2024-07-03 09:30:56
Scientists have built an artificial intelligence system using memristors, electrically programmable electronic components that can store information. Through tests, they demonstrate that it can work with a miniature solar cell in dim light.
Artificial intelligence (AI) is now widely deployed in various embedded applications such as patient monitoring, building security and industrial safety. However, their high energy consumption is an obstacle to their deployment in environments requiring their autonomy. A team of researchers1, including those from IM2NP (Institut des Matériaux, de Microélectronique et des Nanosciences de Provence), has managed to overcome this difficulty thanks to memristors, these programmable electronic components capable of storing information by modulating the value of their resistance. Their work, published in the revue Nature Communicationspave the way for the deployment of AI in energy-autonomous embedded systems.
Memristors were used here because they have the advantage of consuming very little energy, which offers the possibility of self-powered operation when combined with energy harvesters, such as miniature solar cells. Except that most AI circuits based on this type of electronic component rely on analog memory calculation and exploit the classical laws of electricity (Ohm’s and Kirchhoff’s laws) to perform the fundamental operation of neural networks, namely multiplication and accumulation (also called MAC for Multiply Accumulate). This concept then requires a stable and precise power supply, which makes it incompatible with energy harvesters which are inherently unstable and unreliable.
To overcome this constraint, the scientists designed a binarized neural network, fabricated using a hybrid process combining CMOS (Complementary Metal Oxide Semiconductor) technology and memristors. For this, an integrated circuit was designed consisting of four arrays of 8,192 memristors each, for a total of 32,768 memristors, and which uses an amplifier strategy, combining two transistors with two memristors, for optimal robustness. This alternative approach is particularly resistant to fluctuations in unreliable power supplies, it also eliminates the need for compensation or calibration and allows efficient operation under a variety of conditions. This circuit was powered by a miniature high-bandwidth solar cell, which is optimized for cutting-edge applications.
Calculation accuracy even when energy becomes scarce
Simulations of image classification neural networks were performed to test this new system. The different tests reveal that under intense lighting, the circuit achieves inference performances comparable to those of a conventional laboratory power supply. While in low-light scenarios, it appears that the circuit remains functional, suffering only a modest drop in the accuracy of the neural network. More precisely, the misclassified images are mainly cases that are difficult to classify and that they are generally atypical or extreme cases.
The researchers demonstrate that for simpler tasks such as mining Modified National Institute of Standards and Technology (MNIST) databases, the circuit maintains its accuracy even when power is scarce. And when the power supply is lost, memristors retain data, unlike static random access memory (RAM), which loses stored information.
This IC can operate with low power supplies of 0.7 volts, which allows it to be powered by a broadband solar cell optimized for indoor applications and with an area of only a few square millimeters. Even under low light equivalent to 0.08 times the average solar flux, it remains functional. “Our approach lays the foundation for self-powered AI and the creation of smart sensors for various applications in health, safety and environmental monitoring,” write the authors of this research work.
1 This research work was carried out by researchers from the Institute of Materials, Microelectronics and Nanosciences of Provence (IM2NP, CNRS / Aix-Marseille University), in collaboration with scientists from the Center for Nanosciences and Nanotechnologies (C2N, CNRS / Paris-Saclay University), the Atomic Energy and Alternative Energies Commission (CEA-Leti) and the Photovoltaic Institute of Ile-de-France (IPVF)
1722364028
#artificial #intelligence #system #powered #solar #energy
