Spintronics appeals to artificial intelligence
Belonging to the field of artificial intelligence, artificial neural networks are computer systems capable of solving complex problems. How they work is inspired by the human brain. Their abilities, acquired through machine learning methods, are now used to implement facial recognition, develop autonomous cars and operate social networks. However, because it is difficult for these systems to learn new skills efficiently, scientists are relying on spintronics (or spin electronics) to unlock their capabilities. This technology, which revolutionises the binary system used in computers, exploits the electrical charge of electrons and a quantum property called spin.
Recognising a face is one of the basic tasks that a human being is capable of performing. Unfeasible for a computer fifty years ago, such a task is today possible thanks to the development of artificial intelligence (AI). Artificial neural networks are the source of this evolution. These special algorithms mimic the way the brain works to solve complex problems. "The human brain has been a source of inspiration," according to Liza Herrera Diez, a researcher specialising in spintronics and Damien Querlioz, a researcher in artificial intelligence at the Centre for Nanoscience and Nanotechnology (C2N - Univ. Paris-Saclay, CNRS, CEA, Univ. Paris Cité). “We have transposed the way the brain computes into electronic components, so that they work in a similar way."
The catastrophic forgetfulness of AI
However, this artificial copy is not yet perfect. "All current artificial intelligences have the same limitation: they have to learn everything all at once. Every time they learn something new, they almost instantly forget what they knew before. This makes them very different from the human brain, which assimilates information throughout its life," explains Damien Querlioz. The two C2N scientists have been working together for years to develop neuromorphic computing, which is computer processing that acts like the human brain, and more efficient artificial synapses. Liza Herrera Diez adds: "We want to find a solution to counteract the phenomenon known as catastrophic forgetfulness. This would mean that an AI trained to recognise images of cats, and then improved to be able to identify images of dogs, would not forget its first skill. This would decrease the volumes of data needed for learning."
Catastrophic forgetfulness hinders the development of tools that exploit neural networks. "One of the most exciting future applications of AI is the design of connected medical devices, for example to help a diabetic adjust their insulin level," predicts Damien Querlioz. For this, the AI will monitor the patient for years and continuously learn to adapt to changes in their medical profile. It will soon be able to do so thanks to the innovations in spintronics developed at C2N.
Mini spinning tops that never stop spinning
Spintronics, a discipline that appeared in the second half of the 20th century, focuses on a quantum characteristic of electrons called spin. The analogy developed by physicists to explain this property, which can be observed on a scale smaller than that of an atom, is to imagine electrons as spinning tops. Spin is a way of characterising this movement, which for an electron, can only take two orientations, either upwards or downwards.
This quantum property has found a practical application in our daily life. A modern computer is made up of electronic components. These are mainly comprised of semiconductor materials, but also of components that exploit the magnetic properties of spin, such as hard disk read heads.
In a computer, information is stored, processed and transmitted in the form of binary code, a succession of zeros and ones. This coded language is adapted to the electrical signals that drive computer activity. Like light messages sent with a torch, there can only be two states of information: the light is off (0) or on (1).
Welcome to the spintronics valley
Manufacturers quickly came up against the limited storage capacities of electronic systems. To remedy this, spintronics, which appeared in computer science in 1997, exploited the discovery of giant magnetoresistance less than ten years earlier. By introducing metallic layers a few nanometres thick in the hard disk read heads, the binary language could be written magnetically rather than electronically. By applying an electric current or a magnetic field, the orientation of the electron spins was forced up or down, creating an alternation of 0 and 1. "The new types of devices developed by spintronics have a novel physicochemical composition. We want to use this to develop new and even more powerful features," explains Liza Herrera Diez.
Storage devices currently generate an electric current or a magnetic field to change the spin of electrons. The researcher is studying a technique based on electric fields. Rather than acting directly on the ferromagnetic metal nanolayer, she adds a layer of ion-rich oxides. When an electrical voltage is applied, the ions in the oxide layer are transferred to the metal layer and change its magnetic properties. "This consumes less energy than acting directly on the metal material. As opposed to a volatile charge effect, magneto-ionics is permanent, since the displaced ions do not spontaneously return to their original position."
Bringing nuance to the binary
The new properties provided by ions create a very interesting possibility for AI: nuance. What if computer systems no longer simply alternated between two binary states, but also incorporated hidden states? Damien Querlioz has shown theoretically that the introduction of nuances, such as 0+, 1- or 1+++, would solve the learning problems of AIs. Liza Herrera Diez's work on spintronics offers the possibility of integrating this abstract capability into concrete artificial neuron networks. Indeed, when storing new information in the spintronic device, changes occur in the binary code: zeros become ones, and vice versa. Ion exchange between the oxide layer and the metal layer slows down or accelerates this change. These variations then produce hidden binary states and bring that much sought after faculty, nuance. "We will create a demonstrator to show that a practical application is possible," say the two C2N scientists.
Interdisciplinarity allows value-creation
To develop this demonstrator, the team is relying on the European METASPIN project, which in May 2022, won funding from the European Innovation Council (EIC Pathfinder) for four years. This project has eight partners, including research laboratories specialising in oxidation chemistry and computational cognitive sciences, a German manufacturer that designs magnetic materials, along with two French start-ups, Spin-Ion Technologies, a C2N spin-off, and HawAI.tech, which study spintronics and AI respectively. The scientists explain: "C2N designs algorithms and nanodevices, but this project requires other interdisciplinary skills: we need to characterise oxidation at the interface and in the ferromagnetic material, design software for the AI, consider the next generation of devices through the study of anti-ferrous materials, etc." This is why the C2N team plans to accommodate PhD thesis students and post-doctoral researchers in relation to this project.
Soon, new artificial synapses will perhaps remedy the amnesia attacks of artificial intelligences.
- L. Herrera Diez, R. Kruk, K. Leistner, J. Sort. Magnetoelectric materials, phenomena, and devices. APL Mater. 9, 050401 (2021).