Speaker
Description
Quantum reservoir systems offer a promising route to mitigate some of the optimization challenges affecting variational quantum machine-learning models, including trainability issues. In reservoir computing, the quantum dynamics are kept fixed, while only a classical readout is trained, making the approach particularly suitable for near-term photonic implementations. However, only a few experimental platforms have so far demonstrated the feedback loops required to provide memory and enable the processing of time-series data.
In this work, we present a photonic quantum reservoir protocol for time-series forecasting implemented on a reconfigurable integrated photonic chip operating in the multiphoton regime. The platform enables programmable feedback and allows us to study how feedback dynamics and quantum correlations affect computational performance. We show that the presence and depth of feedback loops determine the reservoir memory capacity, allowing the system to retain information about past inputs and capture temporal correlations. At the same time, indistinguishable multiphoton states introduce quantum interference effects that enrich the reservoir dynamics and enhance nonlinear expressivity.
This increased expressivity enables the system to approximate higher-order nonlinear functions and improves performance compared with distinguishable photon inputs. We benchmark the reservoir on standard tasks, including temporal XOR, NARMA sequences, and Mackey–Glass chaotic time-series prediction, observing enhanced performance in regimes where both memory and nonlinearity are required.
Overall, these results demonstrate that photonic quantum reservoirs can be tuned to balance memory and expressivity, providing a scalable architecture for time-series forecasting carried out with integrated quantum photonic platforms.
| Topic | Hybrid quantum-classical approaches for learning and optimization |
|---|