Mykola Lukashchuk - Information Geometry and Message Passing - 12th November 2025 скачать в хорошем качестве

Mykola Lukashchuk - Information Geometry and Message Passing - 12th November 2025

Abstract: The recent advancements in information-geometric learning approaches cannot be overstated. Many deterministic algorithms, such as Quasi-Newton optimization methods and the ADMM optimization scheme, have been recovered and unified as special cases of the natural gradient or information-geometric gradient descent, demonstrating the usefulness of this generalization. Moreover, better variants of these schemes—both in efficiency and accuracy—have been proposed. Information geometry concentrates on the geometry of the variational family. However, it overlooks the geometry of the model in message passing terms—specifically, the graph connectivity. In this seminar, I will discuss the paradigmatic differences between these two approaches: where they benefit from each other, where they are incompatible, and how to overcome these differences in certain scenarios to build geometrically informed natural intelligence.