Bayesian Blind Separation of Spike Signals from Noisy Mixtures

CESAR F. CAIAFA; ANDRZEJ CICHOCKI

Shririnkoen, JAPAN

Workshop; RIKEN Brain Science Institute Retreat 2009; 2009

RIKEN BSI

The detection of neural activity is a prerequisite for understanding many types of brain function. However, identifying individual activity of neurons is a difficult task mainly because the measurements are always contaminated with high level noise and, even using the most sofisticated microelectrodes, it is difficult to completely isolate the action potential produced for a single neuron specially when spike shapes are very similar among neurons. In this work, we developed a new approach for the separation of spike signals generated by several neurons in a local population from a set of measurements taken with an array of sensors (microelectrode array). After aplying a prefiltering to the measured signals (mixtures), the underlying sources are transformed to a very particular set of signals called spike trains which are very sparse. Our task is then reduced to solve a Blind Source Separation (BSS) problem with more sources than sensors (underdetermined case) and with high level additive noise which is a difficult problem not solved with previous techniques. We model sparsity through mixed-state random variables and we developed a Bayesian method for estimating sources by the maximization of the posterior distribution (MAP). We have evaluated our approach by applying it to synthetically generated data under several conditions.