Gunay, E. M. | Search bioRxiv

bioRxiv2025

Alignment of human behavior, brain, and AI models in the high-level valence processing of complex social scenes

Humans can evaluate the emotional meaning of complex social scenes in real-life settings. Recent evidence from the human brain and AI models pointed to visual processing as a core substrate for representing emotional valence of natural images, but this conclusion may not generalize to complex social scenes. We implemented experiments with social scenes in which emotional valence is partially dissociated from visual characteristics, objects, and scene settings. Human behavior, neuroimaging, and visual AI models confirm that visual processing captures basic emotional associations of objects and scene elements. However, when the valence of social scenes is incongruent with these basic properties, higher levels of processing are needed in the human association cortex and AI models. Our results show how and when valence processing demands advanced cognitive, neural, and computational processes that extend beyond the encoding of visual features.

neuroscience↗

bioRxiv2025

Common Neural Choice Signals reflect Accumulated Evidence, not Confidence

Centro-parietal EEG signals (CPP and Pe) correlate with the reported level of confidence. According to recent computational work these signals reflect evidence which feeds into the computation of confidence, not directly confidence. To test this prediction, we causally manipulated prior beliefs to selectively affect confidence, while leaving objective task performance unaffected. Behaviorally, we found that manipulating prior beliefs causally affected confidence without corresponding changes in accuracy and a negligible effect on reaction times. The EEG data showed a monotonic relation between the reported level of confidence and both CPP and Pe amplitudes. Importantly, this finding is compatible both with the theory that these signals track confidence as well as with the alternative theory that they track accumulated evidence. Critically, both neural signals were insensitive to the influence of prior beliefs on confidence, showing that they reflect the accumulated evidence that is used by the system to compute confidence, not directly confidence. Likewise, oscillatory activity in alpha and beta band was insensitive to the influence of prior beliefs on confidence. Decoding analyses revealed that the brain does hold shared representations for prior beliefs and confidence, and we identified a frontal signal that is sensitive to both confidence and prior beliefs.

neuroscience↗