---
title: Quantum Optical Neural Networks
url: https://www.emergentmind.com/topics/quantum-optical-neural-networks-qonns
type: topic
---

# Quantum Optical Neural Networks

Searching arXiv for recent and foundational papers on quantum optical neural networks to ground the article.
arxiv_search(query="quantum optical neural network", max_results=10, sort_by="relevance")
Quantum Optical Neural Networks (QONNs) are photonic neural-network architectures in which trainable optical transformations implement the linear stages of a network and genuinely optical or measurement-induced mechanisms supply the nonlinearity. In the foundational formulation, a QONN acts on optical modes by alternating parameterized linear interferometers with site-wise nonlinearities, so that a depth-$N$ network has the layered form $S(\vec\Theta)=\prod_{i=1}^N[\Sigma(\phi)\,U(\vec\theta_i)]$ [1808.10047]. Subsequent work broadened the term to include discrete-variable and continuous-variable photonic circuits, all-optical feed-forward systems, Hong–Ou–Mandel and Mach–Zehnder interference neurons, bosonic reservoirs, and hardware-inspired architectures with programmable nonlinearities, photon subtraction, quantum emitters, or atom–cavity dynamics [2409.02533], [2410.17702].

## 1. Definition, scope, and terminological variants

The core idea of a QONN is a mapping between neural-network primitives and optical quantum hardware. Linear weights are realized by interferometric meshes built from beam splitters and phase shifters, while activation is supplied by a non-Gaussian gate, a material nonlinearity, a detection event, or an interference-derived nonlinear observable. In the original 2018 proposal, arbitrary $m\times m$ unitaries are decomposed into arrays of beam splitters and phase shifters, inputs are Fock states across modes, and outputs are measured by single-photon detectors that resolve photon number in each mode [1808.10047].

The optical substrate admits multiple information encodings. Discrete-variable encodings include single-photon Fock states and dual-rail qubits, while continuous-variable encodings include coherent states, squeezed states, and more general Gaussian transformations with symplectic constraints of the form $U U^\dagger-V V^\dagger=I$ and $U V^T-V U^T=0$ [2409.02533], [2410.17702]. This heterogeneity has made QONNs a family of architectures rather than a single canonical model.

The acronym is also overloaded. In one line of work, “QONN” denotes “Quantum Orthogonal Neural Network,” where an orthogonal linear map $W\in O(n)$ is implemented by a parametrized network of two-mode Reconfigurable Beam Splitter gates arranged in a triangular “pyramid” pattern [2411.13520]. In the present usage, however, QONN refers primarily to *quantum optical neural networks* in the photonic sense introduced in the earlier quantum-optical literature [1808.10047].

## 2. Optical primitives and sources of nonlinearity

Across the literature, QONNs are assembled from a relatively stable set of photonic primitives: beam splitters, phase shifters, Mach–Zehnder interferometers, squeezers, detectors, and mode-selective state preparation. What differentiates architectures is the mechanism used to obtain a nonlinear response, since purely Gaussian or purely linear-optical networks do not reproduce the role of an activation function [2409.02533].

| Mechanism | Representative formulation | Representative use |
|---|---|---|
| Kerr or Kerr-like nonlinearity | $U_{\rm Kerr}(\kappa)=\exp[i\kappa\,a^{\dagger2}a^2]$ or $\hat{\rm NS}_m(\chi)=\exp[i\frac{\chi}{2}(\hat a_m^\dagger)^2\hat a_m^2]$ | Layered variational QONNs |
| EIT activation | $I_p^{\rm out}=\sigma(I_c)\,I_p^{\rm in}$ in cold $^{85}$Rb | All-optical state tomography |
| Measurement-induced nonlinearity | Conditional transformation $_{\rm anc}\langle m|U(|\psi_{\rm in}\rangle\otimes|{\rm resource}\rangle_{\rm anc})$ | Optical perceptrons, non-Gaussian gates |
| HOM interference | $p_{\rm coinc}=\frac12[1-\sum_i w_i|\langle I|W_i\rangle|^2]$ or $P_{\rm coinc}=\frac12(1-|\langle W_\lambda|X\rangle|^2)$ | Quantum optical neurons and shallow networks |
| Photon subtraction | $\rho\propto A_K\,\rho_G\,A_K^\dagger$ with $A_K=\prod_{k\in K}a_k$ | CV QONNs with adaptive activations |
| Atom–cavity activation | $a_i^l(z)=\frac{g|z_i^l|}{\Omega_i^l}|\sin(\Omega_i^l t_l)|$ | Fully optical classification networks |
| Quantum-emitter saturation | $f_i(z)=t(|z_i|^2)\,z_i$ | All-optical deep learning |

