{"id":"5c3ea6d8-8322-44f2-99e8-1c800c108d1a","arxiv_id":"2412.08816","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A greedy algorithm that adds neurons one at a time, each chosen to maximize joint mutual information, improves spike-encoding classification accuracy in two neuromorphic tasks.","lead":"This paper presents an algorithm that tunes the firing thresholds of a population of spiking neurons by maximizing the mutual information between a signal and the resulting spike trains. In two classification tasks, the tuned populations capture more information and classify more accurately than randomly configured populations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on a quadratic extrapolation defining 'true' MI; without fit diagnostics or error bars, the selected thresholds and reported MI gains may be artifacts of the estimator.","rationale":"The reader's weakest assumption correctly identifies the mutual information extrapolation as the most load-bearing component: the entire algorithm selects thresholds by maximizing a fitted intercept, and the headline claims about population gains and near-optimal accuracy are only as strong as that estimate. My review found no independent evidence—parameter-free derivations, machine-checked proofs, or released code—that would reduce this risk. The two applications reproduce the qualitative trend, which is genuine support, but they use the same estimation pipeline, so they do not independently validate the extrapolation. I do not see a more fundamental internal inconsistency: the greedy recursion and the use of joint MI are mathematically coherent, and the claim that adding a neuron cannot decrease joint MI is correct. The weakest point remains the unvalidated mapping from finite-window estimates to 'true' MI. Because this is an addressable empirical concern rather than a demonstrated error, the reader's CONDITIONAL verdict is appropriate; no verdict change is needed. My agreement is 'agree' because the reader's weakest assumption and my load-bearing concern are the same.","tokens_in":13101,"tokens_out":3253,"duration_ms":40475,"concrete_test":"Re-run the m=2 threshold selection for Application 1 using bootstrap resampling of the 260 waveforms and two alternative sets of window durations for the Eq. (5) fit (e.g., T={10,20,40,80} and T={15,30,60,120}). For each candidate spike density, compute the intercept a with 95% bootstrap confidence intervals and report the fit residuals. If the argmax density moves outside the 15% value reported, or if the confidence intervals for a overlap across densities from 10% to 30%, then the extrapolated MI is not reliable enough to support the threshold selection and the population-coding claim would need to be re-evaluated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that maximizing the extrapolated mutual information yields spike-encoding parameters whose population-level MI and classification accuracy both increase. The linchpin is Section 2.2, Eq. (5): I(X;W_T)=a+b/T+c/T^2, with the 'true' MI taken as the intercept a (Eq. 6). Every reported MI value and every threshold choice depends on this fitted intercept. If the assumed 1/T, 1/T^2 expansion is not accurate for the k-nearest-neighbor estimator used, or if the fit is unstable across the chosen window durations, then the value a is not the true MI and the greedy threshold selections may simply maximize a fitting artifact. The paper gives no fit diagnostics (residuals, number of T values, T range, goodness of fit) and no error bars or bootstrap uncertainty on a. Since MI is estimated from only 260 (Application 1) or 1000 (Application 2) waveforms and the selected threshold is then applied to the full dataset, sampling variability in the extrapolated intercept is also unquantified. All downstream claims—population MI gains, near-optimal accuracy, and the MI-as-proxy conclusion—inherit this uncertainty. The consistency of the two applications is encouraging, but it does not by itself establish that the intercept a is a reliable estimate of the quantity the algorithm claims to optimize.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a greedy algorithm for tuning the parameters of a population of leaky integrate-and-fire (LIF) neurons that encode a real-valued stimulus into spike trains. Starting from a single neuron whose firing threshold is chosen to maximize an estimate of mutual information I(X; W_1), the algorithm recursively adds neurons to maximize the joint mutual information I(X; {W_1, ..., W_m}), treating previously selected neurons as a fixed aggregate. The mutual information is estimated with a k-nearest-neighbor estimator and extrapolated to infinite observation-window duration via a quadratic fit in 1/T, with the intercept taken as the 'true' mutual information. The method is evaluated on two tasks: 13-way classification of blood pressure