There is a fairly rare form of matter in the world called spin glass. For a long time, this form was a purely theoretical concept, but recently a group of researchers from Los Alamos National Laboratory finally succeeded in creating spin glass in reality, and, moreover, using electron beam lithography technology to fabricate a variant of artificial neural networks. In other words, spin glass can be used to produce algorithms implemented as hardware rather than software code. This, in turn, will allow heavy computational tasks, such as image recognition, to be performed with minimal energy consumption.
Spin glass can be thought of as a system of randomly oriented nanomagnets whose orientation is the result of random interactions of attraction and repulsion between the opposite poles of the nanomagnets and the poles of the same name. However, the whole system is not in a stable and balanced state, so any changes in it are still possible. Moreover, as the temperature drops, the dynamic and thermodynamic characteristics of the system increase so much that it becomes possible to use the entire system for computational operations.
"Theoretical models involving spin glass are widely used in complex systems describing, for example, some of the functions of the brain, the principles of information coding and error correction or the dynamics of changes in stock exchanges," the researchers write, "This interest in spin glass has provided us with a strong motivation and we have been successful in creating the first samples of such artificial material."
The creation of spin glass was the result of complex preliminary theoretical research and experiments. While creating this form of matter, scientists simultaneously created so-called Hopfield's neural network in its structure, whose principle of operation resembles that of associative memory. Such networks usually encode two or more memory images which are associated with the same object. So when a piece of information about an object, a small fragment of a portrait for example, is fed to the network, the network immediately "recalls" the whole object, a complete portrait of the person in that case.
Memory images in Hopfield networks are encoded in the form of the structure of the network itself, so already trained (pre-trained) neural networks can be embodied in a spin-glass. However, such neural networks are less susceptible to interference and noise, and although they produce "dirtier" results than traditional algorithms, they are more flexible for applications of some kinds from the field of artificial intelligence.
