A Shared Space in Thought: Similarity Approaches to Analyzing fMRI Data

Abstract

Most experiments in psychology rely on stimuli far removed from real life. By isolating very specific features, we are also able to isolate very specific functions in the brain. However, something important is lost in this abstraction. Ultimately, we are interested in mechanisms of thought as they exist in our everyday lives. In this paper, I explore ways to analyze fMRI data in complex, naturalistic stimuli—in this case an episode of a TV show—that will allow us to see how the brain represents features of this data. First, I examine how inter-subject similarity changes over time and in response to movie features. I explore how different similarity metrics demonstrate different structural features in the shared representation among viewers. A Multi-Voxel Classification Analysis is used to extend this result of common representational structure in similar stimuli by evaluating the relative dissimilarity of opposing feature representations. By showing reliable classification of scenes with only the activation patterns of other participants, this establishes that the shared neural patterns do correspond to shared representation. Finally, I turn to representational similarity analysis (RSA) in order to quantify the specific contributions of the individual features. Also discussed are complexities in measuring significance with permutation tests which, when ignored, can lead to massively overestimating significance in these analyses.