Neuromorphic Online Clustering and Its Application to Spike Sorting

James E. Smith

REVIEW 4 major objections 6 minor 14 references

Neuromorphic Online Clustering and Its Application to Spike Sorting

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 ↗ Review JSON

Pith's one-line read A single neuromorphic dendrite, using only additions, clusters streaming spike waveforms online and matches or beats the offline, multi-pass k-means algorithm on a synthetic benchmark.

desk verdict A clean online clustering mechanism with a genuinely useful search-off ablation, but the k-means comparison is under-rigged and the missing variance erodes confidence in the headline numbers. read the letter →

arxiv 2506.12555 v1 pith:AUYTC7V6 submitted 2025-06-14 cs.NE cs.AI

James E. Smith This is my paper · ORCID

classification cs.NEcs.AI

keywords neuromorphicdendriteonlineclusteringspikesortingactivedendritesk-meansbaselinelow-precisionhardwaresyntheticwaveformssingle-passlearning

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 tries to establish that a single "neuromorphic dendrite"—a small unit inspired by active dendrites that reads feature vectors and updates a set of weight templates as each input arrives—can perform online clustering of action-potential waveforms, and that on a synthetic spike-sorting benchmark this single-pass, addition-only method matches or beats the offline, multi-pass k-means algorithm. The authors argue that this matters because spike sorting in real experiments could benefit from a low-cost method that learns while the data stream arrives, adapts to changing inputs, and needs no multi-pass retraining. They also claim that the "search" update is the key to the dendrite's advantage, and that a probabilistic version of search preserves accuracy while reducing arithmetic cost.

What carries the argument

The central object is the neuromorphic dendrite: a set of $p$ templates, each an $m \times n$ array of integer weights. For an input feature vector $\mathbf{x}$, inference computes, for each template, the sum of the weights addressed by the input feature values (expanded to a radius $r$ of neighboring values under similarity coding) and outputs the identifier of the template with the largest sum. Immediately afterward, the winning template increments the addressed weights by a capture amount, decrements the unaddressed weights in the same positions by a backoff amount, and every losing template raises its addressed weights by a small search amount up to a base weight. This update rule, using only additions and subtractions, is what lets a single dendrite form, hold, and adapt clusters in one pass.

What would settle it

Run the same synthetic spike sorting benchmark but tune all six hyperparameters separately for every combination of neuron count and instance deviation, then compare per-point tuned nD accuracy against the same k-means baseline; alternatively, run nD on extracellular recordings with known ground truth from simultaneous intracellular and extracellular measurements. If per-point tuning does not preserve the advantage, or if nD accuracy collapses on real waveforms, the central claim would be refuted.

Watch

Extended reading notes

Core claim

On a benchmark where six synthetic features define canonical spike shapes, 4–12 base neurons are generated as deviations from a canonical shape, and 10,000 spike instances stream in with controlled instance-to-instance variability, the proposed neuromorphic dendrite (nD) is claimed to sort spikes more accurately than offline k-means for small instance deviations and about as accurately for larger deviations, while making only one pass over the stream compared with k-means's 4–20 passes. The same qualitative result holds when neurons spike at unequal rates drawn from a zipf distribution. The paper further claims that the nD recovers within about 1,000 inputs after an abrupt change in the generating neurons, that its accuracy falls to k-means-like levels when the search update is disabled, and that applying the search increment randomly only one sixteenth of the time leaves accuracy essentially unchanged.

Load-bearing premise

The comparison depends on the claim that hyperparameters tuned on one or two (neuron count, instance deviation) settings also give near-optimal accuracy for all other settings tested; if that correlation claim fails, the accuracy advantage reported for the dendrite over k-means could be an artifact of tuning rather than a property of the algorithm.

Editorial extensions

If this is right

With fixed, roughly 5-bit precision and no multiplications, the nD could be implemented in very small, low-power neuromorphic hardware for real-time spike sorting.
The search mechanism is load-bearing: switching it off reduces nD accuracy to about k-means level, so any hardware implementation must preserve some form of slow background weight growth.
Because overprovisioning the number of clusters and then merging them does not degrade potential accuracy, online spike sorters need not know the number of neurons in advance.
With zipf-distributed firing rates, using fewer clusters than neurons raises average accuracy because rare, hard-to-sort neurons are simply not assigned their own cluster.

