PaperlayerView exact revision on arXiv

Source-linked AI summary

Axonal delay dispersion decides whether a neuron detects an event or a sequence, and predicts cortical column diameter

Cheng Bi, Jipeng Sun

arXiv:2609.04195v2q-bio.NCcs.NE

TL;DR

Sparse firing makes temporal coding necessary, yet prior accounts do not explain which temporal feature a neuron becomes selective for. The paper models delay-compensated matching on an integrator neuron and finds that delay dispersion selects event versus sequence detection, with consequences for coding limits and cortical geometry.

  • Problem

    Although conduction delays can convert firing order into synchrony, prior accounts do not explain what governs whether a neuron detects a coincident event or an ordered sequence.

    Source: §1 · p. 2

  • Method

    The framework treats the delays converging on a dendritic branch as a physical key for matching input spike-time differences, with the integrator neuron checking the match against separate content.

    Source: §1 · p. 3, §2.1 · p. 6

  • Results

    Delay dispersion moves populations from event to order-selective sequence detection, maps that distinction to myelinated versus unmyelinated projections, and bounds sequence coding at roughly the gamma period.

    Source: §1 · pp. 3–4

  • Takeaways & Limitations

    Because axonal length and myelination set the delay dispersion, the framework makes detector class and code limits testable against projection delays, coincidence windows, and column geometry.

    Source: §3.8 · p. 22

    The framework has not yet been tested against recorded neural data.

    Source: §3.7 · p. 21

Abstract

from arXiv · show

Cortical neurons fire sparsely -- often fewer than one spike per sensory window -- making rate coding insufficient and temporal coding a necessity. That conduction delays convert firing order into synchrony is long established. What governs which class of temporal feature a neuron detects -- one volley of coincident input, or two in a particular order -- has not been examined. We propose a delay-signature framework in which the axonal conduction delays converging on a dendritic branch constitute a physical key: only input sequences whose spike-time differences the delays compensate arrive synchronously, and coincidence detection, via calcium plateau thresholds, converts that synchrony into an all-or-none output. In simulations of an integrator-neuron model we report three results. First, a single physical scalar -- the dispersion of the delay set -- moves a population from event detection to order-selective sequence detection. The transition is emergent under random delays and connectivity: at narrow dispersion sequence detectors do not exist, and the dispersion at which they overtake event detectors tracks the inter-event interval with a slope statistically indistinguishable from one. This maps a computational distinction onto the anatomical one between myelinated and unmyelinated projections, making myelination a switch on what a neuron computes, not only a regulator of speed. Second, the same dispersion sets the code's limits: it bounds the longest codable interval and fixes an absolute timing tolerance of about a millisecond, with slowing better tolerated than speeding. Third, that millisecond window and horizontal conduction velocity together predict cortical column diameter, and the two areas with direct measurements fall where the relation puts them. One anatomically measurable parameter thus sets what a neuron detects and the limits of what it can represent.

1. Introduction

Sparse firing makes temporal coding necessary, and although conduction delays are known to turn firing order into synchrony, prior accounts do not explain which temporal feature a neuron detects. The paper identifies delay dispersion as a physical parameter that selects between event and sequence detection and also bounds what the code can represent.

1. Introduction

  • Sparse cortical firing makes temporal coding necessary, and established delay-line accounts explain how firing order becomes synchrony but not which temporal feature a neuron detects.Prior models show sequence-selective readout but do not identify a physical parameter that decides between event and sequence detection.Source: §1 · pp. 2–3
  • The delay set’s dispersion maps the event-versus-sequence distinction onto myelinated and unmyelinated projections, while a calcium threshold supplies a silent, tuned subpopulation and sequence coding is bounded at roughly the gamma period.The downstream spike count accompanying the silent subpopulation is reported in S3 Fig.Source: §1 · pp. 3–4
  • A branch’s delay signature is a physical configuration that matches pairwise arrival differences, so it selects an equivalence class of input patterns rather than one pattern.The framework contrasts fixed anatomical delays with synaptic matching and accounts in which delays are optimization variables.Source: §1 · p. 3
  • A delay set spanning Δ cannot compensate for spike-time separations longer than Δ, making the same parameter that supports selectivity a bound on sequence duration.Selectivity and its limit are two consequences of the same mechanism.Source: §1 · p. 3
  • Delay dispersion moves populations from event detection to order-selective sequence detection under random delays and connectivity, while also setting timing tolerance and the longest codable interval.The transition is emergent, and the same dispersion bounds the longest interval the code can represent.Source: §1 · p. 4

Box 1. Framework foundations

The framework rests on four established observations linking sparse firing, fixed conduction delays, myelination, and narrow coincidence detection to their computational consequences.

Box 1. Framework foundations

  • Sparse cortical firing makes rate coding insufficient for the information demands described, leaving temporal coding as the necessary alternative.The passage also notes sparse yet temporally precise codes across modalities.Source: Box 1. Framework foundations · p. 4
  • Because conduction delays are physically fixed and myelination sets their dispersion Δ, learning selects pathways from a redundant projection pool rather than tuning delay values.Source: Box 1. Framework foundations · p. 4
  • A millisecond coincidence window requires a calcium threshold and fast feed-forward inhibition, with sparse activation as its visible signature.Source: Box 1. Framework foundations · p. 4

2. Results

The results examine how axonal-delay dispersion shapes coincidence-based detection and coding limits, and how conduction velocity relates to cortical column diameter. Predictions for V1 and mouse barrel cortex are of the same order as observed sizes, based on only two test points.

2.1. The delay-signature principle and the integrator neuron

  • Sequence decoding requires sparse firing and a temporal reference frame, while a narrow coincidence window lets local inhibition sharpen synchrony detection.The framework adopts τw ≈ 1 ms; widening the window reduces selectivity between nearby intervals, while narrowing it increases the timing precision required of conduction delays.Source: §2.1 · pp. 5–6
  • An integrator neuron matches an upstream spike sequence against branch-specific conduction delays, then combines distal matching with proximal content to determine its output.In layer-5 distal dendrites, coincident arrival can trigger a calcium plateau and local potentiation, while chronically mismatched branches are weakened and pruned. Stable delay structures can also arise from relay neurons or other physical processes, not only axonal conduction.Source: §2.1 · pp. 5, 7

2.2. Delay-signature dispersion sets detector class

  • Increasing delay dispersion shifts populations from event to sequence detectors: classes cross at Δ = 16.4 ms, with crossover slope 0.92 [0.83, 1.01] versus interval.At Δ = 2 ms, sequence detectors are 0.00% [0.00, 0.00], while 9.76% respond to a single event. At Δ = 18.6 ms, sequence detectors peak at 0.180% [0.110, 0.255]; 99.0% of classified detectors in the Δ = 14–30 ms range are silent to the reversed sequence.Source: §2.2 · pp. 7, 9, Figure 2 · p. 10

2.4. A scaling prediction: column diameter tracks conduction velocity

  • With orthogonal axon crossings, the model predicts cortical column diameters of 300–600 μm from horizontal conduction at 0.3 m/s and a 1–2 ms coincidence window.Because afferent velocities vary, column boundaries may be graded over a distance comparable to that velocity spread; the proposed relation between velocity uniformity and boundary sharpness has yet to be tested against anatomy.Source: §2.4 · p. 13

2.3. The same dispersion sets the codable interval and the precision it requires

  • The longest sequence interval recognized tracks delay dispersion, while shorter-than-reference intervals produce a broad population response.The upper limit is roughly Δ and barely shifts across thresholds: 14 ms for Δ = 10 ms, 22 ms for Δ = 20 ms, 31 ms for Δ = 30 ms, and 48 ms for Δ = 50 ms. For shorter intervals, 51–54% respond at any given interval, but only 1.8–9.2% respond at every shorter interval tested when Δ = 20 ms, falling to 0.4–3.1% when Δ = 30 ms.Source: §2.3 · pp. 11–12, Figure 3 · p. 14
  • Timing tolerance is about a millisecond in absolute terms, and speeding conduction impairs sequence-detector retention more than comparable slowing.At Δ = 20 ms, 87% [86, 88] survive 2% velocity jitter, versus 61% [59, 63] at 5% and 43% [42, 44] at 10%. A 5% slowing retains 59% [57, 60] against 48% [46, 49] for a 5% speeding; tolerated absolute error is 1.46 ms at θ = 2.5 and 0.82 ms at θ = 4.5, while tolerated fractional error falls as 1/Δ. The stricter gate changes the constant but not the asymmetry or 1/Δ relationship.Source: §2.3 · pp. 12–13, Figure 3 · p. 14

2.4. A scaling prediction: column diameter tracks conduction velocity

  • At 0.33 m/s, V1 predicts 330–660 μm columns versus observed ∼400 μm; at 142 ± 76 μm/ms, mouse barrel cortex predicts 142–284 μm versus ∼200–300 μm.The velocities are from other laboratories, not fitted to the column measurements.Source: §2.4 · p. 15
  • The two areas differ 2.3-fold in conduction velocity and their column sizes vary in the predicted direction, but two points do not establish a regression.More areas with horizontal conduction measured in the layer receiving the projection are needed to test the scaling.Source: §2.4 · p. 15
  • Accounting for uncertainty in conduction velocity and afferent crossing angle broadens the predicted diameter range to 100–1000 μm, still the same order as measured column sizes.The horizontal direction used in the construction is the slower one: interlaminar conduction was ∼71% faster than intralaminar conduction.Source: §2.4 · p. 15

3. Discussion

The discussion frames delay dispersion as a determinant of temporal selectivity and its limits, with consequences for state sensitivity and cortical column geometry. It also relates the framework to other temporal models and identifies important assumptions and untested claims.

