Multifold Fusion Attention Variant for Emotion Recognition
Developed an end-to-end framework that processes raw multimodal data without feature preprocessing, addressing challenges in emotion recognition for real-world applications including AI, healthcare, and human-computer interaction.
Key Contributions:
- Novel attention-based intermediate fusion mechanism
- Dynamic modality weighting for adaptive data quality handling
- End-to-end processing of raw audio, video, and EEG data
- Captures both short- and long-range dependencies
Results:
91.2%
Accuracy on EAV Dataset
7.39%
Improvement over TransNet