Reading between the lines

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

The decisive test this paper leaves open is a full hyperparameter sweep per condition; if the reported advantage shrinks when each neuron-count and instance-deviation pair is tuned independently, the advantage may be a tuning artifact rather than a property of the algorithm.
Because the dendrite unit is defined at the sub-neuron level, a neuron with multiple dendrites could cluster several feature streams in parallel, an extension the paper gestures toward but does not simulate.
The same capture/backoff/search update could be applied to online clustering of other streaming feature data, such as gesture or sensor signatures, provided the features are discretized to a small integer range.

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.

Desk Editor's Note

read the letter

Colleague,

Bottom line: this is a clearly written simulation study of a simple online clustering rule inspired by active dendrites, applied to spike sorting. The claim that the nD beats offline k-means is plausible but rests on a comparison that is less tightly controlled than the prose suggests.

What is genuinely new: the nD update rule with capture/backoff/search and similarity coding is a clean, low-precision formulation. The spike-sorting application is appropriate, and the paper includes a solid accuracy metric, a search-off ablation, a probabilistic search variant, and an operation-count analysis. The search-off ablation in §6.4 is the most convincing part: it shows that the search mechanism is responsible for the improvement at low instance deviations. The adaptability demonstration is a nice qualitative illustration.

Where the soft spots are: §6.2 is the main one. Hyperparameters are chosen by optimizing accuracy for one or two points, and the paper asserts that these choices transfer because 'high correlations in the accuracy curves' exist. That correlation is not shown. I agree with the reader's stress-test note: this is a real risk. A fair comparison would tune k-means on the same data or show that a grid of nD parameters yields the same qualitative advantage. The absence of variance or confidence intervals for any of the reported means (all are averaged over 16 seeds) makes it impossible to assess robustness. There is also only one baseline, offline k-means, and no code or data to reproduce the numbers. These are not fatal flaws—the mechanism holds together and the qualitative pattern appears across all neuron counts—but they are real limitations. I also note the paper is honest about its scope: synthetic spikes only, and no detailed hardware analysis.

Who it is for: neuromorphic hardware engineers and spike-sorting algorithm developers will get value from the nD formulation and the cost analysis. A serious referee should not desk-reject it.

My recommendation: send it to peer review, but ask the authors to add error bars, at least one more baseline (e.g., an online/streaming clustering method), and either release code/data or describe the hyperparameter selection more rigorously. The central idea is worth engaging with.

Referee Report

4 major / 6 minor

Summary. The paper proposes a neuromorphic dendrite (nD), an online clustering unit inspired by active dendrites, and presents a machine-learning-style formulation with integer weights updated via capture, backoff, and search operations. The nD is evaluated on a synthetic spike-sorting benchmark where six-feature action-potential shapes are generated from a canonical waveform with controllable base and instance deviations. The central empirical claim is that the nD, operating in a single online pass, outperforms an offline k-means baseline at small instance deviations and matches it at larger deviations, while also offering adaptation to stream changes, lower operation counts, and a probabilistic variant that reduces search cost. The paper includes ablations (search on/off), a probabilistic search implementation, non-uniform firing-rate experiments, and an analysis of overprovisioning cluster counts.

Significance. If the performance claim survives closer scrutiny, the paper makes a useful contribution to neuromorphic online clustering: it provides a self-contained algorithmic description, demonstrates a concrete application domain (spike sorting), and supports the mechanism with a search-ablation experiment. The operation-count analysis is a strength, and the adaptability illustration, though anecdotal, suggests a capability that offline methods lack. The main value lies in the algorithm's potential for low-precision, addition-only hardware, not in any deep theoretical result. However, the evaluation methodology currently falls short of supporting the strong comparative claim: hyperparameters are tuned on the same benchmark used for the comparison, no variance or confidence information is reported, and the baseline is limited to a single k-means variant. These issues must be addressed before the claimed advantage can be considered established.

major comments (4)

