{"id":"c8cb67e5-f3f1-45bd-809e-36bdc2bba1ee","arxiv_id":"2411.14692","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Class II FitzHugh-Nagumo neurons filter inputs by frequency and encode the input waveform in interspike intervals near resonance.","lead":"A computational study shows that a class II neuron model transmits information about an input only when the input's frequency matches the neuron's own rhythm. This makes the neuron a frequency filter that encodes the input's slower waveform in the timing between spikes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Low normalized prediction error does not establish that interspike intervals encode S(t); NPE can be small for near-periodic firing even with no stimulus-dependent modulation.","rationale":"The reader's weakest assumption was the same gap: NPE used as a proxy for stimulus encoding. I agree with that diagnosis. The central claim is plausible and qualitatively supported by the resonance-dependent dips in NPE, but the AM encoding statement requires a direct measure of dependence between S(t) and the interval sequence. Low NPE can arise from near-periodic firing even in the absence of causal coupling to the stimulus, so the paper's evidence is consistent with, but does not uniquely support, the claim that interspike intervals carry information about the temporal waveform. The proposed mutual-information test with surrogate significance would settle whether intervals contain stimulus information beyond what a constant-input oscillator would produce. Because no code or numerical data are provided for independent audit, conditional acceptance remains appropriate; the reader's verdict does not need adjustment.","tokens_in":6732,"tokens_out":4381,"duration_ms":54116,"concrete_test":"Compute, for the quasi-periodic setup with alpha=0.05, the mutual information I(t_k; S(tau_k)) between each interval and the stimulus at the associated spike time, with a bias-corrected estimator and time-shifted surrogates, for f=0.6 (resonant) and f=0.2 (nonresonant). In parallel, run an uncoupled FHN neuron driven by the constant mean of S(t) with matched firing rate and compute NPE(1) on its intervals. If the surrogate-corrected MI is not significantly positive in the resonant band, or the constant-input oscillator yields the same low NPE(1), then the small NPE reflects near-periodic intrinsic firing rather than encoding of S(t).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the inference from Eq. (1): low NPE(1) is read as 'interspike intervals encode S(t).' NPE measures local predictability of the interval sequence {t_k}, not dependence of {t_k} on the stimulus. A FHN neuron that is phase-locked or firing near-periodically will have a nearly constant interval sequence and therefore a small NPE even if S(t) is causally irrelevant to the intervals; normalization by the error of the mean prediction does not remove this confound. The spread in instantaneous frequencies in Fig. 2(b) is consistent with cycle-to-cycle variability, but the paper never directly estimates correlation, mutual information, or transfer entropy between S(t) and the intervals. In the quasi-periodic case, the AM claim requires that the slow amplitude profile, not the carrier phase, is recoverable from the intervals; broad NPE dips around f and its harmonics do not establish that. Since no code or numeric data are provided, the resonance-dependent NPE dips are evidence for predictable firing, but the central encoding claim remains conditional.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":6888,"tokens_out":3311,"duration_ms":37972,"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":[{"comment":"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.","section":"Eq. (1), Figs. 1 and 3"},{"comment":"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.","section":"Eq. (2), Fig. 3"},{"comment":"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.","section":"Figs. 1-3"}],"minor_comments":[{"comment":"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.","section":"Fig. 1 and text"},{"comment":"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.","section":"Fig. 2"},{"comment":"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.","section":"Text after Eq. (2)"},{"comment":"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.","section":"Methods, prediction parameters"},{"comment":"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.","section":"Availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the qualitative effect is plausible, but the load-bearing step—interpreting low NPE as evidence of stimulus encoding—needs direct validation. The requested control analyses and error bars are feasible within the existing simulation framework, so major revision rather than rejection seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the claim that a class II FHN neuron can both filter inputs by frequency and encode signal strength in its interspike intervals, working as an AM processor. That gives class II neurons a functional role beyond fixed-frequency resonators, and it is worth taking seriously. The paper backs it up with simulations across three intrinsic frequencies (alpha), two input families (Rossler chaotic and quasi-periodic), and additive noise. The NPE resonance dips shift with alpha as expected, and the qualitative picture is consistent. The authors also build honestly on earlier chaos-prediction work, so the citation pattern is not a problem.