Abstract
Mental imagery has fascinated philosophers, scientists and the public alike since antiquity. Neuroimaging now offers a window into this most personal of experiences. Multi-voxel pattern analysis (MVPA) of fMRI signal has successfully distinguished between a small number of imagined items/categories, and the neural activity pattern during visualization of an object has been shown to share similarities with that during perception. However, neuroimaging has not yet been used to characterise the content of complex mental images in any detail. Which stimulus features are being represented? Which aspects of the representation are shared between mental imagery and perception? Here we combine pairwise pattern-analysis of a rich stimulus set, with novel, adaptive stimulus selection, to provide precise, simultaneous characterisation of multiple facets of particular mental images, as represented by human occipitotemporal cortex. Across a large set of 200 naturalistic images, we identified those that when viewed evoked the most similar activation pattern to a predefined referent image. These stimuli we term the “neural neighbourhood,” and by examining which features they share, the neural representation of the referent was characterized along various feature dimensions. In two critical manipulations, we varied the identity of the referent image, and whether it was freely viewed or merely imagined. We find coding of precise semantic category and of emotional content, abstracted from low-level sensory properties. The balance of tuning along the various feature dimensions is dependent on the particular referent, and whether it was being viewed or imagined.