Empowering Intelligence

Researchers have developed self-powered edge AI hardware, using solar cells and memristors, that can run neural networks even in low light.

Nick Bild
1 month ago β€’ Machine Learning & AI
A self-powered, energy-efficient computing platform for edge AI (πŸ“·: F. Jebali et al.)

Edge AI refers to the deployment of artificial intelligence (AI) algorithms directly on local devices, such as smartphones, cameras, sensors, and other Internet of Things devices, instead of relying on a centralized cloud server for processing. This approach brings computation and analysis closer to the source of data, offering several advantages in terms of speed, privacy, and efficiency.

One significant benefit of edge AI is its enhanced inference speed. By processing data locally on the device, there is a reduction in the latency associated with sending data to a remote server and waiting for a response. This is particularly crucial in applications where real-time decision-making is essential, such as autonomous vehicles, smart surveillance systems, and augmented reality.

Privacy is another key advantage of edge AI. Since data is processed locally, sensitive information can often remain on the device and not be transmitted over a network. This is especially pertinent in applications like healthcare, where patient data confidentiality is of utmost importance.

Despite these advantages, powering large, distributed networks of AI devices at the edge poses numerous challenges. One of the primary challenges revolves around the fact that the energy demands of AI computations can lead to rapid battery drainage in edge computing platforms. A team led by researchers at Aix-Marseille University have come up with a potential solution to this problem that could enable a whole new set of applications to utilize edge AI techniques. They have developed a self-powered, energy-efficient hardware platform that can run even complex image classification neural networks.

The researchers relied on a very energy-efficient emerging technology in their design β€” memristor-based computing. The most efficient of these systems utilize analog-based in-memory computing, which was also utilized in this work. These circuits can perform the multiply-and-accumulate operation, which is critical to neural network inferences, directly in memory, eliminating the need for data transfers between memory and processing units. This not only reduces power consumption, but is also faster.

In total, four arrays of 8,192 memristors were incorporated into the device. It is powered by a miniature wide-bandgap solar cell that can generate electricity even under low-light conditions, such as those encountered indoors. Traditionally, this arrangement would be problematic β€” memristor-based computing circuits depend on complex peripheral circuits that are tuned for a particular voltage to overcome issues with the inherent high variability of memristors. But energy harvesters, like solar cells, vary significantly in the voltage levels they generate in a way that depends on present environmental conditions.

The team’s way around this problem involved the use of a logic-in-sense-amplifier and two-transistor/two-memristor strategy. This made the device robust against changes in power supply voltage, and did not require any complex supporting circuits. It was demonstrated that this hardware configuration allowed the device to operate even when little energy was produced by the solar cell β€” it could operate at light levels as low as 0.08 suns. Under these conditions, the computing unit would transition from a high-precision mode to a more approximate mode to allow for continued computation.

To assess the performance of this system, a binarized neural network was implemented in the hardware. The network was tasked with recognizing handwritten digits, with training from the MNIST dataset, and also general image recognition tasks after being trained on the CIFAR-10 dataset. When assessing the more challenging general image recognition task, it was found that images were classified correctly in over 86% of cases with full power. Under low-light conditions (0.08 suns), this accuracy only dropped to 73%, showing the adaptability of the technology to difficult conditions. Further evaluation showed that the misclassifications that did occur were due to difficult-to-classify cases.

These initial results are very encouraging. Based on the performance that has been observed thus far, the researchers hope to see their system used in the development of intelligent sensors for health, safety, and environmental monitoring applications in the near future.

Nick Bild
R&D, creativity, and building the next big thing you never knew you wanted are my specialties.
Latest articles
Sponsored articles
Related articles
Latest articles
Read more
Related articles