CUHK
News Centre

A collaborative research team led by The Chinese University of Hong Kong (CUHK)’s Department of Electronic Engineering has successfully developed a lightweight, error-tolerant stochastic computing architecture by harnessing the inherent stochastic switching behaviour of memristors. The new technology significantly enhances the visual processing efficiency of edge devices[1], with potential applications in autonomous driving, virtual and augmented reality wearables, medical imaging equipment and more. It offers a novel hardware solution for the deployment of artificial intelligence in resource-constrained environments. The findings have been published in the renowned international journal Nature Communications.

Memristors, short for memory resistors, are emerging microelectronic devices that look set to enable next-generation AI computing, due to their ability to simultaneously store and process data with minimal energy cost. However, the inherent stochastic electrical switching behaviour, the unpredictable switching problem, of memristors typically hinders their use in high-precision computing. Prior research mainly focused on suppressing this characteristic to improve computational accuracy.

Reverse thinking: turning stochasticity into an advantage

The research team took a different approach by using the inherent switching stochasticity of memristors to propose a probability-based stochastic computing architecture, enabling more lightweight and error-tolerant image processing.

To validate this concept, the team performed edge detection using probabilistic logic constructed from memristors. Edge detection[2] extracts fundamental visual cues such as contours and textures from images, helping humans and machines understand images and make decisions. Extracting these visual cues from images typically involves intense matrix multiplication and gradient computation. Implementing this in binary computing can lead to excessive computational workload and latency, making it undesirable for hardware integration and deployment in resource-constrained edge visual scenarios. Moreover, the deterministic and high-precision data representation can be overly redundant but is also highly susceptible to soft errors induced by environmental noise and interference.

In contrast, the new stochastic computing architecture converts data into sequences of randomly distributed 0s and 1s, performing logical operations in a probabilistic manner. This data representation is inherently error-tolerant, as the impact of paired bit-flips, accidental errors in the data where 0s and 1s are inverted due to interference, can be cancelled. This approach more closely aligns with the characteristics of human vision and AI decision-making, providing valuable support for real-time analysis and computing on edge devices.

Reduction of 95% in energy consumption advances autonomous driving and medical imaging diagnosis

The team designed stochastic number encoders (SNEs) that leverage the properties of memristors and integrated them with compact logic gates to develop lightweight stochastic logic units. The SNEs encode data into stochastic numbers with well-regulated probabilities, enabling probability-based logical computations. The team also completed a hardware Roberts cross[3] operator to extract image contours and textures. The new technology achieved a remarkable 95% reduction in energy consumption while withstanding up to 50% of bit-flips, demonstrating its exceptional performance in lowering computational costs and offering error-tolerance.

Professor Hu Guohua, the corresponding author of the paper and Assistant Professor in the Department of Electronic Engineering at CUHK, said: “While these are promising advances, one major challenge lies in the integration of memristors with peripheral circuits and the parallel operation of large-scale circuits. Besides, the success rate and uniformity of the memristors are still key concerns in large-scale manufacturing, despite current technological advances. Additionally, device-to-device non-uniformity in memristor fabrication can significantly affect the overall reliability and performance. A system-level analysis of memristors, SNEs and peripheral circuits could therefore be considered.”

Dr Song Lekai, the paper’s first author, added that further development is needed, with hardware and algorithm co-designs also required to address or accommodate adverse circumstances, such as noises and delays from the memristors and electronic circuits. However, this technology marks an important step towards leveraging the inherent stochasticity of memristors for lightweight, error-tolerant edge visual applications ranging from autonomous driving, virtual reality (VR) and augmented reality (AR) to medical imaging diagnosis and beyond.

For the full research, please visit: https://www.nature.com/articles/s41467-025-59872-2

[1] An edge device is a terminal unit located at a network boundary. It acts as an interface between the real world and the network, enabling data collection or computational processing directly near its source. Examples include sensors, smartphones, routers and autonomous vehicles.

[2] Edge detection includes a variety of mathematical methods that aim at identifying edges, defined as curves in a digital image at which the brightness changes sharply or, more formally, has discontinuities. It is a fundamental tool in image processing, machine vision and computer vision.

[3] Roberts cross operators are used in image processing and computer vision for edge detection. The idea behind them is to approximate the gradient of an image through discrete differentiation, which is achieved by computing the sum of the squares of the differences between diagonally adjacent pixels.