Abstract
Understanding the spatio-temporal neural dynamics underlying the processing of information in complex visual scenes is critical to the study of high-level visual function, decision making, perception as well as many other brain functions (Gosselin&Schyns, 2002). To address this question we used a perceptually bi-stable natural scene (a segment from Dali's "Slave Market") together with the bubbles paradigm (Gosselin&Schyns, 2001; Smith et al. 2006) in an MEG experiment. On each trial we sampled regions of the image in five different 1-octave spatial frequency bands using Gaussian bubbles. The subjects report their perception of the scene in each trial. We performed an ICA-based source analysis on the MEG data. For each source, and at each time point, we quantified (using Shannon information) the strength of the relationship between pixel visibility and the MEG response (bandpassed 1-40Hz) for each pixel in each spatial frequency band. We performed Non-negative Matrix Factorization (NMF) on this set of pixel-MEG information images to obtain a parsimonious parts-based representation of the regions of the visual stimulus that modulate neural activity. Crucially, these components include both behaviorally relevant information (which can be identified by correlating pixel visibility directly with the behavioral response) as well as parts of the stimulus space that are not related to the subject's perceptual responses but are nevertheless represented in the brain. We used these components as spatial filters – applied to the single trial bubble masks – to reduce the dimensionality of the stimulus space and enable us to track directly the neural representations of these stimulus regions. Preliminary results reveal effects such as a lateralization of early neural representation of high spatial frequency behaviorally relevant information and demonstrate the feasibility of this methodology for rigorously quantifying the spatio-temporal dynamics of neural representations of complex natural images during behavior.
Meeting abstract presented at VSS 2014