3. Discussion

  • Delay dispersion selects which temporal feature a neuron detects, while the same delay-compensation construction maps firing latency onto column position and predicts column diameter.The discussion treats axonal delay dispersion as a measurable anatomical quantity linking sequence decoding with column geometry.Source: §3 · p. 15, Figure 4 · p. 16

3.1. State sensitivity as a test of delay compensation

  • A 5% slowing retains 59% of matched detectors, versus 48% for a 5% speeding, an asymmetry proposed as a test of delay compensation.The intact-brain prediction is an isolated-component estimate because temperature and neuromodulatory state also affect excitability, synaptic efficacy, and homeostatic compensation.Source: §3.1 · pp. 15, 17

3.2. The hard threshold: selectivity, sparsity, and failure under jitter

  • Hard calcium gating yields population selectivity of 0.703 on clean input, versus 0.020 for soft gating, and 0.375 at 5 ms jitter.The hard threshold supports a silent tuned subpopulation, but does not itself confer jitter tolerance or reject chance coincidences; jitter also erodes silence to reversed-order inputs.Source: §3.2 · p. 17

3.3. The codable bound and the gamma period

  • At a 20–30 ms delay spread, the longest codable interval approaches one gamma cycle, but that spread is inferred from general conduction-delay ranges rather than measured for horizontal fibres.The proposed relationship suggests that longer associations require cross-cycle phase locking or theta-nested rhythms.Source: §3.3 · p. 17

3.4. Relation to delay learning and to other temporal readouts

  • Learning selects among pre-existing physical delays rather than tuning them, distinguishing the framework from delay-learning models and polychronization.Activity-dependent myelination may partly retune the redundant delay pool, so selection and construction can differ by degree.Source: §3.4 · p. 18

3.5. Modes of circuit activation

  • Sparse input may make circuit startup difficult, so bursts are proposed to open channels through temporal compaction and facilitation; phase-locked spikes can then maintain matching.The framework interprets phase locking as supporting sustained delay compensation, rather than as an epiphenomenon of oscillation.Source: §3.5 · pp. 18–19

3.6. Predictive coding and Transformer attention

  • The attention analogy is structural rather than a claim that the brain implements a Transformer: delay signatures natively match order, whereas Transformer attention requires separate positional encoding.Branch-level dispersion makes order sensitivity local and graded; the framework also separates contextual index matching from completed content checked at the soma.Source: §3.6 · pp. 19–20

3.7. Limitations

  • The model simplifies dendritic geometry and gating, holds several parameters static, uses a short plateau, and omits membrane integration as a delay source in recurrent circuits.In local recurrent circuits, membrane integration could contribute comparably to effective delay; that dynamic extension is not modelled.Source: §3.7 · p. 20

Box 2. Testable predictions

The paper proposes tests linking delay dispersion to temporal tuning, coincidence-window width, and column diameter. Together, these predictions connect measurable neural and anatomical quantities to the framework's claims.

Box 2. Testable predictions

  • Delay dispersion should correlate with target temporal tuning, with narrower dispersion linked to sharper event responses and wider dispersion to broader sequence integration.A roughly 20–30 ms dispersion would also place the codable bound near the gamma period; the authors identify measuring real projection delay statistics as a test.Source: Box 2. Testable predictions · p. 22
  • Column diameter is predicted to scale as D ≈ v × τw; the two areas with direct velocity measurements are consistent, while a broader set of areas is needed for a quantitative test.Where both measurements are already published, the proposed test requires no new experiment.Source: Box 2. Testable predictions · p. 22
  • The framework predicts an effective coincidence window of about 1 ms in L4 stellate cells, where that window has not yet been measured.A window substantially broader than 2 ms in awake animals, together with no tuning change when the window is manipulated, would challenge the prediction.Source: Box 2. Testable predictions · p. 22

3.8. Conclusion

  • The proposed framework links measurable delay dispersion to what neurons detect, sequence duration, and timing precision, and identifies tests in projections, awake coincidence windows, and cortical geometry.The authors state that each proposed test can use techniques that currently exist.Source: §3.8 · p. 22, Box 2. Testable predictions · p. 22

4. Materials and methods

The methods describe discrete-time integrator-neuron simulations with randomized connectivity, delays, and stimulus conditions. Populations, functional selection rules, and seed-level summaries vary across experiments.

