A comment on "A fast L_p spike alignment metric" by A. J. Dubbs, B. A. Seiler and M. O. Magnasco [arXiv:0907.3137]

TL;DR

This paper evaluates the L_p Victor-Purpura spike train metric for clustering zebra finch responses, finding that higher p-values slightly improve performance by approximating windowed coincidence detection.

Contribution

It demonstrates that the L_p Victor-Purpura metric with p>1 can modestly outperform the standard p=1 metric in spike train clustering tasks.

Findings

L_p metrics with p>1 outperform p=1 in clustering accuracy

Higher p-values approximate windowed coincidence detection

The metric's performance varies with data type and p-value

Abstract

Measuring the transmitted information in metric-based clustering has become something of a standard test for the performance of a spike train metric. In this comment, the recently proposed L_p Victor-Purpura metric is used to cluster spiking responses to zebra finch songs, recorded from field L of anesthetized zebra finch. It is found that for these data the L_p metrics with p>1 modestly outperform the standard, p=1, Victor-Purpura metric. It is argued that this is because for larger values of p, the metric comes closer to performing windowed coincidence detection.