Blind Source Separation of Sparse Overcomplete Mixtures and Application to Neural Recordings Michal Natora 1 , Felix Franke 2 , Matthias Munk 3 , and Klaus Obermayer 1,2 1 Institute for Software Engineering and Theoretical Computer Science, Berlin Institute of Technology, Germany {natora,ff,oby}@cs.tu-berlin.de 2 Bernstein Center for Computational Neuroscience, Berlin, Germany 3 Max Planck Institute for Biological Cybernetics, T¨ ubingen, Germany {matthias.munk}@tuebingen.mpg.de Abstract. We present a method which allows for the blind source sep- aration of sparse overcomplete mixtures. In this method, linear filters are used to find a new representation of the data and to enhance the signal-to-noise ratio. Further, “Deconfusion”, a method similar to the independent component analysis, decorrelates the filter outputs. In par- ticular, the method was developed to extract neural activity signals from extracellular recordings. In this sense, the method can be viewed as a combined spike detection and classification algorithm. We compare the performance of our method to those of existing spike sorting algorithms, and also apply it to recordings from real experiments with macaque mon- keys. 1 Introduction In order to understand higher cortical brain functions, an analysis of the simulta- neous activity of a large number of individual neurons is essential. One common way to acquire the necessary amount of neural activity data is to use acute extra- cellular recordings, either with electrodes or, more recently, with multi electrodes (e.g. tetrode arrays). However, the recorded data does not directly provide the isolated activity of single neurons, but a mixture of neural activity from many neurons additionally corrupted by noise. The signal of the neurons is represented by spikes, which have a length of up to 4ms and an occurrence frequency of up to 350 Hz. In order to maximize the information yield, one aims at recording from as many neurons as possible; the number of recording channels in acute recordings, however, is mostly limited to 4 (in tetrodes). Thus, these recordings represent a sparse and overcomplete mixtures of neural signals. The task of so called “spike sorting” algorithms is to reconstruct the single neural signals (i.e., spike trains) from these recordings. There are several reasons to favor realtime online sorting over offline sorting, although more methods are available in the latter category. For example, realtime online spike sorting techniques are particularly desired for conducting “closed loop” experiments and for brain interface devices. The approaches in realtime online sorting (see T. Adali et al. (Eds.): ICA 2009, LNCS 5441, pp. 459–466, 2009. c  Springer-Verlag Berlin Heidelberg 2009