Resolving discrepant conclusions regarding the representation of mirror-symmetric face views in human face-selective areas

Viewpoint invariant face recognition is a remarkable feat of the primate visual system. Traditional theories hold that viewpoint is coded by view-selective mechanisms at early visual processing stages and representations become gradually tolerant to viewpoint changes in higher-level visual areas. Newer theories, based on single-neuron monkey electrophysiological recordings, suggest a three-step architecture revealing a sharp transition from a view-tuned to a mirror-symmetric representation before achieving viewpoint invariance at the highest level of the hierarchy. Consistent with traditional theories, human studies combining neuroimaging and multivariate pattern analysis have provided evidence of view-selectivity in early visual cortex. However, contradictory results have been reported in higher-level visual areas regarding the existence in humans of a mirror-symmetric processing stage.
We recently proposed a unifying network model for these observations [Revsine, Gonzalez-Castillo, Merriam, Bandettini, and Ramírez, 2023. BioRxiv, DOI: 10.1101/2023.02.08.527219]. The model shows that low-level feature imbalances among images of lateral and frontal face views would lead to artefactual observations of mirror symmetry at levels of the visual hierarchy where neuronal receptive fields are large and span both visual hemifields. These artefacts emerge gradually along the visual hierarchy and manifest when the Euclidean distance is used as measure of dissimilarity among brain patterns, or the correlation distance is used instead but computed on mean-centered data. Mirror symmetry is not observed with the correlation distance if computed on uncentered data. Here, we provide empirical evidence in support of the predictions made by our model. We conducted pattern analyses of functional MRI data from early visual cortex, lateral occipital cortex, and the occipital and fusiform face areas [from study by Ramírez, Cichy, Allefeld and Haynes, 2014. The neural code for face orientation in the human fusiform face area. J Neurosci. 34:12155-67]. As predicted, we found no mirror symmetry when relying on the correlation distance on uncentered data. We also observed a gradual increase of mirror-symmetry as a function of the location of a brain area along the ventral stream when relying on the Euclidean distance or when the correlation distance was computed on mean-centered data. These observations suggest that reports of mirror-symmetry in humans are an artefact due to signal imbalances across conditions, and call attention to the influence of common analysis choices on inferences about neural coding based on pattern analyses of neuroimaging data.