Faculty Sponsor: Tsampikos Kottos
Ethan Valentino
Ethan is a rising senior studying Physics and Philosophy from Beacon Falls, CT. He plans to earn an MA in physics from Wesleyan after graduating. In his free time, he enjoys playing basketball, rock climbing, and reading.
Abstract: Artificial neural networks (ANNs) have achieved remarkable success at a variety of machine learning tasks, and have thus become the foundation of modern artificial intelligence. Their success has motivated the study of physical neural networks (PNNs), which replace the heavy digital computational load of ANNs with a physical system capable of performing nonlinear operations at a fraction of the energy and time. Here, we realize a PNN using a compact graph of interconnected coaxial cables that produces wave-chaotic scattering. The rapidly varying frequency dependence of the scattering matrix provides a large set of effectively uncorrelated nonlinear responses within a narrow bandwidth, enabling computational complexity to scale through frequency diversity rather than hardware size. Using probe frequencies as physical trainable parameters, our four-vertex, fully connected network achieves higher accuracy on the MNIST digit-classification task than other PNN proposals. We also demonstrate that the Universal Approximation Theorem holds for our system, meaning it is expressive enough for any machine learning task.
2026_poster_final