Semantic Latent Geometry Reveals Imagination–Perception Structure in EEG

We investigate whether representation-level, semantic diagnostics expose structure in electroencephalography (EEG) beyond conventional accuracy when contrasting perception and imagination and relating outcomes to self-reported imagery ability. Using a task-independent encoder that preserves scalp topology and temporal dependencies, we learn semantic features from multi-subject, multi-modal EEG (pictorial, orthographic, auditory) and evaluate subject-independent decoding with lightweight heads, achieving state-of-the-art or better accuracy with low variance across subjects. To probe the latent space directly, we introduce threshold-resolved correlation pruning and derive the Semantic Sensitivity Index (SSI) and cross-modal overlap (CMO). While correlations between Vividness of Visual Imagery Questionnaire (VVIQ)/Bucknell Auditory Imagery Scale (BAIS) and leave-one-subject-out (LOSO) accuracy are small and imprecise at n = 12, the semantic diagnostics reveal interpretable geometry: for several subjects, imagination retains a more compact, non-redundant latent subset than perception (positive SSI), and a substantial cross-modal core emerges (CMO ≈ 0.5–0.8). These effects suggest that accuracy alone under-reports cognitive organization in the learned space and that semantic compactness and redundancy patterns capture person-specific phase preferences. Given the small cohort and the subjectivity of questionnaires, the findings argue for semantic, representation-aware evaluation as a necessary complement to accuracy in EEG-based decoding and trait linkage.