Maximum entropy decoding of multivariate neural spike trains

Scalable algorithms for decoding the information content of trains of action potentials fired by ensembles of neurons are of increasing interest for several reasons. Firstly, neurophysiological recording approaches such as multielectrode arrays (MEAs) and two-photon calcium imaging are now producing recordings from unprecedented numbers and densities of neurons simultaneously. Secondly, recent work in the area of brain-machine interfaces has opened up the opportunity of taking advantage of such algorithms for purposes such as prosthetic device control, perceptual readout and communication with tetraplegic and "locked-in syndrome" patients. One approach taken for decoding multivariate neural spike trains is Bayesian decoding (see e.g. [1]), where the stimulus s is decoded via the rule s = argmaxs {P*(r|s)P(s)/P*(r)}, where P* refers to a model response probability function fit using a "training" dataset. In the present work, we extend this approach to maximum entropy models of neural population response patterns [2,3]. To preserve decoder temporal resolution, we decode spatial patterns of spikes from time windows short enough that the responses of each cell can be considered binary. We use a training dataset to measure trial-averaged nth order correlations, then apply a numerical optimization algorithm to compute online the likelihood of observing each response pattern in the test dataset. We have examined the cross-validated performance of the algorithm by decoding patterns of activity in a 2D Ising Model in which stimulus dependent statistical structure has been imposed at different orders (see Figure 1). In addition to studying simulated data, we demonfrom Eighteenth Annual Computational Neuroscience Meeting: CNS*2009 Berlin, Germany. 18–23 July 2009