Fernando Ramirez¹ (), Harish Katti¹, Peter Bandettini¹; ¹NIMH

Bilateral symmetry is a fundamental property of organisms that actively navigate the world they inhabit. This property has been implicated in the estimation of face and body orientation, as well as exploited by prominent models of face identification to achieve representations invariant to head rotations. In line with these models, single-cell recordings from the macaque anterior lateral (AL) face-patch have shown a high-concentration of neurons exhibiting reflection invariance, or, in other words, neurons that exhibit similar responses to mirror-symmetric views of a face, like its left- and right-profile views, albeit distinct responses to different face-identities. However, mixed evidence of mirror-symmetry has been reported by neuroimaging studies of human high-level visual areas, and the precise source of findings of mirror-symmetry in monkeys and humans alike remains unclear. A key interpretational challenge regards the confounding role of low-level image properties associated with head rotations [Revsine et al., 2024 J Neurosci]. Nonetheless, a recent study probing convolutional neural networks [Farzmahdi et al., 2024 eLife] concluded that reflection invariant face-representations emerge in fully-connected network layers by pooling of reflection equivariant responses from earlier processing stages, and these properties could not be attributed to low-level image confounds. Here, we show that confounds consistent with those previously described in human neuroimaging studies are present in the images used in this study. More important, we show current methods to control for low-level confounds by equalizing the spectra of images presented to humans and/or artificial networks are limited; cortical magnification of the fovea in primates and a centrality bias of diagnostic features in pre-trained neural networks render spectral equalization ineffective as a control for low-level image confounds. As a way forward, we propose to embrace the variability—instead of controlling it—and suggest explicit experimental manipulations and analyses to identify genuine mirror-symmetric responses within a cross-validated model-comparison framework.