Faces as spectra: implications for adaptation and face coding

Abstract

Studies of face coding have used adaptation to examine whether a facial dimension (e.g. gender or age) is represented by many narrowly tuned mechanisms or a small number of broadly tuned mechanisms. However, these models typically treat variations in the face as unique points along the coding dimension. Many natural stimuli instead correspond to a distribution of values (e.g. light spectra for color or amplitude spectra for space) rather than a punctate level (e.g. a single wavelength or spatial frequency). We examined the implications of reconceiving the face as a distributed spectrum for interpreting neural codes from face aftereffects. As a proof of concept, facial variations were created as the sum of local distortions at either different spatial locations (along the height of the face) or at different spatial scales (corresponding to different spatial frequencies of distortions). For modest distortions these appeared as plausible variations in real faces. Adaptation to the distorted faces produced strong aftereffects in the appearance of the original face. For example, adapting to a face that was contracted around the mouth and expanded around the eyes caused the original face to appear biased in the opposite way. These aftereffects are expected given the similar patterns of adaptation observed for face stimuli defined by functionally equivalent global distortions. However, the implications of the aftereffects are very different, because the implied channel structure depends on the assumptions about the stimulus structure. In particular, adaptation to the broadband face spectrum is predicted to produce broad and global changes in the channel responses even when the channel bandwidths are narrow, and predicts a normalization pattern of aftereffects even when there are many channels spanning the dimension. Our analysis illustrates important limits on the inferences that can be made about the neural representation of faces from how that representation adapts.

Meeting abstract presented at VSS 2018