Decoding the categorization of visual motion with magnetoencephalography

Abstract

Brain decoding techniques are particularly efficient at deciphering weak and distributed neural patterns. Brain decoding has primarily been used in cognitive neurosciences to predict differences between pairs of stimuli (e.g. faces vs. houses), but how distinct brain/perceptual states can be decoded following the presentation of continuous sensory stimuli is unclear. Here, we developed a novel approach to decode brain activity recorded with magnetoencephalography while participants discriminated the coherence of two intermingled clouds of dots. Seven levels of visual motion coherence were tested and participants reported the colour of the most coherent cloud. The decoding approach was formulated as a ranked-classification problem, in which the model was evaluated by its capacity to predict the order of a pair of trials, each tested with two distinct visual motion coherence levels. Two brain states were decoded as a function of the degree of visual motion coherence. Importantly, perceptual motion coherence thresholds were found to match the decoder boundaries in a fully data-driven way. The algorithm revealed the earliest categorization in hMT+, followed by V1/V2, IPS, and vlPFC.