The 2024 architecture with programmable nonlinearities replaces the standard alternation of universal interferometers and fixed nonlinearities by meshes of two-mode nonlinear Mach–Zehnder interferometers programmable through adjustable Kerr-like elements. Its multimode variational unitary is $\hat U_{\rm core}(\{\chi\})=\prod_{\ell=1}^L\prod_{j\in\Omega_\ell}\hat U_{\rm NMZI}^{(j)}(\chi_{2j-1}^{(\ell)},\chi_{2j}^{(\ell)})$, and the full network adds input and output linear Mach–Zehnder layers for dual-rail compatibility [2410.07868].

A distinct all-optical route uses electromagnetically induced transparency. In the integrated all-optical neural network for quantum state tomography, the hidden layer consists of $20$ spatially separated EIT “neurons” in a cold Rb vapor cell, and the transmitted probe intensity implements a smooth, convex activation $\sigma(\cdot)$ [2103.06457].

Another distinct route avoids material nonlinear optics at inference time by exploiting multiphoton interference. In HOM-based quantum optical neurons, the probability of coincidence after a balanced beam splitter depends on the squared overlap of an input photon state and a learned photon state; the square modulus of an inner product then functions as the effective nonlinearity [2507.21036], [2603.28879].

## 3. Encodings, forward models, and training procedures

QONNs differ substantially in how classical or quantum data are embedded into optical modes. The original layered QONN uses Fock inputs and dual-rail qubits, where $\lvert0\rangle\equiv\lvert10\rangle$ and $\lvert1\rangle\equiv\lvert01\rangle$ [1808.10047]. Mehta and Roy’s quantum-optical neuron instead phase-encodes classical input and weight vectors into single-photon states,
$$
|\psi_x\rangle=\frac1{\sqrt N}\sum_{k=0}^{N-1}e^{i\theta_k}|1\rangle_k,\qquad
|\psi_w\rangle=\frac1{\sqrt N}\sum_{k=0}^{N-1}e^{i\phi_k}|1\rangle_k,
$$
so that $|\langle\psi_w|\psi_x\rangle|^2$ becomes the activation [2507.17349].

In all-optical tomography, the network input is a polarization qubit prepared as $|\psi\rangle=(|H\rangle+e^{i\theta}|V\rangle)/\sqrt2$. The measured Pauli expectations are transformed to nonnegative inputs
$$
x_1=1-\langle X\rangle,\quad x_2=1-\langle Y\rangle,\quad x_3=1-\langle Z\rangle,
$$
and the trained model computes
$$
f_{\rm AONN}(x)=W_2\cdot\sigma(W_1x+b_1)+b_2,
$$
with a $3\times20$ first linear layer, $20$ EIT activations, and a $20\times1$ second linear layer [2103.06457].

Spatial encoding is prominent in quantum optical image-processing neurons. In the shallow-network proposal based on HOM interference, a feature vector $x=(x_0,\dots,x_{N-1})$ is encoded in the transverse spatial profile of a single photon, while hidden-layer parameters are encoded in a mixture $\rho_{\rm hidden}=\sum_{i=0}^{M-1}w_i|W_i\rangle\langle W_i|$. The network output is
$$
f(x;\theta)=\sum_{i=0}^{M-1}w_i\,|\langle I|W_i\rangle|^2,
$$
and the measured coincidence probability is $p_{\rm coinc}=(1-f)/2$ [2507.21036]. The 2026 experimental quantum optical neuron adopts the same principle for image templates $X$ and $W_\lambda$, with $P_{\rm coinc}=\frac12(1-|\langle W_\lambda|X\rangle|^2)$ and a sigmoid readout $f_\theta(X)=\sigma(z(X)+b)$ [2603.28879].

