-08-07 cs.NE

Wilkie Olin-Ammentorp This is my paper · ORCID
classification cs.NE
keywords spikingneuralnetworksstate-spacemodelsresonate-and-firephaseencodingrelativehyperdimensionalcomputingparalleltrainingsurrogate-freelearning
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

This paper aims to give resonate-and-fire (R&F) spiking networks—neurons whose complex-valued potentials oscillate and emit spikes—the same parallel-training advantage that state-space models bring to recurrent networks. The author's central move is to lock layers of R&F neurons to a shared resonant frequency, making the relative phases between neurons constant over time and allowing spikes to be read as phase values rather than as discontinuous events. On this basis the paper defines the Phase State Space Model ($\varphi$-SSM), claiming that the same network can be executed as a continuous differential equation, a discrete recurrence, or a convolutional/FFT kernel, with outputs identical within numerical precision. If true, spiking networks can be trained in parallel without surrogate gradient kernels while keeping recurrent, biologically plausible inference. The paper demonstrates the idea by building a network that combines a short-time Fourier transform adapter, attention, and a codebook readout, all communicating via relative phases.

What carries the argument

The load-bearing object is the relative-phase representation of a frequency-locked R&F layer: $\theta_{\mathrm{rel}}(u,\omega,t)=\angle(u e^{-i\omega t})$, which removes the global rotation of the resonant frequency and leaves a time-invariant phase code. The mechanism that connects the three execution views is the phase-to-time mapping $\delta t(\theta,k)=(2\pi/\Im(k))(1/2-\theta/2)$, which converts a spiking input's decoded phase into the exact offset used in $U'(T,\theta,k)=e^{k\delta t}$. Substituting this into the linear R&F update gives $U[n+1]=e^{kT}U[n]+B\,U'(T,\theta,k)$, and unrolling that recurrence yields the convolution $U[n]=(K*H)[n]$ with kernel $K[n]=e^{k n T}$, which the FFT evaluates in parallel. The same discrete-time kernel also supplies the continuous and recurrent views, since all three are algebraic rearrangements of the same exponential transition.

What would settle it

Run the continuous-time R&F equations (Eqn. 6) with a high-accuracy ODE solver using finite-width current pulses or two spikes within one resonant period, then compare that solution to the discrete recurrence and the FFT convolution; any disagreement beyond numerical tolerance would show the three-view equivalence holds only for ideal single Dirac spikes per period, not for the general inputs the paper claims to support.

Watch

Extended reading notes

Core claim

The central claim is that the phase of an R&F neuron's complex potential is a sufficient output to define a spiking state-space model, and that this phase-based model has three exactly equivalent execution views. By fixing the resonant frequency $\omega$ of a layer, the relative phase $\theta_{\mathrm{rel}}(u,\omega,t)=\angle(u e^{-i\omega t})$ is invariant through time, so a spike emitted at phase $\theta$ can be converted into a contribution at the next sample time $T=2\pi/\omega$ via $U'(T,\theta,k)=e^{k\delta t(\theta,k)}$, with $\delta t(\theta,k)=(2\pi/\Im(k))(1/2-\theta/2)$. This yields the discrete recurrence $U[n+1]=e^{kT}U[n]+B\,U'(T,\theta,k)$, which unrolls into a causal convolution $U[n]=(K*H)[n]$ with kernel $K[n]=e^{k n T}$, computable in parallel by FFT. The paper reports that the continuous, discrete, and FFT implementations give outputs of $U(t)$ that agree within numerical precision, and uses the convolutional view to train networks on a copying task and sequential FashionMNIST. In the author's framing, this is an R&F-native alternative to surrogate-gradient training that retains the SSM's triple view and connects naturally to hyperdimensional computing.

Load-bearing premise

The discrete and convolutional kernels assume each input is a Dirac-delta spike carrying exactly one decoded phase per neuron per resonant period; if inputs are real-valued currents, finite-width pulses, or contain multiple spikes per period, the kernelized views are not proven to match the continuous R&F dynamics.

