Face Pareidolia in the Rhesus Monkey

Abstract

Face perception in humans and nonhuman primates is rapid and accurate [1-4]. In the human brain, a network of visual-processing regions is specialized for faces [5-7]. Although face processing is a priority of the primate visual system, face detection is not infallible. Face pareidolia is the compelling illusion of perceiving facial features on inanimate objects, such as the illusory face on the surface of the moon. Although face pareidolia is commonly experienced by humans, its presence in other species is unknown. Here we provide evidence for face pareidolia in a species known to possess a complex face-processing system [8-10]: the rhesus monkey (Macaca mulatta). In a visual preference task [11, 12], monkeys looked longer at photographs of objects that elicited face pareidolia in human observers than at photographs of similar objects that did not elicit illusory faces. Examination of eye movements revealed that monkeys fixated the illusory internal facial features in a pattern consistent with how they view photographs of faces [13]. Although the specialized response to faces observed in humans [1, 3, 5-7, 14] is often argued to be continuous across primates [4, 15], it was previously unclear whether face pareidolia arose from a uniquely human capacity. For example, pareidolia could be a product of the human aptitude for perceptual abstraction or result from frequent exposure to cartoons and illustrations that anthropomorphize inanimate objects. Instead, our results indicate that the perception of illusory facial features on inanimate objects is driven by a broadly tuned face-detection mechanism that we share with other species.

Keywords: eye movements; face detection; face perception; monkey behavior; visual preference.

Figure 1. Experimental methods

(A). Examples of the three stimulus types used (from left to right: unfamiliar female monkeys, illusory faces, non-face objects). The non-face objects were selected from the public domain on the basis that they matched the examples of illusory faces for object content. (B). The results of the human experiment. Here, the rows represent individual subject data (N = 10) and columns represent the 45 images comprising the stimulus set. Importantly, none of the non-face objects was rated as being “face-like” (>100) on a 200-point scale (M_{non-face objects} =5.24; SEM =.45). Two pairwise contrasts confirmed that the non-face objects had a significantly smaller average score than either the monkey faces (P <.01, η² =.99) or illusory faces (P <.01, η² =.99). (C). The trial procedure for the three conditions of interest in the monkey experiment. Each trial consisted of three time periods: fixation, free viewing, and reward after successful trial completion or time out after trial aborts.

Figure 2. Experimental results

see also Figure. S1. (A). Bar graph indicates the average proportion of time spent looking at stimuli as a function of condition (error bars = +/− SEM). We found the expected advantage for monkey faces over objects, together with the hypothesized advantage for illusory faces over objects. We computed the average mean difference in each condition (monkey faces LT subtracted from illusory faces LT [I-M]; non-face objects LT subtracted from illusory faces LT [I-O]; non-face objects LT subtracted from monkey faces LT [M-O]) and performed a one-way repeated measures ANOVA (P <.01, η_p² =.86) to confirm that illusory face paired with monkey face trials elicited the smallest stimulus preference (paired t-tests, 2-tailed; [I-M] v [I-O], P <.01, η² =.97; [I-M] v [M-O], P =.016, η² =.80; [I-O] v [M-O], P =.300 η² =.24). (B). Bar graph demonstrating the distribution of first fixations in the three conditions of interest (the number of first fixations is expressed as a proportion of the total number of trials in each condition; error bars = +/− SEM). An analysis of the first fixation data indicated that in trials where monkey faces were presented with non-face objects, subjects fixated the monkey faces first, and more often (P =.01, η² =.92). There was a similar advantage for illusory faces over non-face objects (P =.01, η² =.96). As with the LT data, this analysis also revealed a significant preference for illusory faces over monkey faces (P =.01, η² =.92).

Figure 3. Fixations calibrated in degrees of visual angle and superimposed on stimuli

see also Figure. S2–S3. (A). Average number of fixations (≥150 ms) in 2-dimensional density plots (3 examples from each stimulus type; top row, monkey faces; middle row, illusory faces; bottom row, non-face objects). Data were normalized to each subject’s maximum fixation count, then averaged across subjects before being smoothed and superimposed on the corresponding stimulus for illustration using MATLAB’s surf function with interpolated shading. Unsmoothed data for every stimulus, together with individual subject maps (before averaging) are available in the supplementary material (Figure. S3). (B). The range of grand r-values as a function of stimulus type. After vectorizing the normalized fixation count data for each subject, we cross-correlated across subjects. This process yielded 10 r-values that were then averaged together to yield a single “grand r-value” per stimulus. The lower r values evident for the non-face objects reflect the greater variance among individual subjects. (C). Classifier performance as a function of subject; the classifier was trained with 93.33% of the data (i.e. 14 out of 15 illusory face/non-face pairs) and tested with the remaining content-matched pair. Chance performance is 50%.