ULER: An Ultra-Lightweight Joint Attention Model for Emotion Recognition from Multimodal Physiological Signals

Area of Science:

• Affective Computing
• Biomedical Signal Processing
• Machine Learning

Background:

• Multimodal physiological signals offer objective emotion recognition.
• Current models are computationally intensive, limiting wearable applications.

Purpose of the Study:

• To propose ULER, an efficient framework for lightweight, multimodal emotion recognition.
• To enable practical, real-time emotion monitoring on resource-constrained devices.

Main Methods:

• Integration of lightweight multi-scale convolution (LMSC) and a joint attention mechanism (LJA).
• Utilizing a dynamic feature fusion network (DFFN) for effective multimodal fusion.
• Evaluation on DEAP, DREAMER, and WESAD benchmarks with subject-dependent and total dataset training.

Main Results:

• ULER achieves high accuracies (e.g., 99.34% on DEAP binary valence) with minimal parameters (0.60M) and low latency (31.10ms).
• Outperforms state-of-the-art methods on multiple benchmarks.
• Demonstrates superior performance even in a reduced channel (11 EEG) wearable setup.

Conclusions:

• ULER provides an efficient and effective solution for multimodal emotion recognition.
• The framework is suitable for real-time applications on wearable devices.
• Enables practical emotion monitoring in personal healthcare and other fields.