[Section 6.2, Table 1] The hyperparameter selection procedure is not controlled and is load-bearing for the main comparison in Figure 16. The paper states that capture/backoff values were chosen by 'simulation sweeps with subjective assessment' and by 'optimizing accuracy for one data point,' relying on the assertion that 'because of the high correlations in the accuracy curves, optimal hyperparameters for one data point are also optimal (or nearly so) for many of the others.' No evidence for this correlation is presented, and no separate validation set or held-out regime is used. Since k-means is run with fixed conditions and is not given any comparable hyperparameter tuning, the nD advantage at small instance deviations could be an artifact of tuning to that region. Please provide a principled protocol: for example, tune on one half of the deviation range and test on the other, or show full accuracy surfaces over the hyperparameter grid demonstrating that the chosen values are near-optimal across the range.
[Figure 16 and Section 5] The central quantitative comparison reports only mean accuracies over 16 rng seeds, with no standard deviations, error bars, or per-seed distributions. The paper itself notes in Section 5 that k-means can converge to local optima that differ significantly in quality depending on the seed, and the nD could be similarly seed-sensitive. Without measures of spread or paired statistical tests, it is impossible to judge whether the apparent nD advantage at small instance deviations is robust or driven by a few runs. Please report per-seed results, confidence intervals, or a paired test (e.g., Wilcoxon signed-rank) for each configuration in Figure 16.
[Abstract and Section 6.3] The abstract claims 'the dendrite outperforms k-means,' but the results in Section 6.3 show that nD outperforms k-means only for the smallest four instance deviations, while for larger deviations the accuracies are described as 'virtually the same.' This overstates the empirical finding. Either qualify the central claim to specify the low-deviation regime, or provide statistical evidence that the difference is significant across a broader range. The current phrasing could mislead readers about the scope of the demonstrated improvement.
[Section 5 and Section 6.3] The comparison is restricted to a single baseline, k-means, on synthetic clusters that the paper itself describes as 'spherical' and well-suited to k-means. While k-means is a reasonable reference point, the conclusion that 'nD out-performs k-means' would be considerably strengthened by at least one additional baseline—for example, an online variant of k-means, a Gaussian mixture model, or an adaptive streaming clustering method. If the authors intend the claim to be specific to k-means only, that scope should be stated explicitly in the abstract and conclusions. As it stands, the paper's title and framing suggest a more general clustering claim.

minor comments (6)

[Figure 6 and Section 4.2] Figure 6 is described inconsistently: the text says 'Figure 6a is the result of experiments with pyramidal neurons [1]' and 'Figure 6b shows ... variation ... [13]', but the caption says 'a) from Figure 2C in [13] b) from Figure 3B in [1]'. Please correct the mismatch.
[Section 6.2] The phrase 'subjective assessment' is not reproducible. Please define the sweep grid (ranges, step sizes) and the criterion used to select hyperparameters, or provide the accuracy surfaces so readers can verify the selection.
[Section 6.5, Figure 18] The claim that the probabilistic search results are 'virtually identical' is not quantified. Report the maximum difference or a similarity metric (e.g., mean absolute difference across all configurations) to support this equivalence.
[Section 7.2] The observation that fewer clusters than neurons can increase accuracy is attributed to the zipf distribution's least-frequent neurons. This is plausible, but the explanation would benefit from a table or explicit per-neuron accuracy breakdown for the 8-neuron case.
[Section 6.6] The operation-count formulas assume a specific implementation of similarity coding and bypassing. It would be helpful to clarify whether the quoted 'total of 287 low precision additions' includes all four sub-functions and how bypassing accounts for weights already at boundaries.
[References] Reference [11] appears to be an earlier version of this work by the same author; please cite it as such (e.g., as a preprint or prior conference paper) rather than as a general reference.

Circularity Check

1 steps flagged · score 3.0 of 10

nD advantage is partially in-sample: hyperparameters are tuned on the same synthetic benchmark used for the k-means comparison, though no formal derivation-loop or load-bearing self-citation is present.

fitted input called prediction [Section 6.2 (Hyperparameters), Table 1, compared with Figure 16]
"Hyperparameters were determined via simulation sweeps with subjective assessment of the results. ... The sweep process consisted of selecting a specific neuron count and instance deviation and optimizing accuracy for the one data point. Because of the high correlations in the accuracy curves, optimal hyperparameters for one data point are also optimal (or nearly so) for many of the others."