4. Materials and methods

  • Simulations used 1 ms steps, local dendritic integration and threshold-triggered plateaus, followed by somatic integration and spiking.Two-input simulations used a 5 ms membrane constant and 10 ms plateau; population simulations used 10 ms and 15 ms, respectively.Source: §4 · pp. 22–23
  • Each feature projected to each branch with probability 0.1, and connection delays were independently drawn uniformly from [0, Δ], making Δ the maximum delay.Unit weights made the calcium threshold correspond to a required number of coincident inputs; experiments also varied threshold conditions.Source: §4 · p. 23
  • Two-event stimuli used disjoint feature groups with within-volley offsets, while controls and retimings reused spike identities; jitter and spurious input were separately manipulated.Delay-precision experiments perturbed individual delays or scaled them together.Source: §4 · p. 23
  • Experiments used 20 independent seeds, with population sizes and trial counts varying by sweep; neurons were selected by firing behavior rather than connectivity.Reported means and 95% bootstrap intervals were computed over seeds, not trials or neurons.Source: §4 · pp. 23–24

Declaration of generative AI in the writing process

The authors report that AI-assisted tools were used only for manuscript language editing and formatting. They state that they reviewed the output and take responsibility for the publication.

Declaration of generative AI in the writing process

Supporting information

The supporting information expands on the framework’s gating, plasticity, and population behavior, while also documenting model extensions, comparisons, and limits. Several proposed extensions are not simulated.

S1 Fig.

  • Redundancy-dependent plasticity retains only the branch whose delays compensate sequence timing, producing synchronous arrival and a calcium plateau.The figure contrasts matched and unmatched branches on a shared time scale.Source: Supporting information · p. 24

S2 Fig.

  • S2 Fig. examines detection and false-positive rates across Poisson noise densities for 200-neuron pools with delays uniform on [0, 20] ms, alongside an analytic match-rate account.The figure title frames this as population coding and Q-sparsity dependence.Source: Supporting information · p. 24

S3 Fig.

  • Hard gating is thirty-fivefold more selective than soft gating at the population level, emits fewer downstream spikes, and alone has a silent baseline; 5 ms jitter roughly halves its selectivity.The figure also compares a two-input example with a leaky integrate-and-fire control whose trajectories superimpose across input orders.Source: Supporting information · pp. 24–25

S1 Text

  • Proposed links to V1 gamma rhythms, cortical feedback, and retinogeniculate topography are conjectural and were not simulated.Source: Supporting information · p. 25

S5 Text

  • A conjunctive soma reaches a veto index of 0.763 while preserving 100% matched firing; a weighted sum never exceeds 0.000 at any inhibition strength.The result supports consistency with the architecture, not that real pyramidal neurons implement a conjunctive soma.Source: Supporting information · p. 25

S3–S6 Text

  • Supplementary comparisons cover delay-line, dendritic, attention-based, and spiking-network accounts, predictive coding, and index–content separation with hierarchical Q-cycles.The latter is demonstrated with a reduced CA3 model and biological circuit examples.Source: Supporting information · p. 25

S7 Text

  • Parallel relays are proposed to generate delay signatures for features at variable sequence positions, with the cerebellar granular layer discussed as a candidate substrate; this was not simulated.Source: Supporting information · p. 25

S8 Text

  • Uncertainty in velocity, timing tolerance, and crossing geometry widens the column-diameter prediction to 100–1000 μm; the construction also predicts perpendicular boundaries and approximately circular cross-sections.Source: Supporting information · p. 26

Data availability

The authors make the code for reproducing reported numbers openly available. The package organizes simulations and figures into scripts and documents the exact runs behind reported outputs.

Code availability

  • Code needed to reproduce every reported number is openly available in the delaykey-main package.Source: Data availability · p. 26

Package contents

  • The package includes the integrator-neuron model, experiment scripts, figure scripts, and a README mapping manuscript experiments to script names.Source: Data availability · p. 26

Reproduction

  • A run_all.sh records exact invocations behind reported CSVs, while each figure is regenerated from those CSVs alone.Source: Data availability · p. 26

Funding

The authors report receiving no specific funding for this work.

Funding

Loading 2609.04195v2…