Editorial extensions

If this is right

  • The same network can be executed as a continuous differential system, a discrete recurrence, or an FFT convolution, with outputs matching to numerical precision, so one network serves both real-time hardware and parallel GPU training.
  • Gradients flow through relative phases at every discrete time step for every neuron, including silent ones, without surrogate spiking kernels.
  • Spikes in a $\varphi$-SSM layer can be interpreted as sparse phase communication, because relative phases of frequency-locked R&F neurons are invariant through time.
  • The $\varphi$-SSM recurrence is equivalent to hyperdimensional computing operations (bundling, binding, similarity), linking recurrent SSM training tools to HD computing techniques.
  • The inclusion of an STFT adapter, attention, and codebook readout improves performance on the copying task and sequential FashionMNIST relative to ablations.

Reading between the lines

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

  • Editorial inference: if the three-view equivalence is exact for Dirac-delta inputs, the same construction should transfer to other linear oscillatory neuron models whose envelopes follow a complex exponential, giving a general recipe for parallel training of resonant recurrent networks beyond resonate-and-fire neurons.
  • Editorial inference: the phase-based spike code suggests a testable hardware prediction—an analog physical oscillator implementing the continuous view should match the FFT training view as long as the resonator remains linear and frequency-locked.
  • Editorial inference: because the FFT view computes all timesteps in $O(L \log L)$ work per layer, the $\varphi$-SSM could make very long spiking sequences trainable end-to-end, a regime where backpropagation-through-time is impractical.
  • Editorial inference: treating spikes as phase symbols turns the spiking network into a differentiable vector-symbolic architecture, so techniques from memory capacity and error correction in HD computing may apply directly to the trained representations.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces the Phase State Space Model (phi-SSM), a reformulation of resonate-and-fire (R&F) neurons as a complex-valued linear state-space model with a diagonal transition matrix. It derives three execution modes: a continuous differential view, a discrete recurrence sampled at the resonant period, and a convolutional/FFT view. The paper further interprets the discrete update in terms of hyperdimensional (HD) computing operations (bundle, bind, similarity) and presents a sequence-processing network combining an STFT adapter, an attention-like module, and a codebook readout. The central claimed contribution is that the three views produce identical neural potentials, enabling parallel, surrogate-free training of R&F networks, and the paper reports qualitative ablation results on a copy task and sequential FashionMNIST.

Significance. If the three-view equivalence were correct, the paper would provide a useful bridge between spiking R&F networks and parallel SSM training, avoiding surrogate gradients while retaining biological features of R&F neurons. The derivation is transparent and builds on a standard linear-systems core, and the open-source release is a practical strength. However, the central equivalence as written contains a causal-index error in the convolution, and the HD interpretation discards magnitude information without proof. The empirical section reports no quantitative results, so the practical claims are not yet substantiated. The underlying idea is promising and likely repairable, but the present manuscript does not establish its main claim.

