Russian scientists have developed a trainable neural network made of plastic

The creation of new systems with artificial intelligence remains one of the most promising areas of modern science. Joint efforts of Russian physicists from the Kurchatov Institute, Moscow Phystech, St. Petersburg, Moscow Universities and their Italian colleagues from the University of Parma have created a neural network based on polymer memristors.

For information, memristor and resistor are analogs. The main difference between a memristor and its “brother” is that it is capable of changing its properties depending on the magnitude of the external signal. In this, they remarkably resemble synapses - connections between two neurons in the brain, which also flexibly change the efficiency of signal transmission.

Polyaniline polymer was chosen as a material for the manufacture of memristors. Then the scientists started training the artificial neural network they created. For this, impulses were fed to its input in a random sequence. In case of an incorrect answer, a special corrective impulse was immediately given. The process of training the network consisted of multiple repetitions - at least 15 times - of such operations, after which the system was adjusted to the "desired mode" and was able to perform basic logical operations.

Researchers are only at the beginning of the journey. Ahead, they have to minimize the size of memristors to a size of 10 nanometers, to increase the speed of their response to incoming signals. Only after that it will be possible to talk about the practical application of the development.