Abstract
Over the past decade, hundreds of scientific papers have attempted to decode the multi-voxel patterns underlying distinct perceptual and cognitive states. For the multi-voxel pattern investigator, a large number of methodological decisions are required, many of which can impact classification results. We present a meta-analysis, with two goals: 1.) To discover and quantify various influences on pattern detection from results across many studies, and 2.) To determine the neural regions implicated in representing different types and classes of visual entities. We collected all peer-reviewed papers from Google Scholar, PubMed, Web of Science, and Scopus that either included relevant sear-terms (e.g., "MVPA", "classification"), or cited a seminal study by Haxby et al. (2001). We employed inclusion criteria to reduce this set to papers examining multi-voxel patterns for visual items in the occipital and/or temporal cortices of healthy adults. By coding these papers on a series of method-related variables (e.g., voxel resolution, experimental design, classification technique), brain-related variables (e.g., region), and classification results, we can predict and then test which variables influence multi-voxel pattern discriminability, and quantify their influence. For example, within the set of method variables, classification accuracy is improved with a greater number of acquisition runs. Within the brain-related variables, patterns become less discriminable from posterior to anterior retinotopic regions. This meta-analysis will provide a comprehensive summary of the relevant published research, and also point to various methodological variables that can help or hinder attempts to decode neural representations in the human brain.
Meeting abstract presented at VSS 2014