A probabilistic model of coherent spatiotemporal dynamics in neuronal data

We propose a representation for analyzing a collection of time series data that is suited to the detection and representation of temporal or spatial coherence structure among the elements of the collection. Our method is principally different from any of the existing method in that it focuses on the statistical structure of the “generators” of the observed discrete data. We introduce the concept “feature dimension” and the “structure shape” and use them to parameterize the family of generators. The generators are useful for finding coherences among time series without the need for correlating a base structure to the time series data. Data of this nature is commonplace in brain imaging modalities such as fMRI, EEG, and MEG. As a proof of principle, we use the observed values of voxels’ signal intensity in the case of fMRI to construct a system of generators for the observed values and show how our method reveals the coherent feature of the data.