Tuesday 19 Nov 2019: Dynamics Seminar: Stochastic Separation Theorems and One-trial Learning and Error Correction of Legacy AI systems
Ivan Tyukin - University of Leicester
LSI Seminar Room A 13:30-15:30
In this talk we shall consider the problem of mistakes in Artificial Intelligence (AI) systems and motivate a mathematical framework for simple, real-time, computationally-efficient, and non-iterative improvements of these systems. The improvements are, in essence, shallow networks constructed on top of the existing AI computational architectures. Theoretical foundation of the technology is based on Stochastic Separation Theorems and the ideas of measure concentration. We show that, subject to mild technical assumptions on statistical properties of internal signals in the original AI, with probability close to one the technology enables instantaneous ''learning away'' of spurious and systematic errors. The method will be illustrated with applications to image processing and face/human shape recognition/detection. In addition to these, the framework has a potential to explain rapid encoding of new memories by individual neurons in the human brain.