Efficient human activity recognition with spatio-temporal spiking neural networks

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Human Activity Recognition (HAR) is crucial for health applications, typically using wearable sensor data.
• Current Artificial Neural Networks (ANNs) for HAR face challenges with high computational load and limited temporal feature extraction.
• Existing ANNs' activation functions hinder efficient processing on resource-constrained wearable devices.

Purpose of the Study:

• To investigate the efficacy of Spiking Neural Networks (SNNs) for Human Activity Recognition (HAR).
• To address the computational and temporal feature extraction limitations of ANNs in wearable HAR systems.
• To evaluate SNNs' potential for low-power, high-performance activity recognition.

Main Methods:

• Proposed the application of Spiking Neural Networks (SNNs), inspired by biological neurons, for HAR tasks.
• Utilized SNNs' ability to accumulate input activation and generate binary spikes for spatio-temporal feature extraction.
• Conducted experiments on three distinct HAR datasets to compare SNN performance against ANNs.

Main Results:

• SNNs demonstrated competitive or superior performance compared to traditional ANNs in HAR tasks.
• Achieved significant energy consumption reduction, up to 94%, by leveraging binary spikes for computation.
• SNNs effectively addressed limitations in temporal feature extraction present in conventional ANN approaches.

Conclusions:

• Spiking Neural Networks (SNNs) present a viable and efficient alternative for Human Activity Recognition (HAR) on wearable devices.
• The proposed SNN approach offers substantial energy savings without compromising recognition accuracy.
• SNNs provide a promising direction for developing next-generation, low-power wearable health monitoring systems.