Photonic Quantum Convolutional Neural Networks with Adaptive State Injection

Léo Monbroussou, Beatrice Polacchi, Verena Yacoub, Eugenio Caruccio, Giovanni Rodari, Francesco Hoch, Gonzalo Carvacho, Nicolò Spagnolo, Taira Giordani, Mattia Bossi, Abhiram Rajan, Niki Di Giano, Riccardo Albiero, Francesco Ceccarelli, Roberto Osellame, Elham Kashefi, Fabio Sciarrino · 2025

Read Photonic Quantum Convolutional Neural Networks with Adaptive State Injection by Léo Monbroussou, Beatrice Polacchi, Verena Yacoub, Eugenio Caruccio, Giovanni Rodari, Francesco Hoch, Gonzalo Carvacho, Nicolò Spagnolo, Taira Giordani, Mattia Bossi, Abhiram Rajan, Niki Di Giano, Riccardo Albiero, Francesco Ceccarelli, Roberto Osellame, Elham Kashefi, Fabio Sciarrino (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-20989.txt.

Introduction

Monbroussou and a large team build and run in twenty twenty-five the first light-based quantum neural network of the convolutional kind with adaptive state injection. Quantum circuits made of light keep a fixed particle count, a quirk lately used to design leaner networks. Adding injection, swapping fresh states in mid-circuit, gives extra control without breaking that thrift.

The chip used a semiconductor single-photon source and programmable interferometers of two sizes. It classified simple images in a binary task on real hardware, matching the design's promise. Simulations on larger datasets probed how the design scales. The authors also show a simple adaptive trick for a harder sampling task, aimed at near-term quantum devices.

Catalog record

CREATORLéo Monbroussou, Beatrice Polacchi, Verena Yacoub, Eugenio Caruccio, Giovanni Rodari, Francesco Hoch, Gonzalo Carvacho, Nicolò Spagnolo, Taira Giordani, Mattia Bossi, Abhiram Rajan, Niki Di Giano, Riccardo Albiero, Francesco Ceccarelli, Roberto Osellame, Elham Kashefi, Fabio Sciarrino
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE77 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-20989