The capture/backoff values (3/2 for small instance deviations, 4/1 for large ones) are selected by optimizing accuracy on the same synthetic benchmark that is later used to claim that 'the dendrite outperforms k-means' (Figure 16). The paper asserts that the chosen values transfer across the tested range based on 'high correlations in the accuracy curves,' but no held-out regime or correlation evidence is provided. Since k-means is not given comparable hyperparameter tuning (it uses fixed Lloyd iterations and a known k), the reported nD advantage at small instance deviations is not a clean, parameter-free out-of-sample prediction; it is a post-selection comparison on the evaluation set.

full rationale

The nD inference and update equations (Sections 3.2-3.4) are not derived from the labeled benchmark; clustering operates without labels, so there is no self-definitional loop in which the output equals the input by construction. The comparison with k-means is empirical, and the k-means baseline is a standard offline method. The self-citations [9][10][11] describe the author's broader spiking-neuron architecture but are contextual, not load-bearing for the current accuracy results. The main circularity concern is confined to Section 6.2: capture and backoff are chosen by optimizing accuracy for one data point on the same synthetic benchmark used to demonstrate superiority over k-means, and the asserted transfer of optimal hyperparameters across the whole instance-deviation range is unsupported by shown evidence. The search-off ablation (Section 6.4) and probabilistic-search comparison (Section 6.5) are independent controls that reduce the severity of this issue, and the paper explicitly calibrates no free parameters against external benchmarks. Thus the work is not fundamentally circular, but the headline 'nD outperforms k-means' claim is partially in-sample and should be read with that qualification.

Assumptions & free parameters 7 free parameters · 4 assumptions · 1 invented entities

The central comparison rests on several hand-set hyperparameters and on the synthetic spike model. None of these are derived from first principles or externally validated, so the reported performance is conditional on the benchmark choices.

free parameters (7)

capture = 3 (small deviations), 4 (large deviations)
Weight increment for winning template's input feature values; set by simulation sweeps with subjective assessment (Section 6.2, Table 1).
backoff = 2 (small deviations), 1 (large deviations)
Weight decrement applied to non-input feature values in the winning template; tuned in the same sweeps (Section 6.2, Table 1).
search = 1/16 (or 1 applied 1/16 of the time)
Weight increment for losing templates up to wbase; tuned by sweeps, and the probabilistic variant is introduced to reduce cost (Sections 3.3 and 6.5).
wmax = 32
Maximum weight; fixed after preliminary study rather than swept (Section 6.2, Table 1).
wbase = 28
Cap for search-driven weights; fixed after preliminary study (Section 6.2, Table 1).
radius (r) = 3
Similarity coding range around each feature value; fixed after preliminary study (Section 6.2, Table 1).
base deviation = 0.375
Standard deviation used to generate base neuron shapes; chosen because synthetic waveforms looked similar to experimental waveforms (Section 4.4).

assumptions (4)

domain assumption Spikes emitted by the same neuron have similar extracellular waveforms, so waveform clustering is a valid basis for spike sorting.
This is the premise of the benchmark (Section 1) and is known to be imperfect in real data, as the paper itself notes via Stratton et al. [12].
domain assumption Synthetic spike shapes generated as independent normal deviations from a canonical six-feature shape represent realistic neuron-to-neuron and spike-to-spike variability.
Section 4.3 and Figure 8; this representativeness is asserted, not tested against real extracellular recordings.
ad hoc to paper Hyperparameters tuned on one or two points remain near-optimal across the tested range of neuron counts and instance deviations.
Section 6.2 asserts high correlations in accuracy curves without statistical support; the main comparison relies on this.
domain assumption The number of clusters is set to the known number of neurons in the main comparisons.
Sections 5 and 6; the mismatched-count case is only explored later with an oracle for merging.

