HSE University researchers have mathematically substantiated an easy way to check how confident a classical machine learning system is in its result. The method requires fewer computational resources and helps to distinguish a reliable answer from an assumption that should not be trusted without additional verification.

The development concerns stochastic gradient descent – a popular algorithm that searches for an optimal solution with elements of randomness. Therefore, a single output number is not enough: it is important to know the confidence interval, i.e., the range where the correct result is highly likely to be found.

Previously, such verification required complex statistical calculations, and sometimes even retraining the program. The new approach allows avoiding this. It has already been applied in practice, and now researchers have explained why it works and determined the limits of its reliability.

This is especially important for medicine, finance, and autonomous systems, where it is dangerous to rely on a confident but erroneous answer. The method can simplify and reduce the cost of many classical machine learning algorithms that do not use neural networks.