major comments (4)
  1. [2.4, Eq. (14)] Equation (14) is not the unrolled form of Eq. (13). With U[0]=0, Eq. (13) gives U[1]=B U'(theta_0) and U[2]=e^{kT} B U'(theta_0) + B U'(theta_1), i.e. U[n] = sum_{j=0}^{n-1} e^{k(n-j)T} B U'(theta_j). Equation (14) instead sums j=0..n, which adds the non-causal term B U'(theta_n) to U[n]. Because the FFT/parallel training view is built directly on Eq. (14), the claimed exact equality of the discrete and convolutional views is not established as written. Please correct the convolution limits (or set K[0]=0) and re-derive the kernel accordingly.
  2. [2.5, Eq. (20)] Equation (20) presents the HD bundle/bind form as an equality, but the underlying complex recurrence in demodulated coordinates is V[n+1] = e^{lambda T} |U[n]| e^{i pi Y[n]} + |B| e^{lambda delta t} e^{i pi (1-theta)}, where the two terms generally have different magnitudes. The phase of this weighted sum is not equal to bundle(Y[n], 1-theta) as defined in Eq. (17), which operates on unit phasors. The phrase 'subject to magnitude decay' does not remove the magnitude weighting; the equality holds only if magnitudes are equal or the bundle operation is redefined to include magnitudes. Please either prove the equality under an explicit normalization condition or reframe Eq. (20) as an analogy rather than an equivalence.
  3. [3.4 and Fig. 1(d-f)] The paper asserts numerical agreement 'within numerical precision' and improved performance from the STFT adapter and attention module, but the submitted text contains no figures and no quantitative results: no accuracies, losses, error bars, or task details are reported. Consequently the demonstration and the three-view equivalence claim are unverifiable. Please include the actual figure content or tabulated results, with task definitions and hyperparameters.
  4. [2.4, input assumption] The exact equivalence among the three views is derived for an input that is a sequence of Dirac-delta spikes, at most one per neuron per resonant period, with spike timing mapped from phase by Eq. (11). This restriction is not stated as a limitation. For general real-valued inputs the continuous-time dynamics are not represented by the discrete/convolutional kernels unless the ZOH adapter of Sec. 3.1 is used. Please state this input assumption explicitly and discuss its scope, since the abstract claims compatibility with both real and spiking inputs.
minor comments (4)
  1. [2.4, Eq. (11)] The boundary convention is unclear: with theta=1, Eq. (11) gives delta t=0 and t_spike=T, which coincides with the next sample point; please clarify how spikes exactly at sample boundaries are handled.
  2. [3.2, Eq. (24)] The attention scores A_ij are not normalized across j, since the softmax is deliberately avoided; calling these 'attention scores' without noting that the sum over j need not be 1 may mislead readers.
  3. [References] Several references are incomplete: [5] lacks a year and venue, and [13] and [14] would benefit from full bibliographic details (venue, publication status).
  4. [Fig. 1(c) caption] The caption says the convolutional/FFT view 'requires the storage of all timesteps previously solved iteratively,' but in parallel training all timesteps are computed simultaneously; the wording should distinguish inference-time storage from training-time computation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the phi-SSM three-view derivation is self-contained and the central mathematical claims reduce only to standard linear-system manipulations, not to fitted predictions.

full rationale

The phi-SSM's central derivation is self-contained. The continuous view is the R&F ODE (Eq. 6); the discrete view (Eq. 13) is obtained by solving that ODE between sample points and adding the exact contribution e^{k*delta_t} of a Dirac spike (Eqs. 11-12); the convolutional/FFT view (Eqs. 14-15) is the unrolled form of Eq. 13. Each step is an algebraic transformation of the preceding one, so the three-view equivalence is a mathematical identity, not an empirical prediction fitted to data. The relative-phase output (Eqs. 9-10) is defined directly from the complex potential and the reference oscillator e^{-i omega t}; the citation to [13] for the invariance is unnecessary to the algebra and does not import an unverified premise. Similarly, the HD interpretation (Eq. 20) is explicitly presented as an interpretation, with the HD operators defined in Eqs. 17-19; identifying the recurrence as bundle/bind is a renaming that does not substitute for a derivation. The attention/codebook modules use those defined operators; their numerical benefits are reported as benchmark results, not as predictions from fitted parameters. The only notable issue in the derivation chain is an apparent off-by-one in Eq. 14 (summing j=0..n with K[0]=1 while the unrolled recurrence from U[0]=0 contains terms only up to j=n-1), which would affect the claimed FFT/recurrent equivalence; but that is an internal correctness defect, not a circularity, since the convolutional view is not being used as its own evidence. No step in the paper reduces to its own inputs by definition, and no fitted quantity is renamed as a prediction.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The framework rests on standard linear systems theory and the cited R&F model. The main ad hoc costs are the phase-to-spike-time encoding (one Dirac spike per period) and the magnitude-free HD bundling approximation. No new physical entities are introduced.

