Related Experiment Video
Updated: Jun 23, 2026
06:37
Artificial Intelligence-Based System for Detecting Attention Levels in Students
Published on: December 15, 2023
UDC-SNN: An Uncertainty-Aware Dynamic Cascading Framework with Spiking Neural Network for Balancing Performance and
Guihao Ran1, Shengzhe Li2, Zhiwen Jiang1
1School of Electronic Science and Engineering, University of Electronic Science and Technology of China, Chengdu 611731, China.
Sensors (Basel, Switzerland)
|May 13, 2026
Summary
This study introduces an uncertainty-aware dynamic cascading framework (UDC-SNN) for emotion recognition, balancing high accuracy with low energy use. The UDC-SNN effectively manages multimodal data for real-time applications.
Area of Science:
- Artificial Intelligence
- Biomedical Engineering
- Affective Computing
Background:
- Multimodal emotion recognition faces challenges in balancing performance and energy efficiency.
- Spiking Neural Networks (SNNs) offer potential for energy-efficient AI, but integrating multimodal data requires novel frameworks.
- Existing methods struggle with negative transfer effects from noisy unimodal data.
Purpose of the Study:
- To propose an uncertainty-aware dynamic cascading framework based on Spiking Neural Networks (UDC-SNN) for multimodal emotion recognition.
- To address the trade-off between recognition accuracy and energy consumption in emotion recognition systems.
- To enable real-time emotion recognition in resource-constrained environments.
Main Methods:
- Developed an asymmetric dynamic routing mechanism for demand-driven activation of electroencephalogram (EEG) and electrocardiogram (ECG) branches.
- Implemented uncertainty quantification using Shannon entropy for preliminary ECG inference.
- Created a parameter-free log-linear aggregation strategy to convert modality-specific entropy into dynamic Bayesian weights via an exponential decay function, mitigating unimodal noise effects.
Main Results:
- The UDC-SNN achieved an average recognition accuracy of 90.75% across valence, arousal, and dominance dimensions on the DREAMER dataset.
- The framework demonstrated remarkably low energy consumption, averaging 4.62 μJ per segment.
- Effectively mitigated negative transfer effects from unimodal noise through the proposed aggregation strategy.
Conclusions:
- The UDC-SNN framework successfully achieves a favorable balance between high emotion recognition accuracy and low energy consumption.
- The proposed uncertainty quantification and dynamic weighting mechanisms are effective in handling multimodal data and noise.
- The framework shows significant potential for real-time emotion recognition in resource-constrained applications like wearable devices.
