Skip to main content
Back

EEG-MoCE Reads Human Emotions 14 Points More Accurately Using Curved Space Math

A new framework maps each brain-signal modality to its own curved geometry, delivering double-digit accuracy gains on emotion, sleep, and cognitive tasks.

  • EEG-MoCE embeds each modality in its own learnable-curvature hyperbolic space for multimodal brain-signal learning
  • Achieves 75.88% on EAV emotion recognition, a 14-point jump over prior best (paper)
  • Curvature magnitude serves as data-driven signal of modality importance during cross-modal attention fusion
  • State-of-the-art on three benchmarks: emotion (EAV), sleep staging (ISRUC), cognitive load (N-back)
  • Delta-hyperbolicity analysis confirms EEG, video, audio, EMG, EOG all show tree-like structure
  • Open-source code at github.com/zhourunhe/EEG-MoCE, 2.71ms inference latency

Reading mental states from brain signals is a messy business. EEG is noisy, subject-dependent, and often needs to be combined with video, audio, or other physiological signals to be useful. A new framework called EEG-MoCE tackles that fusion problem with a geometric twist: instead of embedding every modality in flat Euclidean space, it gives each one its own curved hyperbolic space, then uses the learned curvature as a signal for how much that modality should count during fusion.

On a five-class emotion recognition benchmark, the method jumps from around 62% to 75.88% balanced accuracy, a 14-point gain over the strongest prior baseline. Code is public.

Why flat space breaks down

Recent work suggests the brain's hierarchical cognitive processes can be encoded in EEG and adjacent modalities like facial expressions. Euclidean embeddings struggle to represent the exponential growth of states in hierarchical structures, because distances and areas scale only linearly or quadratically with radius in flat geometry. Hyperbolic space, with its negative curvature and exponential volume growth, fits tree-like data naturally, which is why it has caught on for language models and knowledge graphs.

Comparison of Euclidean and hyperbolic geometries for hierarchical data

The authors verify this empirically using a metric called