Email updates

Keep up to date with the latest news and content from BMC Neuroscience and BioMed Central.

This article is part of the supplement: Eighteenth Annual Computational Neuroscience Meeting: CNS*2009

Open Access Poster presentation

Estimation of spectro-temporal receptive fields based on linear support vector machine classification

Arne F Meyer1*, Max FK Happel2, Frank W Ohl2 and Jörn Anemüller1

Author Affiliations

1 Department of Physics, Carl-von-Ossietzky University, Oldenburg, DE, Germany

2 Leibniz Institute for Neurobiology and Institute of Biology, Otto-von-Guericke University, Magdeburg, DE, Germany

For all author emails, please log on.

BMC Neuroscience 2009, 10(Suppl 1):P147  doi:10.1186/1471-2202-10-S1-P147

The electronic version of this article is the complete one and can be found online at:

Published:13 July 2009

© 2009 Meyer et al; licensee BioMed Central Ltd.


The spectro-temporal receptive field (STRF) of a neuron is defined as the linear filter that, when convolved with the spectro-temporal representation of an arbitrary stimulus, gives a linear estimate of the evoked firing rate [1]. A common method for STRF estimation uses the spike-triggered average (STA) to compute the mean stimulus pattern preceding every spike.

Here, we present a method that not only considers stimulus patterns that evoke spikes but also those after which no spikes occur. This results in a binary classification problem. We show that the STRF model is equivalent to the structure of a linear support vector machine (SVM) and propose the use of SVMs for the estimation of the STRF. Based on this approach, we demonstrate that the obtained STRFs are a better predictor for spiking and non-spiking behavior of a neuron.

Methods and results

The SVM is trained using real spike data from anesthetized gerbils [2] and zebra finches [3]. The parts of the stimulus spectrogram preceding a spike are labeled as class 1, whereas the remaining (non spike-evoking) parts are labeled as class 0. We used 80% of the data for training and 20% for prediction. See figure 1

thumbnailFigure 1. The predicted spike rate (bold black line) compared to the actual response smoothed with a Gaussian window of 20 ms (thin red line). The upper panel shows the result for the SVM-based method, the lower panel for the STA method. The mean coherence between actual and predicted rate is 0.23 for both methods.


In comparison to classic STA estimation, the method proposed here is characterized by a notably finer structure in the temporal evolution of spike rate prediction. In particular the non spike-eliciting time intervals are better captured by the novel approach. This behavior is likely a result of the learning procedure employed that is based on a binary classification paradigm with a linear classifier. The averaging approach of the STA results in smoother estimates for the neuronal receptive field (due to the temporal low-pass envelope characteristics of natural stimuli), consequently producing less-detailed spike rate predictions.


  1. Theunissen FE, Sen K, Doupe AJ: Spectral-temporal receptive fields of nonlinear auditory neurons obtained using natural sounds.

    J Neurosci 2000, 20:2315-2331. PubMed Abstract | Publisher Full Text OpenURL

  2. Happel MFK, Müller SG, Anemüller J, Ohl FW: Predictability of STRFs in auditory cortex neurons depends on stimulus class.

    Interspeech 2008, 670. OpenURL

  3. STRFPAK Matlab Toolbox [] webcite