Memristor-Based Circuit Demonstration of Hybrid Gated Recurrent Unit for Edge Computing

Abstract

In industrial IoT and distributed computing environments, edge computing devices empowered by AI have seen increasing deployment in large-scale scenarios, thereby accelerating the demand for time-series data processing. The gated recurrent unit (GRU) outperforms conventional artificial neural networks (ANNs) in tasks such as natural language processing, speech recognition, and machine translation, due to its superior capability in modeling long-range dependencies in sequential data. However, the GRU model is limited by its large parameter count and structural complexity, which presents a bottleneck in hardware circuit implementation. To this end, a memristor-based hybrid gated recurrent unit (HGRU) is proposed, which reduces the parameter count to 67% of the original GRU and shortens the single-step computation latency by 50%, while maintaining complete circuit functionality. Finally, the proposed memristor-based HGRU circuit model is evaluated on the MNIST digit recognition and IMDB sentiment analysis tasks, achieving recognition accuracies of 97% and 86.2%, respectively. Under equivalent parameter settings, it achieves runtime reductions of 37% and 52% compared to the standard GRU, thereby significantly enhancing computational efficiency.