# Techniques: Photonic Quantum Convolutional Neural Networks with Adaptive State Injection

## What the paper contributes (one paragraph)
The paper designs and experimentally demonstrates the first photonic quantum convolutional neural network (PQCNN), a quantum machine learning architecture built from particle-number-preserving linear optical circuits with adaptive state injection (arXiv:2504.20989). It maps the subspace-preserving QCNN of prior work onto photonics: a tensor-encoding quantum data loader, a beam-splitter convolutional layer, a state-injection-based pooling layer that supplies nonlinearity, and a linear optical dense layer. The architecture is argued to give polynomial speed-ups over classical CNNs and to mitigate barren plateaus through Hamming-weight preservation. It is validated on a quantum-dot single-photon source with 8-mode and 12-mode integrated interferometers, performing binary image classification on a custom Bars-and-Stripes dataset, with state injection emulated by post-selection. Scaling is studied numerically on larger datasets including MNIST, and the authors release an open-source PyTorch toolkit for subspace-restricted simulation of photonic QML.

## Techniques offered
- **Subspace-preserving photonic quantum convolutional neural network (PQCNN architecture)** - a particle-number-preserving linear optical QCNN with tensor encoding, beam-splitter convolution, state-injection pooling, and a linear dense layer, supplying a near-term photonic QML classifier with reduced depth and parameter count.
  - guarantee: none proven; claims a polynomial running-time speed-up over classical CNNs and argues (not proves) barren-plateau mitigation via Hamming-weight/subspace preservation.
  - quote: "The proposed architecture offers polynomial speed-ups over classical deep learning architectures and uses convolutional and measurement-based pooling layers to preserve quantum state symmetries while introducing nonlinearity."
- **Adaptive state injection (measurement-based nonlinearity for linear optics)** - a measurement-and-feedforward gadget that injects fresh single photons conditioned on detection outcomes, supplying nonlinearity and added controllability to otherwise linear optical circuits while preserving particle number.
  - guarantee: none (the pooling nonlinearity is structural; no formal bound or soundness property is established for the injection itself).
  - quote: "The resulting architecture alternates linear optical layers with adaptive state injections that can act as the nonlinear gadget of the protocol."
- **Subspace-restricted classical simulation toolkit for photonic QML** - an open-source PyTorch library that simulates these architectures efficiently by computing in the relevant fixed-particle-number subspaces, supplying a reusable classical simulation capability for photonic QML circuits.
  - guarantee: none (an efficiency-oriented simulation tool; no certified accuracy or complexity bound is claimed for it).
  - quote: "This Pytorch \cite{paszke2019pytorchimperativestylehighperformance} based toolkit could be of independent use for photonic simulation and can be found in \cite{PhotonicSubspaceQMLToolkit}}."

## Where it could apply
- Target primitive(s): sampling and variational parameterised-circuit routines (linear-optical / boson-sampling-style evaluation, measurement-based pooling); classical simulation of fixed-photon-number circuits.
- Target application group(s): quantum_ml; platforms_tooling.
- Code family / hardware assumptions: photonic (discrete-variable linear optics, single-photon sources, integrated reconfigurable interferometers, threshold/SNSPD detection, adaptive feedforward); no quantum error correction code family assumed.

## Caveats
This is primarily a new-architecture and proof-of-concept demonstration paper, not a supplier of accreditation, decoding, error correction, or error mitigation primitives. The headline speed-up is a resource-counting claim and the barren-plateau benefit is argued by analogy to Hamming-weight-preserving circuits rather than proven here; none of the three techniques carries a formal statistical, soundness, or static-proof guarantee. State injection is introduced in cited prior work (monbroussou2024quantum) and the subspace-preserving QCNN in monbroussou2024subspace, so this paper supplies their photonic instantiation, integration, and a simulation toolkit rather than the underlying constructions. In the experiment the adaptive state injection is emulated via post-selection, so genuine feedforward is not demonstrated. The toolkit quote is copied verbatim and retains a stray closing brace from the source markup. Full LaTeX source was available; no paywall or partial access.