M2RU: Memristive Minion Recurrent Unit for Continual Learning at the Edge

2025-12-19

AI Summary

Edge continual learning is hindered by the high energy consumption of conventional recurrent networks and frequent off-chip data transfers. This paper proposes M2RU, a memristor-based mixed-signal architecture tailored for edge deployment, which— for the first time—integrates weighted bit-stream computing with on-chip, hardware-level experience replay to enable low-power, real-time online temporal modeling and domain adaptation. Through crossbar optimization, device reliability modeling, and precision-enhancing encoding, M2RU effectively mitigates memristor non-idealities and significantly alleviates catastrophic forgetting. Experimental results demonstrate an energy efficiency of 312 GOPS/W—29× higher than state-of-the-art CMOS implementations—and maintain over 95% of software baseline accuracy on sequential MNIST and CIFAR-10 tasks. Projected device lifetime reaches 12.2 years.

Technology Category

• Machine Learning: Learning on the Edge & Model Compression
• Intelligent Robots: Multimodal Perception & Sensor Fusion
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Systems and Infrastructure for Web, Mobile and WoT: Energy management for devices in mobile Web and WoT environments

Abstract

Continual learning on edge platforms remains challenging because recurrent networks depend on energy-intensive training procedures and frequent data movement that are impractical for embedded deployments. This work introduces M2RU, a mixed-signal architecture that implements the minion recurrent unit for efficient temporal processing with on-chip continual learning. The architecture integrates weighted-bit streaming, which enables multi-bit digital inputs to be processed in crossbars without high-resolution conversion, and an experience replay mechanism that stabilizes learning under domain shifts. M2RU achieves 15 GOPS at 48.62 mW, corresponding to 312 GOPS per watt, and maintains accuracy within 5 percent of software baselines on sequential MNIST and CIFAR-10 tasks. Compared with a CMOS digital design, the accelerator provides 29X improvement in energy efficiency. Device-aware analysis shows an expected operational lifetime of 12.2 years under continual learning workloads. These results establish M2RU as a scalable and energy-efficient platform for real-time adaptation in edge-level temporal intelligence.

Problem

• Enables efficient continual learning on edge platforms
• Reduces energy-intensive training and data movement
• Maintains accuracy with high energy efficiency

Innovation

• Mixed-signal architecture for on-chip continual learning
• Weighted-bit streaming for multi-bit input processing
• Experience replay mechanism for stable domain adaptation