28.11 Pareidolia⧉
Pareidolia is the human tendency to see faces and patterns in things: the two screws and a slot of a wall outlet, the grille-and-headlights "face" of a car, a startled expression in a cloud or a piece of toast. It is an everyday quirk of perception, and Seeing Faces in Things (Hamilton, Stent, DuTell, Harrington, Corbett, Rosenholtz & Freeman, 2024, arXiv:2409.16143) Hamilton et al. 2024 turns it into a computational object. The authors assemble a dataset of roughly five thousand human-annotated "faces in things" (everyday objects whose arrangement triggers a face percept) and use it to compare human and machine face detection on images that contain, by construction, no actual face (Figure 28.11.1). The collection here is not a medium for art but an instrument for measurement: gathering thousands of these accidental faces lets you probe the face detector, both kinds, directly.
Machine face detectors partly fire on these illusory faces, which already says something (the same learned features that find real faces also catch the things that merely look like them), but they diverge from humans, and not randomly. The authors argue part of the divergence may be explained by an evolutionary pressure to detect animal faces, not just human ones: a perceptual system shaped to never miss a predator or prey will be tuned to over-detect faces, accepting false alarms (a face in a rock) to avoid the costly miss. Human pareidolia, on this reading, is the visible residue of a detector deliberately biased toward false positives. Training a model on human faces alone gives it a different operating point, and the dataset permits comparing these behaviors. The evolutionary account is a hypothesis supported by the experiments, not a direct measurement of evolutionary history.
This is the cleanest case of the previous chapter's thesis pointed inward, at perception itself. The Salavon average revealed a genre's convention; the anticliché camera revealed a place's distribution; here a photo collection of illusory faces reveals a property of the human visual system (the shape and bias of its face detector) that no single image of a startled-looking electrical outlet could establish on its own. It ties this part squarely back to Human vision: the collection does not just hold pictures of the world, it holds, in aggregate, a portrait of the eye that looks at it. And it has a practical edge: every face detector deployed in a camera, a photo-organizer, or a content filter inherits some version of this over-firing, so the "faces" it finds in upholstery, foliage, and rock formations are false positives a real system has to reckon with.
A collection of accidental faces measures the viewer, not the world. None of the images contains a face, yet both human and machine detectors fire, so the dataset is a portrait of the face prior, and the human–machine gap is a measurement of how that prior is tuned. The evolutionary reading (over-detect to never miss an animal face) explains why people see more faces than a human-face-trained model does. It is the part's recurring creed, the data is the prior, turned around: here the data is used to read out a prior that already lives in the visual system, and to remind every pipeline that runs a face detector that some of what it finds was never a face at all.
Recap: big lessons of this chapter
A collection of accidental faces measures the viewer, not the world. None of the images contains a face, yet both human and machine detectors fire, so the dataset is a portrait of the face prior, and the human–machine gap is a measurement of how that prior is tuned. The evolutionary reading (over-detect to never miss an animal face) explains why people see more faces than a human-face-trained model does. It is the part's recurring creed, the data is the prior, turned around: here the data is used to read out a prior that already lives in the visual system, and to remind every pipeline that runs a face detector that some of what it finds was never a face at all.