Temporal confounds emulate multivariate fMRI measures of perceptual learning
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Human perception inherently involves learning. We can experience the same stimulus completely differently depending on our prior knowledge. Understanding the neural basis of perception therefore requires measurements that capture these temporal dynamics. Multi-voxel pattern analysis (MVPA) approaches are widely used to characterise changes in neural representations over time. A popular example is the perceptual reorganisation paradigm, which investigates the neural correlates of enhanced recognition of distorted images after cueing with the undistorted version. Studies typically report increased representational similarity of these two images in early visual cortex. However, as these paradigms include an inherent ordering of stimuli that precludes trial randomisation or counterbalancing, they are vulnerable to temporal confounds common to fMRI. Here, we investigate how these confounds could influence current understanding of perceptual reorganisation. We tested different perceptual reorganisation paradigm designs derived from published fMRI studies, and found substantial design-driven order effects at the single-subject level for all paradigms. For certain designs, these effects artificially amplified neural indices of perceptual reorganisation at both the single-subject and group levels, emulating widespread signatures of perceptual reorganisation across the brain. To disentangle perceptual learning processes from measurement artefacts, we recommend (i) selecting designs that minimise the effect of stimulus order confounds on contrasts of interest, (ii) correcting for these confounds, and (iii) confirming results are perceptually driven via negative controls, e.g., stimuli or brain areas not expected to produce perceptual effects. Our work demonstrates how current understanding of perceptual learning mechanisms based on multivariate neuroimaging approaches could be influenced by non-obvious design confounds that misdirect interpretations towards distributed neural processing, and offers practical solutions to address this.