Hybrid Neuromorphic‐Federated Learning for Activity Recognition Using Multi‐modal Wearable Sensors
Ahsan Raza Khan, Habib Ullah Manzoor, Fahad Ayaz, Muhammad Ali Imran, Ahmed Zoha · 2024
In this chapter, the authors proposed an hybrid neuromorphic federated learning framework that synergizes the computational efficiency of spiking neural networks with the dynamic temporal learning capabilities of long short-term memory (LSTM) networks for human activity recognition (HAR) using multi-model data from wearable sensors. This integrated spiking-LSTM (S-LSTM) model capitalizes on LSTM layers to adeptly capture temporal dependencies within time-series sensor data while incorporating spiking layers to facilitate event-driven processing, thereby enhancing energy efficiency in federated settings. The approach has been tested on two publicly available HAR datasets – UCI and Real-World. The simulation results show concrete evidence that the S-LSTM model outperforms the LSTM model in accuracy. Moreover, our findings highlighted a significant 32.30% improvement in energy efficiency compared to the LSTM model.