Berlin 2024 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 2: Machine Learning & Physics

AKPIK 2.4: Vortrag

Mittwoch, 20. März 2024, 15:45–16:00, MAR 0.002

PSeudocode Projective Simulation (PS^2) — •Marius Krumm and Hans J. Briegel — University of Innsbruck, Institute for Theoretical Physics, Technikerstr. 21a, A-6020 Innsbruck, Austria

The rise of deep learning has enabled significant progress in technology and science. However, the opaque oracle-like nature of artificial neural networks severely limits their potential to discover new qualitative scientific insights. This motivates the exploration of methods from eXplainable AI (XAI) that allow to understand the reasoning process of accurate predictions. In this talk, I present a new XAI method called PSeudocode Projective Simulation (PS^2) which represents chains-of-thought in the form of pseudocodes. Here, a thought is modeled as a small data processing module acting on the agent's memory, which can be a trainable neural network. These subroutines are selected in a trainable random walk, making our method an extension of Projective Simulation. On a technical level, methods from hierarchical and safe reinforcement learning are adapted and integrated into our setting. These modified methods help to model domain knowledge about the nature of the thoughts and the environment. The framework is applied numerically to the Highway-Env.

Keywords: eXplainable AI (XAI); Reinforcement Learning; Hierarchical Reinforcement Learning; Projective Simulation; Modular neural networks