Russian Scientists find way to speed up neural network training up to 20-fold

MOSCOW, September 9 (Sputnik) – Russian scientists have found a way to speed up neural network training by 15-20 times by creating an automatic mechanism for analyzing and correcting errors during training, Russian AI software developer Cognitive Pilot told Sputnik.

“Experts from Cognitive Pilot [a subsidiary of Sber and Cognitive Technologies] were able to solve a complex problem that allows for a significant increase in the efficiency of neural network training by creating an automatic mechanism for analyzing and correcting errors made by the neural network during training. The technology also allows for a 15-20-fold reduction in training time (taking into account data selection),” the company said.

The new technology is called Cognitive Neural Network Hospital because it is capable of “curing the sore spots” of a neural network. This happens in two stages: first, the technology identifies the weakest areas, and then, thanks to the information received, new data is revealed that does not exactly match, but is similar to the errors made. It helps to analyze the types of errors and classify them into groups.

Cognitive Neural Network Hospital allows increasing the efficiency of AI training by 40%, the experts noted.