Pith. sign in

REVIEW 3 major objections 5 minor 28 references

Filtered interspike interval encoding by class II neurons

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A class II neuron, exemplified by the FitzHugh–Nagumo model, simultaneously filters time-varying inputs by frequency and represents the stimulus waveform in its interspike intervals.

desk verdict Plausible model-based evidence that class II neurons can do frequency-filtered ISI encoding, but the NPE metric does not yet prove the intervals carry stimulus information. read the letter →

arxiv 2411.14692 v1 pith:JGPC3G4P submitted 2024-11-22 cond-mat.dis-nn

Naoki Masuda , Kazuyuki Aihara This is my paper · ORCID
classification cond-mat.dis-nn PACS 87.19.La05.45.Xt87.18.Sn
keywords classIIneuronFitzHugh-Nagumomodelinterspikeintervalfrequencyfilteringamplitudemodulationresonancenormalizedpredictionerrorchaoticinput
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 claims that a class II neuron—specifically the FitzHugh–Nagumo model—can simultaneously act as a frequency filter and as a signal encoder. The neuron transmits information about a time-varying stimulus only when the stimulus's dominant frequency is close to the neuron's intrinsic firing frequency; outside that band, the interspike intervals are governed by the neuron's own rhythm and carry little input information. Inside the band, the intervals vary in a deterministic way that tracks the slow temporal envelope of the input, so a downstream reader could recover the stimulus waveform from the spike train. The paper demonstrates this by computing a normalized prediction error of the interspike intervals for chaotic and quasi-periodic inputs and by showing that the error drops near resonance across three different intrinsic frequencies and in the presence of noise.

What carries the argument

The load-bearing objects are the FitzHugh–Nagumo model, a two-variable model of a class II neuron (a neuron whose firing frequency is set mostly by its internal dynamics rather than by input strength), and the normalized prediction error $\mathrm{NPE}(1)$, obtained by delay-embedding the interspike intervals and predicting each next interval from its nearest neighbors in the embedding. The small parameter $\alpha$ sets the neuron's intrinsic oscillation frequency, and $\mathrm{NPE}(1)$ measures how deterministically the intervals follow past intervals; the paper treats a low value as evidence that the intervals encode the input. The comparison of $\mathrm{NPE}(1)$ against the input frequency, the actual firing frequency, and the intrinsic frequency range of the neuron is what carries the filtering claim.

What would settle it

Measure an information-theoretic quantity, such as mutual information or transfer entropy, between the input $S(t)$ and the decoded interspike intervals over the same frequency sweep; if these quantities show no peak near the resonance where $\mathrm{NPE}(1)$ drops, the claim that the intervals encode the input would be falsified. The same sweep run with a constant input or with white noise sharing the same dominant frequency would also separate intrinsic predictability from genuine stimulus encoding.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that a FitzHugh–Nagumo class II neuron works as an amplitude-modulation processor: the oscillatory component of the input acts as a carrier, and the input's amplitude profile is encoded in the interspike intervals. Using the Rössler chaotic input and a quasi-periodic input with a slow information component, the authors vary the input's dominant frequency and measure how well the next interspike interval can be predicted from past intervals. They find low prediction error only in the resonant range where the input frequency matches the neuron's intrinsic frequency, and for regular inputs also at integer multiples of that frequency. The authors conclude that class II neurons pass information about a stimulus only when the carrier frequency is resonant with the neuron's inherent oscillation, while simultaneously representing the signal strength through the modulation of interspike intervals.

Load-bearing premise

The argument assumes that a low normalized prediction error for the interspike intervals means the intervals are encoding the input, rather than merely being predictable because the neuron is firing in a nearly periodic, input-independent rhythm.

Editorial extensions

If this is right

  • Within its resonant band, a class II neuron's spike train carries analogue information about the stimulus envelope, not just a one-bit fired/not-fired signal.
  • Inputs whose dominant frequency lies outside the neuron's intrinsic range are filtered out, so downstream neurons receive spike packets only from the resonant channel.
  • For regular inputs, subharmonic resonances at twice or three times the intrinsic frequency also support encoding, widening the usable passband.
  • The degree to which a class II neuron retains class I-like frequency sensitivity sets a trade-off between filter selectivity and the fidelity of interspike-interval encoding.
  • The tested level of additive dynamical noise does not destroy the resonant encoding, indicating the mechanism can operate in noisy biological settings.

