Humans possess remarkable face-processing abilities, with dedicated cortical regions. However, we also experience face pareidolia – perceiving faces in non-face objects. This raises the question of the neural mechanisms underlying this illusory perception. Is it a rapid process driven by basic visual features, or a slower, cognitive interpretation of ambiguous visual cues? This study aimed to investigate the spatial and temporal dynamics of illusory face processing in the human brain to address this question. Previous research in macaque monkeys suggests face pareidolia is a fundamental feature of primate face detection systems, not unique to humans. This study leveraged the high spatial resolution of fMRI and high temporal resolution of MEG to comprehensively analyze brain activity during the perception of illusory faces in objects. A yoked stimulus design was employed, creating matched pairs of images – one containing an illusory face and a visually similar counterpart without. This allowed direct comparison of brain responses in the presence or absence of the illusory face. The study's critical contribution lies in understanding the timing of the neural response to illusory faces and whether it is consistent with fast, automatic feature-based processing or slower, more deliberate cognitive processes.

Existing research has highlighted the specialized nature of human face processing, with regions like the fusiform face area (FFA) and occipital face area (OFA) showing selectivity for faces. However, the neural substrates and temporal dynamics of face pareidolia remain unclear. Behavioral studies, including those on macaque monkeys, indicate that the perception of illusory faces is likely rooted in the fundamental workings of the primate visual system. While some studies suggest abstract face representations in face-selective regions, it is unknown whether this sensitivity is exclusive to these regions or extends to other category-selective areas. The visual diversity of naturally occurring illusory faces offers a valuable opportunity to explore the behavioral tuning of face detection systems without relying on pre-defined features. Understanding the spatial and temporal aspects of processing is crucial to distinguish between a rapid, automatic, feature-based response and a slower, cognitive process that interprets the visual input.

This study utilized a multi-faceted approach combining fMRI and MEG neuroimaging with behavioral ratings and model-based analysis. Two experiments were conducted. **Experiment 1 (fMRI):** 21 participants viewed 96 images (32 human faces, 32 illusory faces, 32 matched non-face objects) presented for 300ms with a 3.7s interstimulus interval. Participants performed a secondary task of judging whether the images were slightly rotated left or right. fMRI data were acquired using a 3T Siemens Verio MRI scanner. Multivariate pattern analysis (MVPA) was applied using a linear support vector machine to classify brain activation patterns across voxels in regions of interest (ROIs) including the FFA, OFA, lateral occipital (LO), and parahippocampal place area (PPA). **Experiment 2 (MEG):** 22 participants underwent a similar experiment, with images presented for 200ms, a variable interstimulus interval (1-1.5s), and a secondary rotation judgment task. MEG data were recorded using a 160-channel KIT MEG system. Time-varying MVPA was employed using linear discriminant analysis (LDA) to assess the temporal dynamics of brain activity related to illusory face processing. Behavioral ratings of "faceness" were obtained from 20 Amazon Mechanical Turk workers for all images. These ratings were used to create a behavioral representational dissimilarity matrix (RDM). Two computational models, Graph-Based Visual Saliency (GBVS) and GIST, were also used to generate RDMs based on visual features. RDMs were constructed from fMRI and MEG data to assess representational similarity across conditions and time points. Data analysis involved cross-decoding (training and testing classifiers on different stimulus exemplars), representational similarity analysis (RSA), and correlation analysis between neuroimaging RDMs and behavioral/computational model RDMs. fMRI-MEG fusion analysis was performed to integrate findings from both modalities.

The study yielded several key findings: 1. **Spatial specificity:** The perception of illusory faces modulated brain activity primarily in face-selective cortex (FFA and OFA). Other category-selective regions (LO and PPA) did not show this sensitivity. 2. **Rapid temporal dynamics:** MEG data revealed a rapid and transient face-like response to illusory faces, peaking around 160ms post-stimulus onset. This early response was comparable to the response to real faces, suggesting a rapid, feature-based process. However, within ~250ms, this representation shifted, becoming indistinguishable from non-face objects. This indicates a rapid resolution of the face detection error. 3. **Model comparison:** The MEG response correlated early with visual feature models (GBVS and GIST), but stronger and more sustained correlations were found with the behavioral "faceness" ratings. This suggests that low-level visual features may trigger the initial face-like response, but the overall perception is influenced by more complex factors captured by the behavioral ratings. 4. **fMRI-MEG fusion:** Integration of fMRI and MEG data suggested that the early face-like response observed in MEG is primarily localized to the FFA. 5. **Differences between real and illusory faces:** While illusory faces elicited an initial face-like response, their brain representation differed substantially from real faces, exhibiting less within-category similarity in both fMRI and MEG analyses.

These findings support the hypothesis that face pareidolia arises from a rapid, feature-based face detection mechanism, rather than a slower cognitive reinterpretation. The initial face-like response, possibly originating from a subcortical pathway involving the amygdala, is quickly resolved through cortical processing in areas involved in face recognition. The study highlights a dynamic interplay between sensitivity and selectivity in face processing, with a broadly tuned face detector leading to initial errors that are rapidly corrected. The temporal resolution of MEG was critical in revealing this rapid transformation in neural representations. The results also shed light on the organization of object representations in ventral visual cortex, showing that responses to objects with dissociated visual appearance and semantic meaning are not uniform across brain regions. The involvement of low-level visual features in driving the illusory face response is suggested by the correlation analysis, but neither GBVS nor GIST models fully captured the complexity of the neural representation.

This study demonstrates the unique representation of illusory faces in the human brain. An initial rapid face-like response, possibly via a subcortical route, is quickly resolved as the brain’s representation shifts to an object-like pattern within a fraction of a second. This highlights the dynamic nature of visual perception and underscores the importance of temporal resolution in understanding cognitive processes. Future research could explore the specific visual features that trigger this rapid response and further investigate the role of subcortical structures in illusory face perception.

The study's limitations include the relatively small sample size, the use of a secondary task (rotation judgment) that might have influenced brain activity, and the lack of direct eye-tracking to confirm attentional focus. Additionally, the computational models employed, while widely used, may not fully capture the intricacies of human visual processing. The generalizability of these findings to naturalistic settings needs further investigation.