TITLE:

SPATIAL-TEMPORAL COMMON MODES ANALYSIS OF NEUROPHYSIOLOGICAL DATA: AN INITIAL APPROACH (No. 159)

DATE:

Friday, August 31st, 2007

TIME:

3:30 PM

LOCATION:

GMCS 214

SPEAKER:

Mark E. Pflieger, Source Signal Imaging Inc.

ABSTRACT:

The spatiotemporal statistical organization of natural brain activity (reflected by neurophysiological measures such as scalp EEG, extracranial MEG, intracranial EEG, or local field potentials) may not conform to the ideal assumptions of some analyses, such as exclusive Gaussianity (PCA), exclusive non-Gaussianity (ICA), trial-to-trial homogeneity of phase-locked responses (event-related averages), additive non-phaselocked activity (event-related variances), and one-one correspondences between spatial and temporal modes (PARAFAC). Spatial-temporal common modes analysis (CMA, for short; a method currently under development) is a data-driven alternative which avoids these assumptions by seeking “most efficient” re-representations of multi-epoch data. Each epoch originally is represented as a matrix of measurements (such as electric potential differences or oriented magnetic field components) with rows corresponding to spatial channels (such as EEG electrodes using a common reference, or MEG sensors) and with columns corresponding to time points (such as latencies relative to stimulus or behavioral events). A change of spatial and temporal coordinates will alter the representation of each epoch so that, compared with the original representation, more or less information (as quantified by some measure) may be required to describe the signal. Thus, the problem of CMA is to find fixed spatial and temporal coordinate systems (basis vectors) that minimize the average information required to represent epochs in a collection. By solving this problem: (a) the complete dataset is represented most efficiently (whether statistics are Gaussian, non-Gaussian, or mixed); (b) the spatial and temporal mode vectors encode maximum common information across epochs (without averaging); (c) the minimum information representations retain essential trial-to-trial variations (without computing variances); and (d) the mutual information between spatial and temporal modes is maximized (without requiring exact one-one correspondences). Variants of the CMA problem depend on the specific measure of representational information, and whether modes are orthogonally constrained. Each variant requires that a continuous nonlinear constrained optimization problem be solved jointly for spatial and temporal matrices of basis vectors. An initial approach using a gradient projection algorithm will be presented.

HOST:

Jose Castillo

DOWNLOAD: