Rapid and dynamic processing of face pareidolia in the human brain

The study investigates how and when the human brain misclassifies inanimate objects as faces (face pareidolia). The central question is whether illusory face perception arises from a rapid, broadly tuned face-detection mechanism driven by visual features, or from a slower cognitive reinterpretation of object parts as facial features. The work is motivated by extensive evidence for specialized face-processing systems in primate visual cortex and behavioral evidence of pareidolia in humans and macaques. By combining fMRI (spatial precision) and MEG (temporal precision) with a yoked stimulus design (illusory face images paired with visually/semantically matched nonface objects) and model-based analyses, the study aims to characterize the cortical loci and temporal dynamics of illusory face responses and to determine the contributions of low/mid-level visual features versus higher-level perceptual interpretations.

Prior research has identified face-selective regions in occipitotemporal cortex, notably FFA and OFA, which show abstract sensitivity to faces and face-like configurations. There is debate about whether ventral temporal representations are governed by appearance-based visual properties or higher-level semantic categories (e.g., animacy, real-world size). Evidence suggests rapid face detection in humans and possible subcortical pathways (superior colliculus, pulvinar, amygdala). Work on lookalike objects shows that appearance can override category identity in VTC. These findings frame the open questions: whether illusory faces engage face-selective cortex specifically versus broader category-selective cortex (e.g., LO, PPA), and whether pareidolia reflects fast, feature-based detection or slower reinterpretation.

Participants: MEG N=22 (8 male, 14 female; mean age 26.2), fMRI N=21 recruited (16 analyzed; 11 male, 10 female; mean age 25.4), Behavioral ratings N=20 (online). Ethics approvals obtained; informed consent provided. Stimuli: 96 images: 32 illusory faces in inanimate objects, 32 matched nonface objects (same identity, visually similar), 32 human faces. Images were 400×400 px, square-cropped. Human faces varied in age, gender, race, expression, and head orientation; counts of multi-face images matched to illusory set. Behavior: MTurk workers (n=20) rated “how easily you can see a face” on 0–10 scale. Experimental design: fMRI event-related design (images 300 ms; ISI 3.7 s; 32 blank trials of 4 s randomly inserted per run; 7 runs). MEG design: images 200 ms; ISI 1–1.5 s, 6 runs with 4 repeats per stimulus per run (total 24 repeats/stimulus). In both modalities, participants performed an orthogonal tilt judgment (3° left/right) to maintain attention. fMRI acquisition: 3T Siemens Verio, 32-channel coil; structural T1 MPRAGE (1 mm isotropic). Functional EPI: TR 2.5 s, TE 32 ms, voxel 2.8 mm isotropic, 42 slices. Two independent functional localizer runs defined ROIs (FFA, OFA, LO, PPA) using faces, places, objects, scrambled objects. Preprocessing: slice-time and motion correction, coregistration; no smoothing for experimental runs; analyses in native space. MEG acquisition: 160-channel axial gradiometer KIT system; 1000 Hz sampling; online 0.03–200 Hz filter. Trials epoched −100 to 1000 ms; downsampled to 200 Hz; PCA retained components explaining 99% variance; minimal preprocessing. Analyses: - Multivariate pattern analysis (MVPA): fMRI ROI-based decoding with linear SVM using The Decoding Toolbox; leave-one-run-out and leave-one-exemplar-out cross-decoding across categories (faces vs objects, faces vs illusory, illusory vs objects). Whole-brain searchlight decoding (Newton SVM; leave-exemplar-out) for spatial mapping. - MEG decoding: time-resolved LDA classifiers at each time point; leave-one-exemplar-out cross-decoding between categories; significance via Threshold-Free Cluster Enhancement (COSMOMVPA). - Representational similarity analysis (RSA): RDMs computed as 1−Spearman correlation over voxel patterns (fMRI) or sensor patterns (MEG) for each stimulus pair; multidimensional scaling for visualization. - Model-based RSA: constructed RDMs from behavioral face-likeness ratings; GBVS visual saliency maps; GIST visual feature descriptors. Compared model RDMs to time-varying MEG RDMs (Kendall’s tau-a), excluding human faces to avoid dominance by strong face signals. Tested category-level distinctions in GBVS/GIST via permutation tests with FDR correction. - fMRI–MEG fusion: Correlated ROI RDMs (excluding human faces) with time-resolved MEG RDMs to localize temporal correspondence (significance assessed across time).