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Photonic Quantum Convolutional Neural Networks with Adaptive State Injection Léo Monbroussou, Beatrice Polacchi, Verena Yacoub, Eugenio Caruccio, Giovanni Rodari, Francesco Hoch, Gonzalo Carvacho, Nicolò Spagnolo, Taira Giordani, Mattia Bossi, Abhiram Rajan, Niki Di Giano, Riccardo Albiero, Francesco Ceccarelli, Roberto Osellame, Elham Kashefi, Fabio Sciarrino Abstract Linear optical architectures have been extensively investigated for quantum computing and quantum machine learning applications. Recently, proposals for photonic quantum machine learning have combined linear optics with resource adaptivity, such as adaptive circuit reconfiguration, which promises to enhance expressivity and improve algorithm performances and scalability. Moreover, linear optical platforms preserve some subspaces due to the fixed number of particles during the computation, a property recently exploited to design a novel quantum convolutional neural networks. This last architecture has shown an advantage in terms of running time complexity and of the number of parameters needed with respect to other quantum neural network proposals. In this work, we design and experimentally implement the first photonic quantum convolutional neural network (PQCNN) architecture based on particle-number preserving circuits equipped with state injection, an approach recently proposed to increase the controllability of linear optical circuits. Subsequently, we experimentally validate the PQCNN for a binary image classification on a photonic platform utilizing a semiconductor quantum dot-based single-photon source and programmable integrated photonic interferometers comprising 8 and 12 modes. In order to investigate the scalability of the PQCNN design, we have performed numerical simulations on datasets of different sizes. We highlight the potential utility of a simple adaptive technique for a nonlinear Boson Sampling task, compatible with near-term quantum devices. I Introduction In 2001, E. Knill, R. Laflamme, and G. J. Milburn demonstrated that universal quantum computing can be achieved by using only beam splitters, phase shifters, single-photon sources, and photo-detectors Knill et al. 2001 . However, the resources needed to fully achieve such a task are demanding, and only noisy intermediate-scale quantum processors can be realized with current photonic technologies. As a consequence, this triggered the study of sub-universal platforms such as Boson Sampling Aaronson and Arkhipov 2011 , which aims at showing a quantum computational advantage with the current technology, as demonstrated by the recent realizations of Gaussian Boson Sampling Zhong et al. 2020 ; Zhong et al. 2021 ; Madsen et al. 2022 . However, so far, it has been difficult to map useful problems in the Boson Sampling scheme. Nonetheless, the development of photonic quantum technologies Flamini et al. 2018 ; Alexander et al. 2025 and, in particular, integrated ones Wang et al. 2020 ; Pelucchi et al. 2022 ; Giordani et al. 2023a , encourages the search for algorithms and architectures that allow for a useful intermediate quantum computational advantage while being compatible with the available hardware. Indeed, photonic integrated interferometers allow for a large number of modes in a compact layout and for intrinsic phase stability, thus significantly increasing the capability of implementing large systems with respect to bulk interferometers. A promising route to extend the expressivity of photonic architectures and the range of applications in the field of quantum machine learning Biamonte et al. 2017 ; Schuld et al. 2014 ; Ciliberto et al. 2018 ; Romero et al. 2017 ; Dunjko and Briegel 2018 ; Biamonte 2021 is represented by the introduction of nonlinearity Spagnolo et al. 2023 ; Steinbrecher et al. 2019 or particular adaptive operations Chabaud et al. 2021 ; Hoch et al. 2025 . This is exemplified by the theoretical work Chabaud et al. 2021 that proposes to apply Boson Sampling equipped with adaptive circuit reconfiguration, to tackle quantum machine learning tasks, for instance kernel estimation Hoch et al. 2025 . The so-called Adaptive Boson Sampling scheme Chabaud et al. 2021 , recently demonstrated in Hoch et al. 2025 , proposes to apply unitary transformations according to the measurement outcomes over subsets of photons and modes, which represent the adaptive parts of the protocol. Within this framework, an alternative approach was recently introduced in Monbroussou et al. 2024a , which substitutes adaptive evolutions with adaptive injection into the optical circuits of new photonic states. More precisely, state injection is a measurement-based operation that triggers the preparation and injection of new photons according to the outcomes of single-photon detections in selected channels. The resulting architecture alternates linear optical layers with adaptive state injections that can act as the nonlinear gadget of the protocol. Such an adaptive linear optic scheme enables a more straightforward and scalable mapping of optical neural networks Steinbrecher et al. 2019 to the operations of an adaptive photonic circuit. Various models for quantum learning attempted to expand the class of machine learning algorithms that can be encoded in quantum hardware Schuld et al. 2014 ; Dunjko and Briegel 2018 ; Cong et al. 2019 , for example, by looking at Classical Convolutional Neural Networks (CCNNs), currently widely used for image, audio, or video recognition Lecun et al. 1998 ; Li et al. 2022 ; Alzubaidi et al. 2021 . A quantum counterpart of CCNNs has been recently proposed in Monbroussou et al. 2025a , offering notable advantages in terms of scalability, trainability, and expressivity. In particular, this scheme introduces a Quantum Convolutional Neural Network (QCNN) architecture based on Hamming weight-preserving quantum circuits. 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. In this work, we introduce and experimentally demonstrate an architecture for Photonic Quantum Convolutional Neural Networks (PQCNNs) which explores the state injection architecture presented in Monbroussou et al. 2024a to realize a photonic implementation of subspace preserving QCNNs Monbroussou et al. 2025a . We realize a proof-of-concept experiment by employing a cutting-edge single-photon source based on a semiconductor Quantum Dot (QD) Somaschi et al. 2016 ; Loredo et al. 2019 , a time-to-spatial demultiplexer, and universal programmable 12-mode and 8-mode interferometers realized with the femtosecond laser-writing technique Ceccarelli et al. 2020 ; Corrielli et al. 2021 . The designed PQCNN scheme is tailored to the experimental platform at hand, with the goal of carrying out a binary image classification. As a complement to the experimental investigation, we provide a systematic study on the scaling and complexity of the protocol, by leveraging numerical simulations on larger quantum systems, demonstrating the potential behind the proposed scheme for PQCNNs. Finally, we offer a tailor-made library that allows for efficient simulation of such architecture by performing the computation in the most appropriate subspaces. II Photonic Quantum Convolutional Neural Networks Figure 1: Classical Convolutional Neural Networks (CCNN) and Photonic Quantum Convolutional Neural Networks (PQCNN) based on adaptive state injection. a) CCNNs alternate convolutional and pooling layers to process the data features and reduce their dimensions. The last layer is a dense layer that connects the remaining features to the output nodes. b) PQCNN circuit representation for an input image of size d_{1}\times d_{2} . The entire flow is composed of a Quantum Data Loader (QDL), a convolutional layer with a Quantum Convolutional Filter (QCF) of size K\times K , a pooling layer, and a dense layer before the detection. In the described scenario, the QDL takes as input two separated registers, indicated with m_{i} and n_{i} , that encode, respectively, the rows and the columns of

Read the complete text