TL;DR
Macaque face cells respond to illusory faces in objects, but this response is driven by local "eye" features, not the overall face-like configuration. This contrasts with human perception, which relies heavily on the global configuration, suggesting a dissociation between neural processing and conscious perception of illusory faces.
What Problem Does This Paper Address?
Vision neuroscience aims to understand how neural activity translates to perception. A key question is whether neurons in face-selective brain regions (like the inferotemporal cortex, IT) directly encode the high-level percepts they are associated with, such as faces. While humans perceive illusory faces in everyday objects (face pareidolia) based on a global, face-like configuration, it's unclear if macaque face cells respond similarly. This study investigates whether macaque face cells' response to illusory faces is driven by the same configural cues that drive human perception, or if it relies on simpler, local features.
Methods
The researchers used a combination of electrophysiology in macaque monkeys and behavioral experiments with human participants.
- Macaque Recordings: Multiunit activity was recorded from the inferotemporal cortex (IT) of macaque monkeys while they viewed images of objects that evoked illusory faces (pareidolia images) and matched control objects without illusory faces.
- Stimulus Manipulations: To dissect the role of configuration and parts, they used:
- Quadrant Scrambling: Images were divided into four quadrants and shuffled to disrupt the global configuration while preserving local features.
- Isolated Quadrants: Individual quadrants (categorized as "eye" or "non-eye" regions of the illusory face) were presented in isolation.
- Encoding Model: A non-pareidolia encoding model, trained on object categorization using only non-face, non-pareidolia images, was used to predict neural responses and pareidolia selectivity.
- Human Behavioral Tasks: Human participants rated the "faceness" of original, scrambled, and isolated quadrant images.
Datasets
- Pareidolia Images: 100 pareidolia images and 100 matched control objects were used in the main experiment. These were previously collected and validated for their ability to evoke illusory faces.
- Human Faceness Ratings: A dataset of 200 pareidolia images and matched controls, with human faceness ratings on an 11-point scale, was used to compare neural responses with human perception.
- Scrambled and Isolated Quadrant Images: Subsets of images were manipulated (scrambled, isolated quadrants) for specific experimental conditions.
- Non-face Object Images: Used for training the non-pareidolia encoding model.
Key Insights and Ideas
- Face Cells are Pareidolia Selective: Macaque face cells in IT cortex do respond more strongly to illusory faces than to control objects, confirming previous findings.
- Human Perception vs. Neural Response Dissociation: The key finding is that the neural basis for pareidolia selectivity in macaque face cells differs significantly from human perceptual experience.
- Local Features Dominate Neural Response: Macaque face cells' pareidolia selectivity is primarily driven by specific object parts, particularly those resembling "eyes," even when these parts are presented in isolation or out of their original configuration.
- Configuration is Secondary for Neurons: Disrupting the global, face-like configuration (via scrambling) significantly reduced human perception of faces but had a less pronounced effect on neural pareidolia selectivity.
- Generic Feature Encoding: The neural response to illusory faces can be predicted by an encoding model trained on general object features, suggesting that face cells might be tuned to generic visual properties that happen to align with illusory face configurations.
Key Results
- Neural Pareidolia Selectivity: Face-selective units in macaque IT cortex showed significant pareidolia selectivity (responding more to pareidolia images than matched controls).
- Faceness Ratings Don't Explain Neural Activity: Human ratings of how "face-like" an image appeared did not correlate with the neural responses of face cells to those images.
- Scrambling's Impact: Quadrant scrambling significantly reduced human faceness ratings but only moderately reduced neural pareidolia selectivity.
- "Eye" Quadrants Drive Selectivity: In both central and anterior IT, the "eye" quadrants of illusory faces accounted for the majority (75-80%) of the pareidolia selectivity observed in face cells.
- Encoding Model Prediction: A non-pareidolia encoding model accurately predicted neural pareidolia selectivity, and this predicted selectivity was also driven by "eye" parts.
Limitations and Future Work
- Interspecies Differences: The study acknowledges that the findings might reflect differences between macaque and human visual processing, as it's unclear if monkeys experience the illusory face percept in the same way humans do.
- Stimulus Selection: The reliance on human-rated stimuli for selecting pareidolia images is a standard challenge in cross-species research.
- Future Directions: The authors suggest further research to disentangle the contribution of face-like configurations from other non-face-like characteristics that might influence neural responses. They also propose investigating whether human face-selective regions show similar sensitivity to non-holistic attributes of pareidolia images.
Summary updated