\n\nThe soft spot is the load-bearing inference. NPE measures local predictability of the interval sequence, not dependence of those intervals on the stimulus S(t). A neuron firing nearly periodically can have low NPE even if the stimulus is causally irrelevant. The paper reads 'successful coding' visually from NPE curves without computing a direct correlation, mutual information, or transfer entropy between S(t) and the predicted intervals. So the claim that 'the neuron passes the signal information' is not fully established by the metric used. For the quasi-periodic AM case, the argument requires that the slow amplitude profile is recoverable from the intervals, and that too is only inferred from NPE dips. The lack of error bars and code is a real but secondary issue; the noisy curves look like single realizations, and the looseness around success thresholds weakens the quantitative claims.\n\nThat said, the qualitative effect is not a fabrication. The contrast between resonant and non-resonant NPE values, and the systematic shift with alpha, points to something real in this model. The paper deserves a serious referee. The key request should be a direct test of stimulus encoding, ideally with mutual information or a similar measure, plus code and error statistics. With those added, the central idea would be much stronger. I would send it to peer review rather than desk reject, and I would cite it if I was working on neural coding, but I would not take the encoding claim at face value yet.","headline":"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.","tokens_in":7444,"tokens_out":1670,"would_cite":true,"duration_ms":20530,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["87.19.La","05.45.Xt","87.18.Sn"],"model":"deepseek-v4-flash","headline":"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.","keywords":["class II neuron","FitzHugh-Nagumo model","interspike interval","frequency filtering","amplitude modulation","resonance","normalized prediction error","chaotic input"],"falsifier":"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.","tokens_in":6481,"feed_emoji":"🧠","tokens_out":8296,"duration_ms":77292,"temperature":0.7,"pith_summary":"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.","feed_headline":"Resonant class II neurons filter inputs, encode in spike intervals","feed_subtitle":"Interspike intervals carry the stimulus waveform only when its dominant frequency matches the neuron's intrinsic rhythm.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Defines class I/II excitability and identifies class II neurons as resonators, the starting point for the filtering claim.","marker":"[11]"},{"why":"Introduces the local-prediction approach for reading deterministic information from interspike intervals.","marker":"[13]"},{"why":"Shows how interspike intervals encode continuous chaotic inputs, setting the baseline for the present prediction-error analysis.","marker":"[14]"},{"why":"Earlier study by the authors on interspike-interval coding that supplies the specific prediction method and the chaotic-drive setup.","marker":"[15]"},{"why":"Introduced the FitzHugh model, one half of the FitzHugh–Nagumo neuron used throughout.","marker":"[27]"},{"why":"Introduced the Nagumo circuit realization, completing the FitzHugh–Nagumo model.","marker":"[28]"},{"why":"Experimental evidence that class II inferior-olive neurons resonate with input frequency, supporting the biological relevance.","marker":"[5]"},{"why":"Review of neuronal resonance that grounds the frequency-filtering interpretation.","marker":"[25]"}],"fun_headline_variants":["Class II neurons act as AM processors, filtering by frequency","Interspike intervals encode amplitude only at resonance","Neurons encode signal strength in spike timing at resonant inputs","Resonant inputs pass, others filtered: spike intervals carry amplitude","AM modulation in neurons: frequency filter, interval encoding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Class II neurons act as AM processors, filtering by frequency","Interspike intervals encode amplitude only at resonance","Neurons encode signal strength in spike timing at resonant inputs","Resonant inputs pass, others filtered: spike intervals carry amplitude","AM modulation in neurons: frequency filter, interval encoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000274,"raw_usage":{"total_tokens":1576,"prompt_tokens":819,"completion_tokens":757,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":435,"completion_tokens_details":{"reasoning_tokens":676}},"tokens_in":435,"tokens_out":757,"duration_ms":7999,"temperature":1.0,"reasoning_tokens":676,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:00:52.435015+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Nagumo, S","cited_arxiv_id":null,"evidence_quote":"Introduced the Nagumo circuit realization, completing the FitzHugh–Nagumo model."},{"cited_title":"Gutfreund, Y","cited_arxiv_id":null,"evidence_quote":"Experimental evidence that class II inferior-olive neurons resonate with input frequency, supporting the biological relevance."},{"cited_title":"Hutcheon, Y","cited_arxiv_id":null,"evidence_quote":"Review of neuronal resonance that grounds the frequency-filtering interpretation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines class I/II excitability and identifies class II neurons as resonators, the starting point for the filtering claim."},{"cited_title":"Sauer, Phys","cited_arxiv_id":null,"evidence_quote":"Introduces the local-prediction approach for reading deterministic information from interspike intervals."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows how interspike intervals encode continuous chaotic inputs, setting the baseline for the present prediction-error analysis."},{"cited_title":"Masuda and K","cited_arxiv_id":null,"evidence_quote":"Earlier study by the authors on interspike-interval coding that supplies the specific prediction method and the chaotic-drive setup."},{"cited_title":"FitzHugh, Biophysical J","cited_arxiv_id":null,"evidence_quote":"Introduced the FitzHugh model, one half of the FitzHugh–Nagumo neuron used throughout."}],"review_version":1}