Researchers at NUST MISIS have developed a mathematical method to verify the reliability of artificial intelligence when transferring knowledge between different tasks. The approach utilizes a new theorem that provides formal proof of the conditions under which knowledge transfer remains accurate, moving beyond traditional experimental testing.
The method adapts Hoare logic, a standard tool for verifying complex computer systems, to machine learning models. By applying this logical framework, developers can mathematically confirm that an AI model will retain its properties and accuracy after being retrained on new datasets, rather than relying solely on empirical observation.
According to Nikita Nikitin, a researcher at the Institute of New Materials at NUST MISIS, this framework is intended for high-stakes sectors such as medicine, industrial automation, robotics, and autonomous transport. The goal is to ensure AI systems are more predictable and secure before they are deployed in real-world environments.
The study results were published in the Scientific Reports journal. The research project was supported by a grant from the Russian Science Foundation under project number 23-73-00131.
