Abstract: Multi-voxel pattern analysis (MVPA), which is based on machine learning theories, has gained great popularity over the past years as a new approach for fMRI data analysis. By training a classifier, MVPA categorizes multi-voxel patterns tuned by different cognitive states. Compared to conventional voxel-wise methods, this new approach provide higher sensitivity for detecting cognitive representations in the brain. It opens up the possibility for “reading out” mental states of human beings from the non-invasive recordings of brain activities.This paper introduce the fundamental principles of MVPA and the basic realization procedures. Scientific questions that may be properly addressed with this new approach and potential problems in its applications are also discussed.

Key words: MVPA, representation, fMRI, classification