pulse waves (PWDB) and 21-way classification of simulated action potential waveforms (SYNTH Monotrode). The authors report that both mutual information and classification accuracy increase as neurons are added, that the mutual-information-selected thresholds achieve accuracy close to thresholds found by an exhaustive search for accuracy, and that they outperform random threshold selection. The central conclusion is that mutual information can serve as a task-independent proxy for tuning spike-encoding parameters in neuromorphic systems.","tokens_in":13420,"tokens_out":4702,"duration_ms":47843,"significance":"If the reported results are reliable, the paper offers a practical and conceptually appealing recipe for configuring neuromorphic encoding populations without task-specific labels, with potential computational savings. The use of two distinct datasets, a comparison against an exhaustive accuracy-oriented search, and a random-parameter baseline are strengths, and the qualitative consistency of the mutual-information/accuracy relationship across both applications is encouraging. However, the central quantitative claims rest on an unvalidated extrapolation of mutual-information estimates, and the accuracy comparisons lack uncertainty quantification; these issues currently prevent the stronger conclusions from being fully supported.","major_comments":[{"comment":"The definition of the 'true' mutual information as the intercept a of the quadratic fit I(X; W_T) = a + b/T + c/T^2 is not validated. The paper does not report the number of window durations T used, the range of T, the residuals of the fit, any goodness-of-fit statistic, or uncertainty estimates for a. Since every threshold selection in Figs. 5 and 7 and every mutual-information value in Tables 1 and 2 is derived from this extrapolated intercept, the claim that the algorithm maximizes true mutual information is not currently supported. Please provide fit diagnostics, a sensitivity analysis of the chosen T values, and error bars or bootstrap intervals for a; without these, the reported information gains and parameter choices may be artifacts of the extrapolation procedure.","section":"Section 2.2, Eq. (5)"},{"comment":"All classification accuracies are point estimates from a single 80/20 split, with no repeated splits, cross-validation, or confidence intervals. The claims that the proposed approach achieves 'near-optimal' accuracy and 'significantly outperforms' random parameter selection—the latter based on only five random trials—are therefore not statistically substantiated. Please report accuracy variability across multiple splits or bootstrap resamples and provide a larger number of random trials with a significance test to support the comparison.","section":"Section 3.3 and Fig. 8"},{"comment":"The 'exhaustive search' used to obtain the accuracy-optimal parameters is not specified: the grid resolution over spike densities, the data split used, and whether the same 260/1000 waveforms were used for parameter selection are all missing. Without these details, the near-optimality claim in Fig. 8 cannot be reproduced or evaluated. Please describe the exhaustive search procedure and, ideally, report the actual accuracy values and grid density.","section":"Section 3.3 and Fig. 8"},{"comment":"The algorithm is motivated by Partial Information Decomposition, and the abstract claims gains 'beyond what could be accounted for by each neuron separately,' but no PID atoms (redundant, unique, synergistic) are ever computed. The observed increase in joint mutual information over single-neuron mutual information does not by itself establish unique or synergistic contributions, because the individual mutual information of each added neuron is not reported and the greedy recursion fixes previous neurons. Either compute the PID decomposition (or at least report the single-neuron mutual information of each added neuron) or soften the synergy-related wording so that the claim matches what is actually measured.","section":"Section 2.1 and Section 4"}],"minor_comments":[{"comment":"The displayed equation for the quadratic extrapolation is garbled in the manuscript ('T b C'); it should read I(X; W_T) = a + b/T + c/T^2.","section":"Eq. (5)"},{"comment":"In the sentence introducing the second neuron, 'J(X; W2)' should be 'I(X; W2)'.","section":"Section 2.1.2"},{"comment":"The database size is given as '437 4 virtual subjects'; this appears to be a typo for 4,374.","section":"Section 3.1.1"},{"comment":"The text says SVMs are implemented with 'sciki t-learn'; the correct name is scikit-learn.","section":"Section 3"},{"comment":"The caption contains a garbled character sequence 'n:,;}' in the expression for the spike trains; please clean this up.","section":"Fig. 1 caption"},{"comment":"The grid search over firing thresholds is described only as spanning the