Babak Rahmani, Microsoft Research

Leuchs-Russell-Auditorium, A.1.500, Staudtstr. 2

Location details

Abstract:

This talk explores research that bridges physical systems and machine learning through two converging directions. The first part describes work on an analog optical computer (AOC) that implements machine learning tasks via a physical dynamical system. In this approach, the equilibrium state of the system is used to encode data for tasks such as classification, nonlinear regression, and combinatorial optimization. Experiments with the AOC indicate that the physical architecture can be made robust to analog noise while providing a representation that generalizes more consistently compared to traditional feedforward networks. The second part of the talk examines how the dynamical system abstraction from the AOC can be translated into large language model architectures [1]. By reformulating a language model as a dynamical equation, it is possible to extend the computation budget available for iterative reasoning tasks such as multi-step problem solving.  A discussion will follow on design choices, algorithmic trade-offs, and the use of physical principles to improve generalization and robustness in both analog platforms and digital models.

[1]: Schöne M, Rahmani B, Kremer H, Falck F, Ballani H, Gladrow J. Implicit Language Models are RNNs: Balancing Parallelization and Expressivity. arXiv preprint arXiv:2502.07827. 2025 Feb 10.