invented entities (1)

neuromorphic dendrite (nD)
purpose: Online clustering building block that assigns feature vectors to cluster IDs using integer template weights and capture/backoff/search updates.
Introduced in Section 3 and demonstrated only on synthetic simulations; no real-data validation or independent replication is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Neuromorphic Online Clustering and Its Application to Spike Sorting." pith.science (2026). https://pith.science/paper/AUYTC7V6

@misc{pith2026250612555,
  author       = {Pith},
  title        = {Pith review of: Neuromorphic Online Clustering and Its Application to Spike Sorting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AUYTC7V6}},
  note         = {Machine review of arXiv:2506.12555}
}

read the original abstract

Active dendrites are the basis for biologically plausible neural networks possessing many desirable features of the biological brain including flexibility, dynamic adaptability, and energy efficiency. A formulation for active dendrites using the notational language of conventional machine learning is put forward as an alternative to a spiking neuron formulation. Based on this formulation, neuromorphic dendrites are developed as basic neural building blocks capable of dynamic online clustering. Features and capabilities of neuromorphic dendrites are demonstrated via a benchmark drawn from experimental neuroscience: spike sorting. Spike sorting takes inputs from electrical probes implanted in neural tissue, detects voltage spikes (action potentials) emitted by neurons, and attempts to sort the spikes according to the neuron that emitted them. Many spike sorting methods form clusters based on the shapes of action potential waveforms, under the assumption that spikes emitted by a given neuron have similar shapes and will therefore map to the same cluster. Using a stream of synthetic spike shapes, the accuracy of the proposed dendrite is compared with the more compute-intensive, offline k-means clustering approach. Overall, the dendrite outperforms k-means and has the advantage of requiring only a single pass through the input stream, learning as it goes. The capabilities of the neuromorphic dendrite are demonstrated for a number of scenarios including dynamic changes in the input stream, differing neuron spike rates, and varying neuron counts.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages

[1]

Spike sorting: new trends and challenges of the era of high-density probes

Buccino, Alessio P., Samuel Garcia, and Pierre Yger. " Spike sorting: new trends and challenges of the era of high-density probes." Progress in Biomedical Engineering 4, no. 2 (2022)

work page 2022
[2]

Accuracy of tetrode spike separation as determined by simultaneous intracellular and extracellular measurements

Harris, Kenneth D., Darrell A. Henze, Jozsef Csicsvari, Hajime Hirase, and Gyorgy Buzsaki. " Accuracy of tetrode spike separation as determined by simultaneous intracellular and extracellular measurements." Journal of neurophysiology 84, no. 1 (2000): 401-414

work page 2000
[3]

Why Neurons Have Thousands of Synapses, A Theory of Sequence Memory in Neocortex,

Hawkins, Jeff and S. Ahmad, “ Why Neurons Have Thousands of Synapses, A Theory of Sequence Memory in Neocortex,” Numenta, Inc, Redwood City, California, Oct. 30, 2015

work page 2015
[4]

Least squares quantization in PCM

Lloyd, Stuart P. , "Least squares quantization in PCM" , IEEE Transactions on Information Theory 28 (1982): 129–137

work page 1982
[5]

Optimization of a multielectrode array (MEA) -based approach to study the impact of Aβ on the SH-SY5Y cell line

Mateus, José Carlos Barreiro. " Optimization of a multielectrode array (MEA) -based approach to study the impact of Aβ on the SH-SY5Y cell line " Master's thesis, Universidade de Aveiro (Portugal), 2015

work page 2015
[6]

A logical calculus of the ideas immanent in nervous activity

McCulloch, Warren S., and Walter Pitts. " A logical calculus of the ideas immanent in nervous activity ." The bulletin of mathematical biophysics 5, no. 4 (1943): 115-133

work page 1943
[7]

Are biological systems poised at criticality? Journal of Statistical Physics

Mora T, Bialek W. Are biological systems poised at criticality? Journal of Statistical Physics. 2011;144:268– 302

work page 2011
[8]

Past, present and future of spike sorting techniques

Rey, Hernan Gonzalo, Carlos Pedreira, and Rodrigo Quian Quiroga. " Past, present and future of spike sorting techniques." Brain research bulletin 119 (2015): 106-117

work page 2015

Show all 14 references

[9]

