Representation-First Emotion Decoding from 7T fMRI

We present a scalable signal processing framework for measuring naturalistic emotional responses in 7T fMRI using 3D convolutional neural networks. Our model learns low-dimensional representations of affective activity from whole-brain recordings during narrative-driven auditory stimulation, recovering structure consistent with valence–arousal dimensions in affective neuroscience. By prioritizing emotional representation learning over anatomical interpretability, the model maps neural activity across individuals into a shared latent space aligned with canonical affective geometry, enabling scalable cross-subject analysis without region-specific assumptions. We also observe subject-specific deviations in these representations that may capture individual differences in emotional processing, suggesting opportunities for downstream interpretation and personalized analysis. This work establishes a scalable deep learning framework for emotion-aware representation learning in fMRI.