Reading between the lines

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

  • A direct extension would be an information-theoretic comparison between the input $S(t)$ and the decoded interspike intervals, such as mutual information or transfer entropy, which would put the encoding claim on a more direct footing than the drop in prediction error.
  • In a population of class II neurons with different intrinsic frequencies, the same mechanism would implement a form of frequency-division multiplexing: several stimulus streams could be carried on one input line and read out by neurons tuned to different carrier frequencies.
  • The resonant-passband behavior may also be testable in biophysical class II models such as Hodgkin–Huxley neurons, and in slice preparations where sinusoidal current injection with a slowly modulated amplitude could reveal whether spike intervals track the envelope only near resonance.
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

3 major / 5 minor

Summary. The manuscript studies a FitzHugh-Nagumo class II neuron driven by chaotic Rössler inputs and by quasi-periodic inputs, and measures the normalized prediction error NPE(1) of the resulting interspike interval (ISI) sequence. The authors report that NPE(1) is small only when the dominant input frequency is near the neuron's intrinsic firing frequency or its harmonics, and they interpret this as evidence that class II neurons act as AM processors: they filter inputs by frequency while encoding the signal amplitude and temporal waveform in the ISIs. The results are shown for three values of the intrinsic time constant alpha, for two input families, and with and without additive Gaussian noise.

Significance. If substantiated, the claim would give a concrete computational role to class II resonance: frequency-selective filtering combined with ISI-based amplitude modulation, potentially relevant to gamma and theta oscillations. The paper is exploratory and simulation-based, with qualitative consistency across several intrinsic frequencies and input types. However, the central inference from low NPE to stimulus encoding is not directly established, and no code or numerical data are provided; the significance is therefore conditional on additional control analyses.

major comments (3)
  1. [Eq. (1), Figs. 1 and 3] The central claim that low NPE(1) demonstrates that ISIs encode the stimulus S(t) is not supported by Eq. (1). NPE measures the local predictability of the ISI sequence from its own past, normalized by the error of the mean predictor; a neuron firing near-periodically with no stimulus dependence would also have a small NPE. The authors should either compute a direct measure relating the ISI sequence to S(t), such as mutual information or transfer entropy between S(t) and t_k, or compare NPE against a control with a constant or statistically independent input. Without such a control, the conclusion that 'the neuron passes the signal information' is not established.
  2. [Eq. (2), Fig. 3] The quasi-periodic-input analysis does not demonstrate that the slow amplitude component of S(t) is encoded. The low NPE dips around f, 2f, and 3f are interpreted as transmission of the information on the slow component, but the interval sequence is predicted only from itself; no reconstruction or spectral comparison between the ISIs and the 0.2 Hz envelope is shown. To support the AM-processor claim, the authors should show that the ISI sequence tracks the slow amplitude profile, for example by correlating smoothed ISIs with the envelope of S(t) or by computing coherence between the two.
  3. [Figs. 1-3] The manuscript does not provide error bars, repeated noise realizations, or a quantitative success threshold for 'successful coding'. The open-circle noisy cases appear to come from a single realization, and regions of small NPE are read visually. Because the NPE values are the only quantitative evidence for the main claim, the authors should state the number of realizations, report variability, and define a criterion for successful encoding relative to a baseline.
minor comments (5)
  1. [Fig. 1 and text] The 'dominant frequency of S(t)' is defined as the inverse of the averaged peak intervals, but the peak-detection method and averaging procedure are not described; please specify them or use the power-spectral peak.
  2. [Fig. 2] The 'instantaneous frequency' used in the density plots is not defined; please state whether it is the inverse of each interspike interval and how many intervals were used to estimate the densities.
  3. [Text after Eq. (2)] The statement that NPE(h) measures how accurately an interspike interval encodes the instantaneous value of 1/S(t) is not justified, especially since S(t) can be small or negative; no calculation relating intervals to 1/S(t) is presented.
  4. [Methods, prediction parameters] The prediction uses fixed values d=4 and l0=12 with no sensitivity analysis; a brief check that the NPE results are robust to these choices would increase confidence in the reported dips.
  5. [Availability] No code or data availability statement is included; given that all conclusions are based on simulations, providing code or at least the generated interval sequences would substantially aid reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the frequency scan is exploratory, no parameter is fitted to the reported error, and the encoding claim does not reduce to the definition of NPE.