- Behavioral validation: Human faces rated most face-like (M=9.96, SD=0.10) > illusory faces (M=6.27, SD=0.79; t(31)=26.15, p<0.001) > matched objects (M=0.70, SD=0.36; human vs matched t(31)=143.57, p<0.001; illusory vs matched t(31)=39.68, p<0.001), Bonferroni-corrected. - fMRI ROI cross-decoding: Human faces vs objects (with or without illusory face) decoded above chance in all ROIs (FFA, OFA, LO, PPA; p≤0.0008). Critically, illusory faces vs matched objects decoded only in face-selective cortex: FFA t(15)=2.54, p=0.015, d=0.64; OFA t(15)=2.34, p=0.020, d=0.59; not in LO t(15)=1.35, p=0.11; PPA t(15)=0.27, p=0.40. Searchlight confirmed peak decoding in ventral temporal cortex overlapping FFA/OFA. - MEG time-resolved decoding: All category pairs decodable within 200 ms. Faces vs objects showed peaks at ~160 ms and ~260 ms. Illusory faces vs objects also peaked at ~160 ms, consistent with rapid processing; the later ~260 ms peak present for real faces was absent or greatly reduced for illusory faces, indicating a transient face-like response for pareidolia. - MEG RSA dynamics: At 130 ms, human faces already clustered; some illusory exemplars near faces. By 160 ms, illusory faces segregated from matched objects and were more similar to human faces; by 260 ms, organization shifted, grouping illusory faces with matched objects, indicating rapid resolution of the initial detection error within ~100 ms. - Model comparisons: Correlation of model RDMs with MEG (excluding human faces) showed GBVS saliency correlated 85–125 ms (peak ~110 ms); GIST visual features correlated from ~95–400 ms; behavioral face-likeness ratings correlated later and stronger overall (120–275 ms and 340–545 ms; peak ~170 ms), approaching the noise ceiling. Category-averaged RDMs: GIST captured that illusory faces were more similar to human faces than matched objects (mean difference −0.0696, p=0.012), whereas GBVS did not; neither model captured stronger within-illusory-face similarity than to matched objects. - fMRI–MEG fusion: Only FFA RDM correlated significantly with MEG at ~160–175 ms (with human faces excluded), linking the early illusory face effect to face-selective cortex. Overall: Illusory faces elicit an early, face-like neural response localized to face-selective cortex and peaking ~160 ms, which rapidly reorganizes to an object-like representation by ~260 ms.

Findings indicate that errors of face detection in objects arise from a rapid, broadly tuned mechanism sensitive to visual features, rather than from a slower cognitive reinterpretation. Spatially, modulation by illusory faces is restricted to face-selective cortex (FFA/OFA), not to object- or scene-selective regions (LO/PPA), underscoring functional specificity within ventral visual cortex when appearance conflicts with category identity. Temporally, MEG reveals a transient face-like representation for illusory faces matching the early timing of real face detection (~160 ms), followed by a rapid reorganization toward object-based representation by ~260 ms, consistent with subsequent recognition-related processes resolving the initial detection error. Model-based RSA suggests low/mid-level visual features (captured by GBVS and GIST) contribute to the earliest stages (85–125 ms and ~95–400 ms), but behaviorally measured face-likeness better explains the overall neural representation starting ~120–170 ms. fMRI–MEG fusion implicates FFA as a likely cortical source of the early illusory face signal. These data inform the debate on appearance vs semantic organization in ventral temporal cortex, showing that when appearance implies a face, face-selective regions transiently represent illusory faces as more face-like despite object identity, before the system resolves toward object representation. The results highlight dynamic temporal weighting between detection and recognition processes in face perception and suggest potential involvement of a rapid (possibly subcortical) detection route that prioritizes sensitivity, followed by cortical mechanisms enhancing selectivity.

The study demonstrates that illusory faces in inanimate objects evoke a rapid, face-like neural response confined to face-selective occipitotemporal cortex, peaking around 160 ms, which is quickly resolved into an object-like representation by ~260 ms. This supports a broadly tuned, high-sensitivity face-detection mechanism that can produce compelling false alarms, followed by recognition-related processes that restore selectivity. Model and fusion analyses indicate contributions from low/mid-level visual features early in time and implicate FFA in the early illusory face response. Future work should: (1) directly probe subcortical contributions (e.g., superior colliculus–pulvinar–amygdala) to rapid face detection, (2) develop richer computational models that better capture the perceptual features driving pareidolia, (3) dissociate detection vs recognition pathways and their interactions using perturbation or lesion approaches, and (4) investigate individual differences and task/context effects on the persistence of pareidolic percepts.

- Eye movements were not recorded, though brief presentations likely minimized saccades. - Human faces were excluded from some RSA/model and fusion analyses to avoid dominance by strong face signals, potentially limiting generality of those specific comparisons. - GBVS and GIST models only partially explained neural/behavioral data, indicating incomplete modeling of the critical visual features. - fMRI temporal resolution precludes direct observation of rapid representational changes; MEG source localization is indirect. - Sample sizes were modest (fMRI N=16 analyzed; MEG N=22). - The persistent subjective experience of pareidolia beyond a few hundred milliseconds was not measured behaviorally during imaging, limiting links between neural dynamics and sustained perception.