AlbaNova and Nordita colloquium

Glasses, Chaos, and Neural Networks: A Unified Physical Perspective 

Prof. Viktor Galitski (University of Maryland)

18 december 2025, 15:15 - The Oskar Klein auditorium (FR4)

This talk will review our recent work on classical and quantum glasses - ubiquitous systems where strong frustrated interactions prevent them from settling into a simple state. I will begin with spin glasses from the perspective of chaos theory and introduce the mean-field formalism of Thouless, Anderson, and Palmer (TAP) to "visualize" the rugged landscape of glassy metastable minima. The central theme of the talk is a one-to-one correspondence between classical spin models and neural networks (NNs), which allows us to transplant spin-glass theory directly into the study of learning. In this mapping, training a NN corresponds to a family of spin Hamiltonians parameterized by training time and physically implies the destruction of the spin glass and the emergence of hidden order associated with the classification task. This provides an appealing, universal physical picture of why certain neural networks work, as well as a natural scheme for their quantization. I will introduce a broad class of quantum neural networks and show their successful experimental realization on current quantum hardware, including IBM transmon systems and two types of trapped-ion quantum computers.