Over the past few decades, the performance of machine learning models in various problems in the real world has improved significantly. However, the training of these models, as well as their implementation, still requires huge energy and computing power.
That is why engineers around the world try to develop alternative hardware solutions that will more effectively launch artificial intelligence models. This can contribute to their widespread use and increase their sustainable development. Some of these solutions are based on memistors, memory devices that can store information without power consumption.
Memristors are electronic components based on nanodles that limit or regulate the flow of electric current in the circuit, and also register in advance how much energy has passed through it. Since they perform both calculations and storage of information, these devices can better perform the processing of information similar to the human brain.
Researchers from Université Paris-Sacplay-Cnrs, Université Grenoble-Alpes-Cea-leti, Hawai.Tech, Sorbonne Université and Aix-Maseille Université-Cnr developed a prototype Bayesian machine (i.e. a device based on Bayes theory) using Memristors. The system they proposed, which was presented in an article published in Nature electronicsIt turned out to be much more energy -saving than currently used hardware solutions.
Image of the complete Bayesian machine under the optical microscope
According to Damien Querlioz, one of the researchers who conducted the study: “Artificial intelligence makes significant progress today, but faces the challenge: its significant energy consumption. It is now well understood that this consumption comes from separation, in computers, between calculations and memory functions. Because artificial intelligence uses a lot of data, it requires high memory, which is expensive in energy. Memory functions are memory functions, because memory functions are memory functions, which are memory functions, which are memory functions, which are memorial functions.
“Until recently, memristors were developing technology and we could not be aware of complete systems with them,” Querliosis explained. “Now our team has built a Bayesian machine, a small artificial intelligence with memristors. The prototype includes 2048 Hafnium oxide memristors and 30080 silicon transistors (MOSFETS).”
Zoomed microscopic optical image of the Bayesian machine on one of the 16 Memristor Matrix. Photo: Damien Querlioz (CNRS/University Paris-Saclay)
The architecture of the machine was designed by writing Bayes's law in such a way that its implementation is natural in accordance with the principles of distributed memory and stochastic calculations, enabling the functioning of the circuit only using local memory and minimum data movement. Scientists use a hybrid complementary process-semiconductor/memristor and prove that a larger machine design is more energy-saving in the practical task of recognizing gestures than the standard implementation of Bayesian inference on the microcontroller. It can recognize some human gestures using thousands of times lower than a traditional hardware solution based on microcontroller. This Bayesian machine also provides immediate/turned off and is resistant to single failures.
“Most research on machine learning based on a memristor is aimed at implementing deep learning,” said Querlioz. “This is obviously an extremely important goal, because deep learning is so successful today. However, deep learning has some restrictions: its results cannot be explained and it does not work well when little data is available. Here we decided to implement Bayesian reasoning, an alternative AI approach, which does not cope well in the applications of large data, in which deep learning works so well, but we stand out in situations of small data, but Full explanations of the results. “
In the future, a Bayesian machine with a memristor created by this group of researchers can help improve the energy efficiency of AI models, and potentially inspire the development of other similar solutions. This can be particularly useful for critical safety applications, such as medical sensors and industrial safety monitoring circuits.
