Technical Program

Paper Detail

Paper: PS-1B.40
Session: Poster Session 1B
Location: Symphony/Overture
Session Time: Thursday, September 6, 18:45 - 20:45
Presentation Time:Thursday, September 6, 18:45 - 20:45
Presentation: Poster
Publication: 2018 Conference on Cognitive Computational Neuroscience, 5-8 September 2018, Philadelphia, Pennsylvania
Paper Title: The impact of noise correlation on multivariate pattern classification in fMRI
Manuscript:  Click here to view manuscript
Authors: RUYUAN ZHANG, Kendrick Kay, University of Minnesota, United States
Abstract: Previous theoretical work has shown that noise correlations (NC) that have the same sign as signal correlations (SC) limit the amount of information encoded in neural population responses. A recent fMRI study replicated this finding in functional magnetic resonance imaging (fMRI) data (Bejjanki et al., 2017). Here, we aim to gain further understanding of how NCs between voxels alter the accuracy of multivariate pattern classification (MVPC), a popular analysis method in fMRI research. In a simulated fMRI orientation experiment, voxel responses were simulated using an encoding model and then classified using linear discriminant analysis. We evaluated two forms of NC: one proportional (i.e., tuning-compatible NC) to the SC between voxels and the other one independent (i.e., tuning-independent NC) of the SC between voxels. Surprisingly, our results show that both the tuning-compatible NCs and the tuning-independent NCs improve MVPC accuracy. We show that these results stem from two major factors: (1) classifiers can “greedily” select voxels that have informative covariance and (2) the SC defined using tuning for two orientations is different from the SC defined using full orientation tuning curves. Taken together, our results provide a theoretical foundation for understanding the effect of NC on MVPC in future experimental studies.