full rationale

The paper's central claim is that class II FHN neurons filter inputs by frequency and encode signal strength in interspike intervals. The analysis does not fit any parameter to minimize the reported normalized prediction error: the input frequency parameters a and f are scanned as control variables, and the neuron's intrinsic frequency range is computed independently from constant-bias simulations. The resonance condition is read off the same NPE-versus-frequency figures, but it is not enforced by construction, and the NPE is computed from the simulated interspike interval sequence without tuning. Self-citations to Masuda and Aihara (refs. 15, 20, 21) support prior results about chaotic interspike interval coding and network behavior, but the prediction algorithm is credited to external references (refs. 13 and 14), so the load-bearing methodological support is not a self-citation chain. The interpretative step from small NPE(1) to 'the interspike intervals encode S(t)' is inferentially bold, and one could question whether low NPE reflects stimulus-dependent modulation rather than mere near-periodic firing; however, that is a validity concern, not a circularity. NPE is not defined in terms of S(t) or of the encoding claim, so the conclusion does not reduce to the metric by definition. The paper is self-contained in its numerical demonstration: it runs uncontrolled simulations over a parameter grid and reports the resulting errors. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no known result is merely relabeled. The observed frequency filtering and interval variability are established by direct simulation, with the resonance interpretation superimposed rather than baked into the measure. Accordingly, no circular step meeting the required evidentiary standard can be identified.

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

The paper is a simulation study built on the FHN model, the Rossler system, and a local prediction method. The central claim depends on the model choices and on the interpretation of NPE as an information metric. No new physical or biological entities are introduced.

free parameters (5)
  • FHN time constant alpha = 0.05, 0.03, 0.02
    Sets the neuron's intrinsic frequency; the paper shows the resonance effect for these three values but not for a continuous range. Chosen by hand, not fitted to biological data.
  • Rossler input bias and amplitude = 0.23, 0.0075
    Keeps the FHN neuron in a suprathreshold oscillatory regime; the modulation depth affects how strongly the input can imprint on interspike intervals.
  • Quasi-periodic input amplitudes and slow frequency = 0.24, 0.015, 0.04, 0.2 Hz
    The slow component is the information signal and the f component is the carrier; the amplitude ratio determines the interspike interval modulation depth and is not varied.
  • Prediction parameters = d=4, l0=12
    Used for local prediction of interspike intervals; NPE values and the apparent success regions could depend on these choices, and no sensitivity analysis is given.
  • Noise standard deviation = 0.0008 per dt=0.0004
    A single noise level is used for the robustness check; the paper does not examine the dependence of the results on noise intensity.
assumptions (5)
  • domain assumption The FitzHugh-Nagumo model with the given parameter values is a valid representation of class II neuronal dynamics.
    Used throughout the model section; the parameters are standard in the literature, but the paper provides no biological validation for this specific regime.
  • domain assumption A Rossler chaotic signal or a quasi-periodic two-tone signal captures the relevant structure of real stimuli like sounds and odors.
    Introduced in the model section; the inputs are chosen for convenience and are not derived from experimental recordings.
  • domain assumption The normalized prediction error NPE(h) computed with d=4, l0=12 measures whether interspike intervals encode the input.
    Borrowed from refs [13-15]; the paper assumes low NPE means successful encoding, without an independent information-theoretic check.
  • domain assumption The dominant frequency of an input can be represented by the inverse of the averaged peak intervals of S(t).
    Used in Fig.1 and Fig.3 to characterize the input; a single scalar frequency may not capture broadband chaotic inputs.
  • domain assumption Additive Gaussian noise with standard deviation 0.0008 approximates biological noise.
    One noise level is tested; the accumulated-noise estimate ignores membrane leak and other sources of variability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Filtered interspike interval encoding by class II neurons." pith.science (2026). https://pith.science/paper/JGPC3G4P

@misc{pith2026241114692,
  author       = {Pith},
  title        = {Pith review of: Filtered interspike interval encoding by class II neurons},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JGPC3G4P}},
  note         = {Machine review of arXiv:2411.14692}
}
read the original abstract

Dynamics of class II neurons, firing frequencies of which are strongly regulated by the inherent neuronal property, have been extensively studied since the formulation of the Hodgkin--Huxley model in 1952. However, how class II neurons process stimulus information and what kind of external information and internal structure firing patterns of neurons represent are vaguely understood in contrast to firing rate coding by class I neurons. Here we show that the FitzHugh--Nagumo class II neuron simultaneously filters inputs based on the input frequency and represent the signal strength by interspike intervals. In this sense, the class II neuron works as an AM processor that passes the information on the carrier and on the temporal waveform of signals.

Figures

Figures reproduced from arXiv: 2411.14692 by the authors.

Figure 1
Figure 1. Performance of interspike interval coding of a FHN neuro [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The distributions of the instantaneous firing frequency w [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) Performance of interspike interval coding in the noise [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages

  1. [1]

    L. S. Benardo, R. E. Foster, Brain Res. Bull. 17 (1986) 773

    work page 1986

  2. [2]

    Llin´ as, Y

    R. Llin´ as, Y. Yarom, J. Physiol. 376 (1986) 163

    work page 1986

  3. [3]

    Bragin et al

    A. Bragin et al. , J. Neurosci. 15 (1995) 47

    work page 1995

  4. [4]

    Steriade, I

    M. Steriade, I. Timofeev, N. D¨ urm¨ uller, F. Grenier, J. Neurop hysiol. 79 (1998) 483

    work page 1998

  5. [5]

    Gutfreund, Y

    Y. Gutfreund, Y. Yarom, I. Segev, J. Physiol. 483 (1995) 621

    work page 1995

  6. [6]

    R. D. Traub et al. , J. Physiol. 493 (1996) 471

    work page 1996

  7. [7]

    C. M. Gray, P. K¨ onig, A. K. Engel, and W. Singer, Nature 338 (19 89) 334

    work page

  8. [8]

    Llin´ as, U

    R. Llin´ as, U. Ribary, Proc. Natl. Acad. Sci. USA 90 (1993) 2078

    work page 1993

Show all 28 references
  1. [9]

    Aihara, G

    K. Aihara, G. Matsumoto, J. Theoretical Biology, 95 (1982) 697

    1982

  2. [10]

    T. Lu, L. Liang, and X. Wang, Nature Neuroscience 4 (2001) 11 31

    2001

  3. [11]

    E. M. Izhikevich, Int. J. of Bifurcation and Chaos 10 (2000) 11 71

    2000

  4. [12]

    Sigeti, and W

    D. Sigeti, and W. Horsthemke, J. Stat. Phys. 54 (1989) 1217

    1989

  5. [13]

    Sauer, Phys

    T. Sauer, Phys. Rev. Lett. 72 (1994) 3811

    1994

  6. [14]

    D. M. Racicot, A. Longtin, Physica 104D (1997) 184

    1997

  7. [15]

    Masuda and K

    N. Masuda and K. Aihara, Neural Computation 14 (2002) 1599

    2002

  8. [16]

    Abeles, Corticonics, Cambridge University Press, Cambridge , 1991

    M. Abeles, Corticonics, Cambridge University Press, Cambridge , 1991

    1991

  9. [17]

    Diesmann, M-O

    M. Diesmann, M-O. Gewaltig, A. Aertsen, Nature 402 (1999) 52 9

    1999

  10. [18]

    M. N. Shadlen, W. T. Newsome, J. Neurosci. 18 (1998) 3870

    1998

  11. [19]

    D. J. Mar, C. C. Chow, W. Gerstner, R. W. Adams, J. J. Collins, P roc. Natl. Acad. Sci. USA 96 (1999) 10450

    1999

  12. [20]

    Masuda, K

    N. Masuda, K. Aihara, Phys. Rev. Lett. 88 (2002) 248101

    2002

  13. [21]

    Masuda, K

    N. Masuda, K. Aihara, Neural Computation 15 (2003) 103

    2003

  14. [22]

    M. C. W. van Rossum, G. G. Turrigiano, S. B. Nelson, J. Neurosc i. 22 (2002) 1956

    2002

  15. [23]

    Longtin, Phys

    A. Longtin, Phys. Rev. E 55 (1997) 868

    1997

  16. [24]

    A. S. Pikovsky, J. Kurths, Phys. Rev. Lett. 78 (1997) 775

    1997

  17. [25]

    Hutcheon, Y

    B. Hutcheon, Y. Yarom, Trends in Neurosciences 23 (2000) 21 6

    2000

  18. [26]

    E. M. Izhikevich, Neural Networks 14 (2001) 883

    2001

  19. [27]

    FitzHugh, Biophysical J

    R. FitzHugh, Biophysical J. 1 (1961) 445

    1961

  20. [28]

    Nagumo, S

    J. Nagumo, S. Arimoto, and S. Yoshizawa, Proc. of the IRE 50 ( 1962) 2061. 5

    1962

Pith tools

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