Yu Chen¹ (y.chen.8@research.gla.ac.uk), Yuening Yan¹, Oliver G. B. Garrod¹, Robin A.A. Ince¹, Rachael E. Jack¹, Philippe G. Schyns¹; ¹University of Glasgow

Facial expressions are multi-component signals composed of Action Units (AUs), from which humans infer both discrete emotion categories (e.g., anger) and affective dimensions (e.g., negative valence, high arousal). Yet the neural mechanisms through which the brain extracts these distinct signals remain poorly understood. Using a generative face model, we independently sampled eight emotion-relevant AUs to create 1,600 animations rendered on random 3D identities. Each AU was either absent (0 intensity, 50% probability) or present with a random non-zero intensity between 0 and 1 (maximum perceptual visibility), enabling us to trace AU-specific neural representations in 8,196 cortical MEG sources. Participants viewed all animations and completed three 7-AFC tasks: emotional category (“happy”, “fear”, “surprise”, “anger”, “disgust”, “sad” and “other”), valence (negative to positive), and arousal (low to high). Each participant saw the same 1,600 animations in each task (dense-sampled, 4,800-trial design), allowing us to isolate task-dependent effects on neural representations. We recorded whole-brain MEG and reconstructed source activity at 8,196 cortical locations. AU representations were quantified using mutual information between single-trial AU intensities and source-level MEG amplitudes every 4ms. After animation onset, AUs were first represented in occipital cortex (~210ms) and then propagated along the third pathway to posterior STS and STG (~310ms), replicated across all participants. Whereas occipital representations were largely task-invariant, pSTS/STG showed systematic task-dependent reweighting. For example, 3×2 ANOVAs (Task×Hemisphere) of one participant revealed stronger STG representations (e.g., Brow Lowerer and Jaw Drop) during category task (e.g., responding “surprise”), with distinct lateralization in pSTS—rightward for Brow Lowerer and leftward for Jaw Drop (relative to other tasks and hemispheres, p<0.05). Thus, while early visual representations of facial movements are task-invariant, higher-level pSTS/STG flexibly and laterally reweights identical facial movement inputs depending on whether observers judge categories or dimensions. This dynamic routing supports divergent emotional interpretations from the same facial inputs.

Acknowledgements: This work was funded by the Wellcome Trust (Senior Investigator Award, UK; 107802) and the Multidisciplinary University Research Initiative/Engineering and Physical Sciences Research Council (USA, UK; 172046-01) (P.G.S.); ERC [FACESYNTAX; 759796] (R.E.J.); the Wellcome Trust [214120/Z/18/Z] (R.A.AI.).