The Moscow Automobile and Road Construction Institute has developed a mathematical tool that allows quantum computers to train neural networks hundreds of times faster than classical systems.

The author of the development is Caesar Pronin, a senior lecturer at MADI. He proposed a method for creating a “quantum oracle” — a navigator algorithm that directs computations directly to the correct answer, bypassing a long enumeration of options.

Grover's algorithm is used to find solutions in quantum computing. It evaluates many options simultaneously but requires an “oracle” — a function that recognizes the correct solution and enhances its probability. The main problem is that quantum operations must be strictly reversible. Auxiliary calculations have to be “folded” without loss of information.

Pronin used a reversible quantum multiplier as the “oracle”. For this, he systematized the matrix apparatus and built quantum analogues of basic logical operations — NOT, AND, OR, XOR, as well as a full adder. On this basis, the concept of training the basic element of a neural network — a quantum perceptron, where the “oracle” is built on an activation function — was formulated.

Calculations showed that Grover's algorithm with the new component finds a solution approximately 326 times faster than classical brute-force search. An example is a 16-qubit register. This lays the theoretical foundation for applied problems in industry where large amounts of data need to be processed in fractions of a second.

The next stage is scaling the approach for 2–3-qubit registers, adapting it for simulators and NISQ devices (intermediate-scale quantum computers), and addressing issues of quantum arithmetic bit depth.