Quantum inference on a classically trained quantum extreme learning machine

Quantum extreme learning machines (QELMs) are unconventional computing architectures that bear remarkable promise in both classical and quantum machine-learning tasks, such as the estimate of quantum state properties. However, the probabilistic nature of quantum measurements demands extensive repetitions for training to precisely estimate expectation values, imposing stringent trade-offs among experimental resources, acquisition time, and signal-to-noise ratio, particularly for large datasets. Here we introduce a paradigm shift by harnessing the correspondence between stimulated and spontaneous emission. The QELM is trained exclusively with intense classical fields, yet it performs inference directly on previously unseen quantum input states to predict their quantum properties. This strategy dramatically reduces acquisition times while substantially enhancing the signal-to-noise ratio. Using frequency-bin-encoded biphoton states, implemented here for the first time in a quantum machine-learning architecture, we demonstrate entanglement witnessing of two-qubit states with (93 +- 4)% accuracy, multi-dimensional entanglement detection, and learning of the Hamiltonian governing photon-pair generation with a fidelity of (96 +- 4)%. By establishing classical training as a scalable route to quantum feature extraction, our results bridge macroscopic observables and nonclassical correlations, opening a new pathway toward faster and more robust quantum neural networks

💡 Research Summary

The paper introduces a novel training paradigm for quantum extreme learning machines (QELMs) that dramatically reduces the experimental overhead traditionally associated with quantum‑machine‑learning architectures. In standard QELMs, the output layer consists of expectation values of observables measured on quantum states after evolution through a reservoir. Because quantum measurements are probabilistic, each training example must be repeated many times to obtain reliable statistics, leading to long acquisition times and low signal‑to‑noise ratios (SNR), especially for large training sets.

The authors exploit the well‑known correspondence between stimulated and spontaneous emission in nonlinear optics. In a third‑order medium, spontaneous four‑wave mixing (SpFWM) generates entangled photon pairs at a very low rate, whereas the same medium driven by a strong coherent seed undergoes stimulated four‑wave mixing (StFWM) and produces a bright coherent beam whose spectral intensity is proportional to the spontaneous coincidence probability. By shaping the seed to match the “asymptotic output field” of a given mode (obtained by back‑propagating the detection mode through the system), the intensity Iₖⱼ measured on the stimulated beam directly maps onto the quantum correlation Cₖⱼ that would be obtained from coincidence counting.

Training therefore proceeds entirely with classical light: for each input mode j a properly amplitude‑shaped seed is injected, the StFWM output spectrum is recorded with an optical spectrum analyzer (OSA), and the set of intensities {Iₖⱼ} is fed into a linear regression to determine the output‑layer weights. Because the classical signal is many orders of magnitude brighter than the spontaneous photon‑pair flux, the required integration time drops from hours (or days) to seconds, and the SNR improves dramatically.

After training, the same reservoir is used for inference on genuine quantum inputs generated by SpFWM. The biphoton state propagates through the reservoir, and coincidence counts Cₖⱼ are collected with superconducting nanowire single‑photon detectors (SNSPDs). The previously learned weights are applied to the vector of measured coincidences, yielding estimates of target quantities such as an entanglement witness ⟨W⟩, high‑dimensional entanglement metrics, or parameters of the photon‑pair generation Hamiltonian. This constitutes a form of transfer learning: a model trained on macroscopic classical data is directly reused for quantum data without retraining.

Experimentally, the authors implement the scheme on a silicon photonic chip. Frequency‑bin encoding is used to define biphoton states across multiple spectral modes. Two electro‑optic modulators (EOMs) and wave‑shapers (WS) constitute the reservoir, providing tight‑binding‑type interactions that scramble the bins and expand the effective Hilbert space up to 64 dimensions. Training with classical StFWM reduces the acquisition time by roughly three orders of magnitude compared with conventional coincidence‑based training.

Performance is benchmarked on three tasks: (i) two‑qubit entanglement witnessing, achieving 93 % ± 4 % accuracy; (ii) multi‑dimensional entanglement detection, reaching 96 % ± 4 % accuracy; and (iii) learning the Hamiltonian governing photon‑pair generation, with a fidelity of 96 % ± 4 %. These results match or exceed those obtained with fully quantum‑trained QELMs, demonstrating that classical training does not sacrifice predictive power.

The work establishes a scalable route to quantum feature extraction: by leveraging the stimulated‑spontaneous duality, one can train large‑scale QELMs with modest experimental resources while retaining the ability to infer non‑classical properties from quantum inputs. This approach opens new avenues for fast, robust quantum neural networks, quantum sensing, and potentially for other platforms (e.g., superconducting circuits, atomic ensembles) where analogous stimulated‑spontaneous correspondences exist. Future directions include optimizing reservoir topologies, extending to multi‑input/multi‑output architectures, and integrating error‑mitigation strategies to further enhance the practicality of quantum‑classical hybrid learning systems.