ECE Seminar: Gerasimos Angelatos

ECE Seminar: Gerasimos Angelatos

Title: Finding Signal in the Noise: Resolvable Capacity and Learning with Physical Systems

Abstract: The ability to perform computation with physical systems is unavoidably limited by the presence of noise in their extracted outputs. Though ubiquitous across classical and quantum regimes, from photonic neural networks to quantum computers, the precise quantification of how noise impacts learning, and how to mitigate its effects, remains poorly understood. In this talk, I will describe a mathematical framework developed in Ref. [1] for evaluating the resolvable expressive capacity (REC) of general physical systems subject to noise in their outputs. An important consequence is that structured noise — such as quantum sampling noise — naturally defines a hierarchical basis of functions, called eigentasks, which a given system can approximate with minimal error. I will describe how the REC of any physical system, and thus its ability to process information in the presence of noise, can be quantified in terms of its finite set of eigentasks. Furthermore, this construction naturally establishes a simple supervised learning rule: learning in a truncated eigentask basis minimizes generalization error for noisy physical systems. I will share results from the application of this framework to quantum machine learning and low-light image sensing.

[1] F. Hu, G. Angelatos et al., Phys. Rev. X 13, 041020 (2023), https://doi.org/10.1103/PhysRevX.13.041020

Bio: Dr. Gerasimos Angelatos is a Senior Research Engineer at RTX BBN Technologies in Cambridge MA, where his research interests are quantum computing hardware, and learning or optimization with physical systems. He earned his PhD at Princeton University in the Electrical and Computer Engineering Department in 2022, studying quantum machine learning with Dr Hakan Tureci.