HOW IT WORK


Modern approaches, using ultra-precise neural nets, make it possible to achieve excellent results when working with good-quality source data, but a lot of quality is lost when the data is noisy (out of focus, blurred, or afflicted with other kinds of noise). We have created our own recognition algorithms for these noisy images, employing a mathematical toolkit based on Markov chains and random metrical spaces.


Unlike the majority of machine learning methods, for us a high dimensionality of the feature space is not a difficulty (the "curse of dimensionality") but rather an opportunity to obtain new results.

WHAT IS THE RESULT


For the well-known MNIST database of handwritten numerals we create a software system that can retain a low error rate (20% or less) even with intensive noise (50-60%) in the source data.



THE MNIST NUMERALS EXAMPLE


Contact Us