RIS-Based Over-the-Air Neural Networks for Low-Power CNN Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing wireless devices face challenges in performing real-time machine learning inference tasks due to the high computational power and latency requirements of convolutional neural networks (CNNs), especially in resource-constrained IoT platforms, necessitating a solution that enables fast inference without specialized hardware.
Innovation Solution
Implementing an over-the-air neural network (OANN) using reconfigurable intelligent surfaces (RIS) to emulate digital convolution operations in the analog domain, allowing signal reflections to determine the output of convolution steps without requiring signal storage, data forwarding, or dedicated processors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional CNN architecture with digital convolution operations is used, then inference accuracy is maintained, but computational power requirements and latency increase significantly
Solution Approach 1:
The patent replaces the mechanical/digital computation system with an electromagnetic wave-based analog system. Specifically, it substitutes digital convolution operations performed by processors with electromagnetic wave propagation and reflection phenomena, where the physical environment itself performs the computation. The wireless channel and RIS elements act as analog computational components, eliminating the need for digital processors while maintaining inference accuracy.
Solution Approach 2:
The patent introduces reconfigurable intelligent surfaces (RIS) as intermediary elements between the transmitter and receiver. These RIS elements mediate the electromagnetic wave propagation by providing configurable reflections that encode computational operations. The RIS acts as a physical intermediary that transforms the wireless channel into an analog computational engine, enabling CNN inference without digital processors.
2Speed
If signal processing is performed at the sensor location with specialized hardware, then real-time inference is achieved, but device size and complexity increase
Solution Approach 1:
The patent enables the wireless environment itself to perform the computational service. Instead of requiring the sensor device to provide specialized processing hardware, the system uses the natural electromagnetic wave propagation characteristics and configurable RIS elements to automatically perform convolution operations. The physical environment provides the computational service that would otherwise require complex onboard hardware.
Solution Approach 2:
The patent makes the wireless communication channel multi-functional by enabling it to simultaneously perform signal transmission and analog convolution computation. The same wireless infrastructure used for data communication is also utilized for neural network inference, eliminating the need for separate specialized processing hardware and reducing device complexity.
3Power
If data is transmitted to remote MEC center for processing, then computational resources are available, but communication overhead and latency increase
Solution Approach 1:
The patent transitions the computation from the digital domain to the analog electromagnetic domain. By performing convolution operations in the analog signal domain through wave propagation and reflection, the system avoids the need to transmit large volumes of digital data for processing. The computation happens in a different dimensional space (analog vs. digital), enabling real-time inference without data forwarding latency.
4Speed
If onboard computing resources are included in IoT platforms, then real-time inference is enabled, but power consumption and device size increase
Solution Approach 1:
The patent replaces energy-consuming digital processing operations with passive electromagnetic wave propagation. The analog convolution computation is performed by the natural behavior of electromagnetic waves interacting with the wireless channel and RIS elements, requiring no active power consumption for the computational operations themselves. This substitution eliminates the need for power-hungry onboard processors in IoT devices.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Enables processor-free, low-power, and latency-free inference for tasks like modulation classification, reducing the need for onboard computing resources and maintaining accuracy comparable to all-digital CNNs.
Implementation Method 1
The OANN leverages the physics of signal reflection to represent digital 'convolution,' a part of a CNN architecture, in the analog domain
Data Source
AI summary
Provided herein are systems and methods for implementing an over-the-air neural network (OANN) including by receiving, at a relay receiver of a relay node, a signal of interest from a transmitter, directionally re-transmitting the signal of interest from each of a plurality of relay transmitters of the relay node to a corresponding one of a plurality of programmable reconfigurable intelligent surfaces (RIS), reflecting, by each of the plurality of RIS, the corresponding re-transmitted signal of interest, and adjusting, by a neural network controller, a reflection angle of each of the plurality of RIS to direct the reflected signals of interest to combine in a deterministic manner at the relay receiver.