A Quandela led European team showed that a programmable silicon photonic processor can run a quantum machine learning classifier and fully reconstruct the quantum states it produces, using one fixed measurement basis instead of the exponentially many that standard quantum state…
A programmable light-based quantum processor has shown it can both run a machine-learning classifier and fully reconstruct the quantum states it produces, using one fixed way of measuring the chip's output instead of the exponentially many that standard characterization would demand.
The result, from French quantum hardware vendor Quandela with the Center for Theoretical Physics of the Polish Academy of Sciences and the University of Warsaw, comes via a Quantum Computing Report summary of work carried out under the EU's Horizon Europe QUONDENSATE Pathfinder project. The genuine news is not that the team ran a quantum machine-learning workload on a chip. Many groups have done that. The portable insight is the cost of measuring what the chip actually computed.
Quantum state tomography, the process of fully reconstructing what a quantum system is doing, is normally the throttling bottleneck in photonic quantum experiments. To pin down the full state of n entangled modes, a researcher typically has to collect measurement statistics across an exponentially growing set of measurement bases, the different "angles" from which a quantum system can be observed. Doubling the system size roughly squares the measurement budget. At the 24-mode scale of Quandela's Belenos chip, the standard approach is already straining.
The Quandela-led experiment takes a different route. Belenos is a programmable silicon photonic processor built from a mesh of Mach-Zehnder interferometers, the standard optical elements that split and recombine light, fed by single photons from a semiconductor quantum-dot source and read out by photon-number-resolving detectors that count how many photons arrive. The team trains only a thin classical software layer on top, in the paradigm known as Quantum Reservoir Computing (QRC), where the chip's fixed physics does the heavy lifting and only the readout is updated.
The single-measurement-basis property falls out of that architectural choice. Because the chip's analog dynamics are fixed and the software layer is a linear map, the team can recover complete tomographic information, including full state reconstruction and multi-mode entanglement characterization, from one fixed measurement basis applied to the chip's output, rather than sweeping through an exponential set of bases. The same hardware can run a classical-style classification task (telling apart classes of quantum states or input signals) and a quantum information-processing task that requires genuine entanglement, and both come with built-in diagnostics.
The caveats sit close to the result. The 24 modes here are a demonstration scale, not a production scale, and the demonstration is one group's experimental work. Quantum reservoir-style machine learning does not carry the same trainability story as parameterized quantum circuits, where the quantum gates themselves are tuned during learning, so comparisons to variational or gate-model quantum machine learning should be drawn carefully. The "scalable" label applies to the architecture and the tomography method, not yet to large-qubit or fault-tolerant workloads, and replication on larger photonic arrays plus head-to-head benchmarks against classical machine-learning baselines are the natural next tests.
The work also has a clear policy wrapper. QUONDENSATE is an EU Horizon Europe Pathfinder project, the EU's early-stage research funding line for high-risk, high-reward ideas, with Quandela, the Polish Academy of Sciences' theory center, and the University of Warsaw as the named partners. That institutional scaffolding is part of why a 24-mode photonic experiment can sit alongside state-reconstruction claims: the consortium is built to push the measurement-cost question, not just ship another chip demo.
What to watch next: whether the single-basis tomography property survives the move from 24 modes to the next-generation photonic arrays the consortium is targeting, and whether independent groups can replicate the cost-scaling claim on chips built outside the Quandela stack.