Stochastic Separation Theorems: How Geometry May Help to Correct AI Errors
Alexander N. Gorban, Bogdan Grechuk, Ivan Tyukin · Notices of the American Mathematical Society · 2022
AI instabilities and adversarial examples—errors due to minor changes in data or structure—have recently been found in many advanced data-driven AI models. Mounting evidence suggests that these errors are in fact expected in such systems and may not always be cured by larger volumes of data or better training algorithms as long as the AI architecture remains fixed. If errors are inevitable in data-driven AI, then how do we deal with them once they occur?