range of spike densities; the number of threshold values and the density resolution should be stated for reproducibility.","section":"Section 3.1.2 and 3.2.2"},{"comment":"The term 'near-optimal' is used without a quantitative definition; please specify a tolerance or margin relative to the exhaustive-search accuracy.","section":"Section 3.3"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the mutual-information extrapolation in Eq. (5), which is the linchpin of the parameter selection and the reported MI values. If the authors can supply fit diagnostics, T-range sensitivity, and uncertainty estimates, the paper could be publishable. The scope (neuromorphic spike encoding) and the two application studies fit the journal; note that the single-neuron MI optimization was introduced in the authors' prior work [9,10], so the novelty is primarily the greedy population extension and its validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the paper proposes something simple and reasonably new — greedily tune the firing thresholds of a population of LIF neurons by maximizing joint mutual information, instead of tuning for a task. The authors' previous work did this for single neurons; here they extend it to populations with a forward-selection heuristic. The PID framing is mostly motivational; the actual algorithm is standard greedy feature selection. The genuinely useful part is the empirical demonstration on two datasets: PWDB blood pressure pulse waves (13 sites) and SYNTH Monotrode action potential waveforms (21 classes). In both, adding neurons increases both MI and classification accuracy, the MI-tuned thresholds reach accuracy close to an exhaustive search for accuracy, and they clearly beat random threshold choices. That is a real result for neuromorphic system designers who want a task-independent way to initialize encoding parameters.\n\nThe soft spots are about the evidence, not the idea. The main one: the 'true' MI is the intercept of a quadratic fit to MI estimates at several window durations (Eq. 5). The paper gives no fit diagnostics — no number of T values, no T range, no residuals, no bootstrap uncertainty. Every threshold selection and every reported MI number depends on that intercept. If the fit is unstable, the greedy choices could be chasing noise. The two datasets agreeing is encouraging, but it is not the same as showing the extrapolation is reliable.\n\nThe accuracy numbers also come from a single 80/20 split — no error bars, no repeated splits. The 'exhaustive search' baseline for near-optimal accuracy is underspecified: we don't know the grid resolution, which matters for how 'near-optimal' looks. And the abstract's synergy claim is overstated: they never actually compute PID atoms. Showing joint MI > single-neuron MI is expected when you add a different neuron; it doesn't by itself show synergistic encoding. That is a wording issue, not a fatal flaw.\n\nMy read: the qualitative story is probably right, and the contribution is modest but real. The audience is people building neuromorphic pipelines who need a principled way to pick encoding parameters without training a task model. The paper needs a robustness pass — bootstrap or repeated splits for MI and accuracy, fit diagnostics for Eq. (5), and a clearer description of the exhaustive search. I'd send it to a referee; with those fixes it would be a solid methods paper, and even as-is it's worth engaging with.","headline":"A plausible, modest contribution — greedy MI-based tuning of LIF population thresholds — whose qualitative results are likely right, but the numbers rest on an extrapolated MI estimator with no fit diagnostics.","tokens_in":13899,"tokens_out":3764,"would_cite":true,"duration_ms":39642,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding neurons to a spike-encoding population, chosen greedily to maximize mutual information, raises both encoded information and classification accuracy, nearly matching task-specific tuning.","keywords":["spike encoding","population coding","mutual information","partial information decomposition","neuromorphic","leaky integrate-and-fire","LIF neuron","classification accuracy"],"falsifier":"Re-run the two protocols using a bias-corrected mutual-information estimator at a single sufficiently long window (or with bootstrap confidence intervals over many data subsamples) instead of the quadratic extrapolation, and compare the resulting parameter rankings and accuracy curves. If the alternative estimator changes which spike densities are selected, or if the extrapolated MI stops tracking held-out classification accuracy as the population grows, the central claim would be contradicted.","tokens_in":12906,"feed_emoji":"🧠","tokens_out":6523,"duration_ms":61742,"temperature":0.7,"pith_summary":"This paper tries to show that the encoding parameters of a population of spiking neurons