A Macrocolumn Architecture Implemented with Spiking Neurons

Smith, James E., “A Macrocolumn Architecture Implemented with Spiking Neurons” arXiv:2207.05081 (2023)

2023 arXiv
[10]

Implementing Online Reinforcement Learning with Temporal Neural Networks

Smith, James E. "Implementing Online Reinforcement Learning with Temporal Neural Networks." arXiv:2204.05437 (2022)

2022 arXiv
[11]

Neuromorphic Online Clustering and Classification

Smith, James E. “Neuromorphic Online Clustering and Classification” arXiv:2310.17797v1 (2023)

2023 arXiv
[12]

Action potential waveform variability limits multi-unit separation in freely behaving rats

Stratton, Peter, Allen Cheung, Janet Wiles, Eugene Kiyatkin, Pankaj Sah, and François Windels. " Action potential waveform variability limits multi-unit separation in freely behaving rats." PloS one 7, no. 6 (2012)

2012
[13]

Analysis of extracellular spike waveforms and associated receptive fields of neurons in cat primary visual cortex

Sun, Shi H., Ali Almasi, Molis Yunzab, Syeda Zehra, Damien G. Hicks, Tatiana Kameneva, Michael R. Ibbotson, and Hamish Meffin. "Analysis of extracellular spike waveforms and associated receptive fields of neurons in cat primary visual cortex." J Physiol 599 (2021): 2211-2238

2021
[14]

The effect of nonstationarity on models inferred from neural data

Tyrcha J, Roudi Y, Marsili M, Hertz J. The effect of nonstationarity on models inferred from neural data. Journal of Statistical Mechanics: Theory and Experiment. 2013; p. 03005

2013

Pith tools

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

[1] [1]

Spike sorting: new trends and challenges of the era of high-density probes

Buccino, Alessio P., Samuel Garcia, and Pierre Yger. " Spike sorting: new trends and challenges of the era of high-density probes." Progress in Biomedical Engineering 4, no. 2 (2022)

work page 2022

[2] [2]

Accuracy of tetrode spike separation as determined by simultaneous intracellular and extracellular measurements

Harris, Kenneth D., Darrell A. Henze, Jozsef Csicsvari, Hajime Hirase, and Gyorgy Buzsaki. " Accuracy of tetrode spike separation as determined by simultaneous intracellular and extracellular measurements." Journal of neurophysiology 84, no. 1 (2000): 401-414

work page 2000

[3] [3]

Why Neurons Have Thousands of Synapses, A Theory of Sequence Memory in Neocortex,

Hawkins, Jeff and S. Ahmad, “ Why Neurons Have Thousands of Synapses, A Theory of Sequence Memory in Neocortex,” Numenta, Inc, Redwood City, California, Oct. 30, 2015

work page 2015

[4] [4]

Least squares quantization in PCM

Lloyd, Stuart P. , "Least squares quantization in PCM" , IEEE Transactions on Information Theory 28 (1982): 129–137

work page 1982

[5] [5]

Optimization of a multielectrode array (MEA) -based approach to study the impact of Aβ on the SH-SY5Y cell line

Mateus, José Carlos Barreiro. " Optimization of a multielectrode array (MEA) -based approach to study the impact of Aβ on the SH-SY5Y cell line " Master's thesis, Universidade de Aveiro (Portugal), 2015

work page 2015

[6] [6]

A logical calculus of the ideas immanent in nervous activity

McCulloch, Warren S., and Walter Pitts. " A logical calculus of the ideas immanent in nervous activity ." The bulletin of mathematical biophysics 5, no. 4 (1943): 115-133

work page 1943

[7] [7]

Are biological systems poised at criticality? Journal of Statistical Physics

Mora T, Bialek W. Are biological systems poised at criticality? Journal of Statistical Physics. 2011;144:268– 302

work page 2011

[8] [8]

Past, present and future of spike sorting techniques

Rey, Hernan Gonzalo, Carlos Pedreira, and Rodrigo Quian Quiroga. " Past, present and future of spike sorting techniques." Brain research bulletin 119 (2015): 106-117

work page 2015