SAMOP 2023 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

QI: Fachverband Quanteninformation

QI 23: Poster II (joint session QI/Q)

QI 23.3: Poster

Wednesday, March 8, 2023, 16:30–19:00, Empore Lichthof

Quantum low-precision neural networks and their classical counterparts — •Felix Soest, Konstantin Beyer, and Walter Strunz — Institut für Theoretische Physik, Technische Universität Dresden, Dresden, Germany

With increasing accessibility of quantum computing devices and the successes of classical machine learning, efforts have been made to combine the two. Whether using quantum resources can provide an advantage to trainability or generalisability remains an open question, as the size of classical neural networks is much larger than what current quantum technologies can offer. Moreover, a clear indication of a quantum advantage is usually hard to identify. An often considered ansatz is that of parametrised unitaries, where the quantum machine learning model comprises multiple layers the parameters of which are trained classically. It has recently been shown that these models have classical surrogates [1], allowing for a classical benchmark to compare these models to. However, classical feed-forward neural networks can in general not be mapped to unitaries, in part due to the lack of irreversibility. Therefore we aim to construct a framework using intermediate measurements which has a classical counterpart. The resulting network is a parametrised quantum channel that allows us to reproduce classical low-precision networks as a special case. Allowing for quantum operations in this framework extends the classical regime, providing a good benchmark.

[1] Schreiber et al. arXiv:2206.11740