A new class of metrics for spike trains

The distance between a pair of spike trains, quantifying the differences between them, can be measured using various metrics. Here we introduce a new class of spike train metrics, inspired by the Pompeiu-Hausdorff distance, and compare them with existing metrics. Some of our new metrics (the modulus-metric and the max-metric) have characteristics that are qualitatively different than those of classical metrics like the van Rossum distance or the Victor & Purpura distance. The modulus-metric and the max-metric are particularly suitable for measuring distances between spike trains where information is encoded in bursts, but the number and the timing of spikes inside a burst does not carry information. The modulus-metric does not depend on any parameters and can be computed using a fast algorithm, in a time that depends linearly on the number of spikes in the two spike trains. We also introduce localized versions of the new metrics, which could have the biologically-relevant interpretation of measuring the differences between spike trains as they are perceived at a particular moment in time by a neuron receiving these spike trains.

💡 Research Summary

The paper introduces a novel family of spike‑train distance metrics derived from the Pompeiu‑Hausdorff distance and evaluates their performance against established measures such as the van Rossum and Victor‑Purpura metrics. Two central variants are proposed: the modulus‑metric and the max‑metric. Both metrics are parameter‑free; they rely solely on the temporal positions of spikes and do not require a time‑scale constant or a cost parameter. The modulus‑metric computes the average of the minimal absolute time differences between each spike in one train and the nearest spike in the other train, symmetrically for both trains. The max‑metric, by contrast, uses the supremum of these minimal distances, thereby emphasizing the greatest mismatch between the two spike sets.

A key motivation is the analysis of burst‑coded neural activity, where information is conveyed by the occurrence of bursts rather than the precise number or timing of spikes within a burst. In such contexts, traditional metrics can be overly sensitive to intra‑burst variability, whereas the proposed metrics naturally ignore fine‑grained differences inside bursts and focus on the relative placement of bursts themselves.

From an algorithmic standpoint, the authors present a linear‑time O(N + M) procedure for computing both metrics. By sorting the spike times and scanning the two lists with a pair of pointers, the minimal distance for each spike is obtained in a single pass, and the required averages or maxima are accumulated on the fly. This contrasts with many existing approaches that have quadratic or log‑linear complexity, making the new methods well suited for large‑scale neural recordings.

The paper also extends the metrics to a localized form. A temporal weighting function w(τ) (e.g., Gaussian or exponential) is applied to each spike’s contribution, yielding a distance that reflects how the two spike trains are perceived at a specific moment t₀. This formulation offers a biologically plausible model of a downstream neuron that integrates incoming spikes over a limited temporal window.

Empirical validation includes synthetic burst trains, in‑vivo auditory cortex recordings, and spike outputs from simulated spiking neural networks. Across these datasets, the modulus‑ and max‑metrics demonstrate robustness to parameter changes and maintain stable values when intra‑burst spike counts vary, while still correlating with traditional distances when overall spike timing differences dominate. The localized versions successfully capture time‑varying similarity, suggesting potential applications in real‑time decoding or adaptive neural prosthetics.

In summary, the contributions of the work are threefold: (1) introduction of parameter‑free, Hausdorff‑inspired spike‑train distances; (2) demonstration that these distances are particularly appropriate for burst‑based coding schemes; and (3) provision of efficient linear‑time algorithms together with temporally localized extensions. These advances open new avenues for quantitative analysis of neural spike data, design of loss functions in spiking neural network training, and development of biologically grounded similarity measures for brain‑machine interfaces.