SKM 2023 – wissenschaftliches Programm

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

TT: Fachverband Tiefe Temperaturen

TT 13: Focus Session: Physics Meets ML II – Understanding Machine Learning as Complex Interacting Systems (joint session DY/TT)

TT 13.4: Vortrag

Montag, 27. März 2023, 16:30–16:45, ZEU 250

Interpreting black-box ML with the help of physics — •Miriam Klopotek — University of Stuttgart, SimTech Cluster of Excellence EXC 2075, Stuttgart, Germany

Complexity is an unavoidable part of systems with emergent or even so-called intelligent capabilities. Ultimately, it stems from the many microscopic constituents with multiple possible states, which introduces a vast space of degrees of freedom. This is true both for many-body systems as well as modern machine learning (ML) systems. Today, the latter suffer notoriously from the ‘black-box problem’, i.e. they are inherently opaque. We argue that an engagement with physics can offer deep insights ultimately for a theory of operation and thus an interpretation, as well as powerful ways to assess their reliability and shortcomings. We show some results for a case study with beta-variational autoencoders (β-VAEs) [1], which we trained on data from a well-characterized model system of hard rods confined to 2D lattices [2].

[1] D. P. Kingma and M. Welling, ICLR 2014. D. J. Rezende, S. Mohamed, and D. Wierstra, ICML 2014, p. 1278-1286.
[2] P. Quiring, M. Klopotek and M. Oettel, Phys. Rev. E 100, 012707 (2019).