do not need to be tuned for a specific task: greedily choosing each neuron's firing threshold to maximize the mutual information between the input signal and the resulting spike trains produces configurations whose classification accuracy is near the best achievable with task-specific tuning. The authors argue that population coding adds information beyond any single neuron—unique and synergistic contributions—and that these gains track accuracy as neurons are added. They demonstrate this in two unrelated tasks, blood-pressure pulse-wave classification and neural action-potential waveform classification, using leaky integrate-and-fire neurons. If the claim holds, mutual information can serve as a task-independent proxy for setting spike-encoding parameters, which matters because spike encoding is a bottleneck for neuromorphic systems and task-specific optimization can be expensive.","feed_headline":"Greedy information tuning makes spike codes near-optimal","feed_subtitle":"Adding neurons raises encoded bits and classification accuracy in two tasks, matching task-specific tuning without training a model.","key_machinery":"The machinery is the recursive greedy maximization of joint mutual information. At each step the algorithm treats the population selected so far as one fixed random variable $W_1$, then searches over the firing threshold of a new leaky integrate-and-fire neuron to maximize $I(X; \\{W_1, W_2\\})$; the new neuron is accepted at the threshold giving the largest joint information, and the process repeats until a stopping condition. The estimated quantity is a discrete-continuous mutual information between signal amplitude and spike patterns, computed with a k-nearest-neighbors estimator and then extrapolated to an infinite observation window through the quadratic fit $I(X; W^T)=a+b/T+c/T^2$, whose intercept $a$ is taken as the 'true' mutual information.","core_discovery":"On the paper's own terms, the central discovery is that a greedy, information-theoretic construction of a neuron population is both practically effective and nearly optimal. Starting with the neuron whose spike train maximizes $I(X; W_1)$, the algorithm fixes that neuron and chooses each next neuron to maximize the joint information $I(X; \\{W_1, W_2\\})$ between the stimulus and the whole population, with the previously chosen neurons treated as a single aggregate variable; this is inspired by the Partial Information Decomposition of joint information into redundant, unique, and synergistic atoms. In the pulse-wave task, tuning five such neurons raises the estimated mutual information from 3.058 to 6.116 bits and classification accuracy from 45.1% to 68.6%; in the action-potential task it raises MI from 1.876 to 3.187 bits and accuracy from 81.8% to 94.8%. The authors report that these accuracy values closely track the MI curve, match an exhaustive task-specific parameter search, and beat random parameter selection, with diminishing returns as neurons are added.","pith_inferences":["A natural extension the paper leaves implicit is to test whether MI-optimal parameters stay near-optimal when the classifier or feature extractor changes, since the current evidence uses one SVM with spike-count bins.","The paper does not report the redundant, unique, and synergistic atoms it invokes; computing those atoms directly would show whether the accuracy gains come mostly from unique or synergistic information, and could predict when adding neurons stops helping.","The quadratic extrapolation is a potential fragility; replacing it with modern bias-correction techniques could make the approach reliable on shorter recordings where the fit is less stable.","If the MI-accuracy link generalizes, the algorithm could serve as a model-free diagnostic for comparing different spike-encoding schemes before committing to a downstream neuromorphic pipeline."],"forward_implications":["Mutual information can be used as a task-independent target when configuring spike encoders, removing the need to train or evaluate a downstream model during parameter search.","Population coding in neuromorphic encoders can yield joint gains beyond single-neuron information, with accuracy rising as neurons are added until diminishing returns set in.","The greedy algorithm can also serve as a principled way to choose how many neurons to include, by watching the MI curve flatten.","The same tuning recipe should transfer to other spike encoders and other signal types, since the objective does not depend on the classification model or task."],"supporting_citations":[{"why":"Earlier work by the same group optimizing single-neuron spike encoding with information theory; this paper extends that approach to populations.","marker":"[10]"},{"why":"Shows information-theoretic efficiency metrics for neuromorphic spike encoders, establishing MI as task-independent and