Over-the-air multi-sensor inference with neural networks using memristor-based analog computing

2024-12-01

Physical Communication

AI Summary

To address the high power consumption and massive data transmission bottlenecks hindering deep neural network deployment on resource-constrained sensor nodes in real-time wireless systems, this paper proposes an over-the-air direct-drive memristive neural network paradigm for analog co-inference. The method bypasses analog-to-digital conversion and digital storage, directly encoding multi-sensor data in the RF domain and driving it into a memristor crossbar array to perform vector-matrix multiplication and end-to-end inference entirely in the analog domain. This work pioneers the deep integration of sensing, communication, and computation within the analog domain, establishing a low-power near-sensor computing architecture. Experimental evaluation demonstrates 92.3% inference accuracy on multimodal sensor streams, an end-to-end energy efficiency of 12.8 TOPS/W, and a 5.7× reduction in inference latency.

Technology Category

• Machine Learning: Hardware-aware ML
• Application Domains: Internet of Things, Sensor Networks & Smart Cities
• Cognitive Modeling & Cognitive Systems: Analogy

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation

Problem

• Energy Efficiency
• Wireless Systems
• Deep Neural Networks

Innovation

• Memristor-based Analog Computing
• Over-the-air Inference
• Energy-efficient Processing