Fig. 7: Pareidolia selectivity is predicted by the feature-tuning estimated from non-pareidolia, non-face images.
From: Face cells encode object parts more than facial configuration of illusory faces
a Design of the non-pareidolia encoding model. This encoding model was based on a CNN trained on object classification. We input into this model the same images that were presented to the monkeys. Using cross-validated subsets of only the 100 matched controls and 40 non-face objects, we estimated a linear mapping from the model to each neural site (see “Methods”). This resulted in an encoding model for each neural site that captures the tuning for characteristics of only non-pareidolia, non-face images. b Scatterplot showing the similarity between each neural site’s observed (i.e., computed from neural responses) pareidolia d’ and its corresponding model-predicted value. Each dot represents a neural site in central IT (CIT; pink; n = 171) or anterior IT (AIT; green; n = 95). The values on the top left corner depict the Pearson’s correlation and the corresponding p-value calculated using the corr function in MATLAB (R2021b). c Scatterplot showing the correlation between neural face selectivity (observed face d’) and model-predicted pareidolia selectivity (pareidolia d’). The black line indicates an ordinary least squares (OLS) linear regression fit, with shaded 95% confidence intervals error bands. Same conventions as (b). d The proportion of mean d’ value (relative to the d’ of the full original images) for all quadrants, eye quadrants only and non-eye quadrants only, computed based on the responses of the non-pareidolia encoding models of face-selective units (observed face d’ > 1) in CIT (pink bar; n = 80), AIT (green bar; n = 58), and human subjects’ faceness ratings (gray bar; n = 100) for a subset of 34 images. Example images shown in (a). adapted from Wardle, S.G., Taubert, J., Teichmann, L. et al. Rapid and dynamic processing of face pareidolia in the human brain. Nat Commun 11, 4518 (2020). https://doi.org/10.1038/s41467-020-18325−8 released under a CC BY license: https://creativecommons.org/licenses/by/4.0/.
