Neural network in the glass. Does not require power, recognizes numbers

We are all familiar with the ability of neural networks, like handwriting recognition. The fundamentals of this technology have existed for many years, but, only relatively recently, a leap in the area of ​​computer power and parallel data processing made it possible to make a very practical solution from this technology. Nevertheless, this practical solution, basically, will be presented in the form of a digital computer repeatedly changing the bits, in the same way as in the performance of any other program. But in the case of a neural network developed by researchers from the universities of Wisconsin, MIT, and Columbia, the situation is different. They created a glass panel that does not require its own power supply, but at the same time capable of recognizing handwritten numbers .

This glass contains precisely positioned inclusions, such as air bubbles, graphene impurities and other materials. When light falls on glass, complex wave patterns arise, as a result of which light becomes more intense in one of ten areas. Each of these areas corresponds to a number. For example, below are two examples showing how light travels when recognizing the number “two”.


')
With a training set of 5,000 images, the neural network is able to correctly recognize 79% of the 1000 input images. The team believes that they could improve the result if they could bypass the limitations caused by the glass production process. They started with a very limited device design to get a working prototype. Further, they plan to continue exploring various ways to improve the quality of recognition, while trying not to overly complicate the technology so that it can later be used in production. The team also has plans to create a three-dimensional neural network in the glass.