Whole-brain MEG decoding of symbolic and non-symbolic number stimuli reveals primarily format-dependent representations

Abstract

The human brain can rapidly form representations of numerical magnitude, whether presented with symbolic stimuli like digits and words or non-symbolic stimuli like dot displays. Little is known about the relative time course of these symbolic and non-symbolic number representations. We investigated the emergence of number representations for three stimulus formats - digits, words, and dot arrays - by applying multivariate pattern analysis to MEG recordings from 22 participants. We first conducted within-format classification to identify the time course by which individual numbers can be decoded from the MEG signal. Peak classification accuracy for individual numbers in all three formats occurred around 110 ms after stimulus onset. Next, we used between-format classification to determine the time course of shared number representations between stimulus formats. Classification accuracy between formats was much weaker than within format classification, but it was also significant at early time points, around 100 ms for both digit / dot and digit / word comparisons. We then used representational similarity analysis to determine if we could explain variance in the MEG representational geometry using two models: a GIST feature model capturing low-level visual properties and an approximate number model capturing the numerical magnitude of the stimuli. Model RSA results differed between stimulus formats: while the GIST model explained unique variance from 100-300 ms for all number formats, the performance of the approximate number model differed between formats. Together, these results are consistent with the view that distinct, format-specific number representations, moreso than a single \"abstract\" number representation, form the basis of numerical comparison.