Use of Pattern-Information Analysis in Vision Science: A Pragmatic Examination
Résumé
MultiVoxel Pattern Analysis (MVPA) is presented as a successful alternative to the General Linear Model (GLM) for fMRI data analysis. We report different experiments using MVPA to master several key parameters. We found that 1) different feature selections provide similar classification accuracies with different interpretation depending on the underlying hypotheses, 2) paradigms should be created to maximize both Signal to Noise Ratio (SNR) and number of examples and 3) smoothing leads to opposite effects on classification depending on the spatial scale at which information is encoded and should be used with extreme caution.