free parameters (3)
  • per-layer resonant frequency omega = not stated
    The phi-SSM layers lock omega to define relative phase; this value is chosen by hand, not derived. The STFT adapter makes its own resonant frequencies trainable.
  • leakage/damping lambda = not stated
    k = lambda + i*omega, with lambda < 0 controlling potential decay; it is a hyperparameter selected by the user, not derived from the problem.
  • attention scale beta = learned
    A learned scalar in Eq. 24 that scales attention similarities; it is fitted to data during training.
assumptions (5)
  • standard math Linear time-invariant systems: the solution of dU/dt = k*U + w*I is obtained via exponential kernels and impulse-invariant discretization.
    Used to derive Eqs. 11-15; this is standard control theory and is not proven in the paper.
  • domain assumption The resonate-and-fire neuron model with complex potential and the firing rule from Izhikevich [7] is correct.
    The entire framework is built on this model, taken as given from the cited reference.
  • domain assumption Fixing the resonant frequency makes relative phase invariant during free evolution.
    The paper cites its own prior work [13] for this property; it is true only for the unforced homogeneous solution, not for driven neurons.
  • ad hoc to paper At most one Dirac-delta spike per neuron per resonant period, with spike timing bijectively mapped from the relative phase theta.
    Needed for Eqs. 11-12; this is not stated as a limitation and is not tested against realistic spikes.
  • ad hoc to paper The HD bundle/bind operation on phases alone (Eq. 20) is equivalent to the complex linear recurrence without magnitude normalization.
    Bundle is defined on unit vectors, but U amplitudes are arbitrary; equivalence holds only under an unstated assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Phase State Space Models: Parallel, Surrogate-Free Training of Spiking Networks." pith.science (2026). https://pith.science/paper/VVMZJH7U

@misc{pith2026260807754,
  author       = {Pith},
  title        = {Pith review of: Phase State Space Models: Parallel, Surrogate-Free Training of Spiking Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VVMZJH7U}},
  note         = {Machine review of arXiv:2608.07754}
}
read the original abstract

State-space models (SSMs) provide a powerful theoretical framework to enable parallel training of recurrent networks. We expand on previous work adapting SSMs to spiking models to provide a novel interpretation of resonate-and-fire (R\&F) neural networks which is compatible both with real and spiking inputs, parallel and recurrent execution, has clear connections to hyperdimensional (HD) computing, and maintains biologically-realistic features. We demonstrate an implementation of this approach which integrates an STFT, recurrent memory, and attentional features within a single spike-compatible network.

Figures

Figures reproduced from arXiv: 2608.07754 by the authors.

Figure 1
Figure 1. The 𝜑-SSM provides three equivalent views which can be used to calculate the neural potentials of R&F neurons through time. The continuous, differential view (a) can be implemented directly via physical oscillators or simulated via differential solvers. The discrete view (b) calculates the same updates at regularly-spaced intervals at multiples of 𝑇 , the resonant period of the neural layer, and can be implemented o… view at source ↗
Figure 2
Figure 2. We demonstrate the efficacy of combining the [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 9 canonical work pages

  1. [1]

    Malyaban Bal and Abhronil Sengupta. 2025. P-SpikeSSM: Harnessing Probabilistic Spiking State Space Models for Long-Range Dependency Tasks. https://doi.org/10.48550/arXiv.2406.02923 arXiv:2406.02923 [cs]

    arXiv doi:10.48550/arxiv.2406.02923 2025

  2. [2]

    Atilim Gunes Baydin, Barak A Pearlmutter, Alexey Andreyevich Radul, and Jeffrey Mark Siskind. 2018. Automatic Differentiation in Machine Learning: A Survey.Journal of machine learning research18, 153 (2018), 1–43

    work page 2018

  3. [3]

    Guillaume Bellec, Franz Scherr, Elias Hajek, Darjan Salaj, Anand Subramoney, Robert Legenstein, and Wolfgang Maass. 2019. Eligibility traces provide a data-inspired alternative to backpropagation through time.Arxiv(2019). https://openreview.net/forum?id=SkxJ4QKIIS

    work page 2019

  4. [4]

    Yu Du, Xu Liu, and Yansong Chua. 2024. Spiking Structured State Space Model for Monaural Speech Enhancement. https://doi.org/10.48550/arXiv. 2309.03641 arXiv:2309.03641 [cs]. Manuscript submitted to ACM 10 Olin-Ammentorp

    work page Pith review arXiv doi:10.48550/arxiv.2309.03641 2024

  5. [5]

    Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. [n. d.]. Combining Recurrent, Convolutional, and Continuous-time Models with Linear State-Space Layers. ([n. d.])

    work page

  6. [6]

    Huber, Jules Lecomte, Borislav Polovnikov, and Axel von Arnim

    Thomas E. Huber, Jules Lecomte, Borislav Polovnikov, and Axel von Arnim. 2025. Scaling Up Resonate-and-Fire Networks for Fast Deep Learning. https://doi.org/10.48550/arXiv.2504.00719 arXiv:2504.00719 [cs]

    arXiv doi:10.48550/arxiv.2504.00719 2025

  7. [7]

    Izhikevich

    Eugene M. Izhikevich. 2001. Resonate-and-fire neurons.Neural Networks14, 6-7 (2001), 883–894. https://doi.org/10.1016/S0893-6080(01)00078-8

    doi:10.1016/s0893-6080(01)00078-8 2001

  8. [8]

    Samy Jelassi, David Brandfonbrener, Sham M Kakade, and Eran Malach. [n. d.]. Repeat After Me: Transformers are Better than State Space Models at Copying Transformers are Better than State Space Models at Copying. ([n. d.])

    work page

Show all 16 references
  1. [9]

    Rachkovskij, Evgeny Osipov, and Abbas Rahimi

    Denis Kleyko, Dmitri A. Rachkovskij, Evgeny Osipov, and Abbas Rahimi. 2021. A Survey on Hyperdimensional Computing aka Vector Symbolic Architectures, Part II: Applications, Cognitive Models, and Challenges. (2021), 1–36. http://arxiv.org/abs/2112.15424 arXiv: 2112.15424

    2021 arXiv

  2. [10]

    Wolfgang Klimesch. 2018. The frequency architecture of brain and brain body oscillations: an analysis.European Journal of Neuroscience48, 7 (Oct. 2018), 2431–2453. https://doi.org/10.1111/ejn.14192

    2018 doi

  3. [11]

    Axel Laborieux and Friedemann Zenke. 2022. Holomorphic Equilibrium Propagation Computes Exact Gradients Through Finite Size Oscillations. http://arxiv.org/abs/2209.00530 arXiv:2209.00530 [cs]

    2022 arXiv

  4. [12]

    Neftci, Hesham Mostafa, and Friedemann Zenke

    Emre O. Neftci, Hesham Mostafa, and Friedemann Zenke. 2019. Surrogate Gradient Learning in Spiking Neural Networks. (2019), 1–25. http: //arxiv.org/abs/1901.09948 arXiv: 1901.09948

    2019 arXiv

  5. [13]

    Wilkie Olin-Ammentorp. 2023. Hyperdimensional Computing Provides a Programming Paradigm for Oscillatory Systems. http://arxiv.org/abs/ 2312.11783 arXiv:2312.11783 [math]

    2023 arXiv

  6. [14]

    Wilkie Olin-Ammentorp and Maxim Bazhenov. 2022. Residual and Attentional Architectures for Vector-Symbols. arXiv: 2207.08953

    2022 arXiv

  7. [15]

    Paxon Frady, Daniel Ben Dayan Rubin, Sophia Sanborn, Sumit Bam Shrestha, Friedrich T

    Garrick Orchard, E. Paxon Frady, Daniel Ben Dayan Rubin, Sophia Sanborn, Sumit Bam Shrestha, Friedrich T. Sommer, and Mike Davies. 2021. Efficient Neuromorphic Signal Processing with Loihi 2. http://arxiv.org/abs/2111.03746 arXiv:2111.03746 [cs]

    2021 arXiv

  8. [16]

    PhasorNetworks.jl

    Christopher Rackauckas, Yingbo Ma, Julius Martensen, Collin Warner, Kirill Zubov, Rohit Supekar, Dominic Skinner, Ali Ramadhan, and Alan Edelman. 2021. Universal Differential Equations for Scientific Machine Learning. https://doi.org/10.48550/arXiv.2001.04385 arXiv:2001.04385 ...

    2021 doi arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.