An international team including QLAB Director Prof. Magdalena Stobińska-Moretto and Michał Siemaszko (PhD student at the University of Warsaw) has demonstrated the first neuromorphic computing architecture built on a photonic quantum memristor. The work has been accepted for publication in PRX Quantum.
Neural networks rely on nonlinearity – something quantum systems, whose evolution is intrinsically linear, provide only at the high cost of entangling gates. The team’s answer is a quantum memristor: a Mach-Zehnder interferometer in a femtosecond-laser-written photonic chip whose internal phase is updated according to the photons detected at one of its outputs. This feedback loop supplies both nonlinearity and a short-term memory, turning the chip into a quantum reservoir; only a simple linear-regression readout has to be trained.
Using single photons at 1550 nm, the researchers tested the device on nonlinear function prediction and on three time series – NARMA, Mackey-Glass and the Santa Fe laser data. Switching off the memristive feedback degraded accuracy markedly (by 34% on average for the nonlinear function), while the quantum reservoir matched or outperformed classical models with just one trainable parameter.
The University of Warsaw’s contribution covered the theory and algorithms behind the time-series predictions. The experiment was carried out at the University of Vienna with photonic chips from IFN-CNR and Politecnico di Milano.
M. Selimović, I. Agresti, M. Siemaszko et al., Experimental neuromorphic computing based on quantum memristor, PRX Quantum (accepted). DOI: 10.1103/jknv-3tx7 · Preprint: arXiv:2504.18694