computationally cheaper than task-specific optimization.","marker":"[9]"},{"why":"Introduces Partial Information Decomposition, whose redundant/unique/synergistic atoms motivate maximizing joint information when adding each neuron.","marker":"[35]"},{"why":"Provides the k-nearest-neighbors estimator used to compute mutual information between continuous signal amplitudes and discrete spike trains.","marker":"[29]"},{"why":"Establishes the expansion of finite-window information estimates in powers of 1/T that underlies the quadratic extrapolation to true MI.","marker":"[32]"},{"why":"Analyzes the upward sampling bias in finite-sample information measures and supports the extrapolation method used to estimate true MI.","marker":"[34]"},{"why":"Supplies the SYNTH Monotrode dataset of simulated extracellular action potential recordings used in Application 2.","marker":"[26]"},{"why":"Supplies the Pulse Wave Database of simulated arterial pulse waves used in Application 1.","marker":"[6]"},{"why":"Provides the scikit-learn SVM implementation used to measure classification accuracy for the near-optimality comparison.","marker":"[25]"}],"fun_headline_variants":["Greedy info tuning near-optimal for spike codes","Population spike codes tuned greedily hit near-optimal accuracy","Maximizing mutual info in neuron populations boosts spike encoding","Greedy neuron selection for spike codes beats random tuning","Information-maximizing spike population nearly optimal in two tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim stands or falls on the reliability of the estimated 'true' mutual information, which is obtained by extrapolating finite-window estimates to infinite window duration with a quadratic fit; if that extrapolation is biased or unstable, the greedy parameter choices, the reported MI gains, and the claimed link to accuracy would not be trustworthy.","fun_headline_variants_meta":{"raw":{"variants":["Greedy info tuning near-optimal for spike codes","Population spike codes tuned greedily hit near-optimal accuracy","Maximizing mutual info in neuron populations boosts spike encoding","Greedy neuron selection for spike codes beats random tuning","Information-maximizing spike population nearly optimal in two tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1325,"prompt_tokens":1027,"completion_tokens":298,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":217}},"tokens_in":643,"tokens_out":298,"duration_ms":3472,"temperature":1.0,"reasoning_tokens":217,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:31:12.378040+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the two protocols using a bias-corrected mutual-information estimator at a single sufficiently long window (or with bootstrap confidence intervals over many data subsamples) instead of the quadratic extrapolation, and compare the resulting parameter rankings and accuracy curves. If the alternative estimator changes which spike densities are selected, or if the extrapolated MI stops tracking held-out classification accuracy as the population grows, the central claim would be contradicted.","supporting_citations":[{"cited_title":"Entropy and information in neural spike trains","cited_arxiv_id":null,"evidence_quote":"Establishes the expansion of finite-window information estimates in powers of 1/T that underlies the quadratic extrapolation to true MI."},{"cited_title":"Optimizing neuromorphic spike encoding of dynamic stimulus signals using information theory","cited_arxiv_id":null,"evidence_quote":"Earlier work by the same group optimizing single-neuron spike encoding with information theory; this paper extends that approach to populations."},{"cited_title":"Efficiency metrics for auditory neuro­ morphic spike encoding techniques using information theory","cited_arxiv_id":null,"evidence_quote":"Shows information-theoretic efficiency metrics for neuromorphic spike encoders, establishing MI as task-independent and computationally cheaper than task-specific optimization."},{"cited_title":"The upward bias in measures of information derived from limited data samples","cited_arxiv_id":null,"evidence_quote":"Analyzes the upward sampling bias in finite-sample information measures and supports the extrapolation method used to estimate true MI."},{"cited_title":"How many neurons can we see with current spike sorting algorithms?","cited_arxiv_id":null,"evidence_quote":"Supplies the SYNTH Monotrode dataset of simulated extracellular action potential recordings used in Application 2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Pulse Wave Database of simulated arterial pulse waves used in Application 1."},{"cited_title":"Scikit-learn: Machine learning in Python","cited_arxiv_id":null,"evidence_quote":"Provides the scikit-learn SVM implementation used to measure classification accuracy for the near-optimality comparison."}],"review_version":1}