Establishing Baselines for Photonic Quantum Machine Learning: Insights from an Open, Collaborative Initiative

TL;DR

This work establishes a baseline for photonic quantum machine learning by convening the open Perceval Challenge, which uses a full multiclass MNIST task within a hardware-feasible linear-optical framework. It aggregates thirteen approaches into end-to-end photonic feature extractors, photonic annotations, and model-tuning strategies, collectively showing strong experimental rigor and reproducibility through a single open repository. Although no heuristic quantum advantage is observed, the study maps the design space of photonic ML, demonstrates parameter-efficient architectures, and highlights hybrid quantum–classical pipelines as a promising near-term path. The findings have practical impact by providing community benchmarks, fostering collaboration across physics and AI, and informing future hardware development and benchmark design for quantum-enhanced learning.

Abstract

The Perceval Challenge is an open, reproducible benchmark designed to assess the potential of photonic quantum computing for machine learning. Focusing on a reduced and hardware-feasible version of the MNIST digit classification task or near-term photonic processors, it offers a concrete framework to evaluate how photonic quantum circuits learn and generalize from limited data. Conducted over more than three months, the challenge attracted 64 teams worldwide in its first phase. After an initial selection, 11 finalist teams were granted access to GPU resources for large-scale simulation and photonic hardware execution through cloud service. The results establish the first unified baseline of photonic machine-learning performance, revealing complementary strengths between variational, hardware-native, and hybrid approaches. This challenge also underscores the importance of open, reproducible experimentation and interdisciplinary collaboration, highlighting how shared benchmarks can accelerate progress in quantum-enhanced learning. All implementations are publicly available in a single shared repository (https://github.com/Quandela/HybridAIQuantum-Challenge), supporting transparent benchmarking and cumulative research. Beyond this specific task, the Perceval Challenge illustrates how systematic, collaborative experimentation can map the current landscape of photonic quantum machine learning and pave the way toward hybrid, quantum-augmented AI workflows.

Figures (43)

• Figure 1: Number of QML+MNIST papers per year, broken down by modality. There is a clear increasing trend in the number of publications that include the terms MNIST and quantum machine learning on their title or abstract over the years. This rise reflects the adoption of MNIST as a benchmark for testing novel quantum approaches.
• Figure 2: Three hybrid circuits trends observed in the challenge. When the photonic interferometer is used as a feature extractor, the model is trained end-to-end. When the photonic interferometer is used as a feature annotator, image or representations are passed through the encoder and their representations through the interferometer are fed as annotations to the encoder. In the case of model fine-tuning, a pretrained encoder is used and the photonic interferometer is used in the projection head, either for transfer learning, model refinement or self-supervised learning.
• Figure 3: Example photonic circuit diagram with $m=6$ modes. Photons are injected, pass through repeated beam-splitter (BS) layers, and accumulate phase shifts set by the PCA features. The measurement yields an $m$-mode photon-number distribution that encodes the feature vector.
• Figure 4: Accuracy vs. number of photonic modes $m$ for the sigmoid-transformed kernel.
• Figure 5: The trainable circuit is made of a first row of beam splitters, then $L$ blocks of Unitary matrices followed by circuits made with generic 2-modes circuits