Training procedures are correspondingly diverse. Fidelity-based objectives dominate quantum-state transformation tasks; the 2018 QONN defines
$$
C(\vec\Theta)=1-\frac1K\sum_{i=1}^K\big|\langle\psi_{\rm out}^i|S(\vec\Theta)|\psi_{\rm in}^i\rangle\big|^2
$$
and uses derivative-free optimization in simulation or in-situ parameter updates from measured fidelities [1808.10047]. Classification models use cross-entropy and standard optimizers. The all-optical tomography experiment uses mean-squared error with Adam at learning rate $\eta=2\times10^{-3}$ and reports that $10\,000$ iterations sufficed to converge [2103.06457]. The single-photon-detector optical neural network trains a Bernoulli activation model with a “physics-aware stochastic training” rule based on $P_{\rm SPD}(a=1\mid\mu)=1-e^{-\mu}$ [2307.15712]. Hybrid photonic models also invoke finite differences, parameter-shift rules, automatic differentiation, or explicit PyTorch autograd for differentiable interference modules [2507.17349], [2509.01784], [2601.01690].

## 4. Representative implementations and demonstrated tasks

QONNs have been studied in simulation, in hardware-aware models, and in laboratory demonstrations. The tasks span quantum information processing, classification, attention mechanisms, reinforcement learning, and reservoir computing.

| Task | Architecture | Reported result |
|---|---|---|
| Single-qubit photonic state tomography | All-optical AONN with SLMs, lenses, and EIT | RMSD of $\theta_{\rm pred}$ vs. $\theta_{\rm true}$ $\lesssim0.05\,{\rm rad}$; fidelity stays above $99\%$ [2103.06457] |
| Black-box Hamiltonian simulation | Layered QONN with Kerr nonlinearity | Seven layers achieved $\sim0.1\%$ test error for Bose–Hubbard simulation at $U/t_{\rm hop}=20$ [1808.10047] |
| Quantum optical autoencoder | Structured QONN encoder/decoder | Local- and global-structured strategies converged to $\sim92\%$ reference fidelity [1808.10047] |
| Quantum reinforcement learning | Depth-$6$ QONN policy for cart-pole | Fitness increased from random ($\sim20$) to $\sim60$ steps over $1000$ generations [1808.10047] |
| MNIST optical classification in the single-photon regime | SPDNN with hidden layer at $\sim1$ photon per neuron excitation | Test accuracy $98.0\pm1.3\%$ [2307.15712] |
| QONN-based attention for jet classification | QViT with quantum orthogonal layers | Test accuracy $0.6755$, AUC $0.7369$ [2411.13520] |
| MNIST and Fashion-MNIST quantum optical neurons | HOM/MZ differentiable neurons | Multiclass MNIST: classical net and HOM-amplitude both reach $\approx98.2\%$ test accuracy [2509.01784] |
| Camera-free image classification | Experimental single QON and two-neuron QOSN | Test accuracy $100\%$ on MNIST “0 vs. 1”; $95\%$ on Fashion-MNIST for the two-neuron QOSN [2603.28879] |

These demonstrations reveal two distinct but connected trajectories. One trajectory treats QONNs as quantum processors for state preparation, tomography, simulation, or variational tasks [1808.10047], [2103.06457], [2410.07868]. The other treats them as photonic classifiers or feature extractors that exploit interference, few-photon detection, or optical saturation to realize neural primitives with low optical energy and small hardware footprints [2307.15712], [2507.21036], [2603.28879], [2601.01690].

The application space has also widened. A QONN-derived “quantum optical convolutional neural network” augments a quantum-optical core with convolution and pooling for MNIST, reporting $99.01\%$ test accuracy and average ROC-AUC $0.998$ [2012.10812]. Atom–cavity QONNs report test accuracy $\approx95\%$ on MNIST and $\approx95\%$ on SAT-6 with a convolutional front end [2511.06167]. A waveguide-QED architecture based on coherent transient dynamics reports $97.60\%$ on MNIST and $92.32\%$ on a nine-colored-object dataset [2605.17752].

## 5. Trainability, parameter scaling, and simulation frameworks

A central theoretical question is whether photonic variational networks inherit barren plateau pathologies. For linear-optical continuous-variable modules, the answer is conditional. Coherent light in $m$ modes can be “generically compiled efficiently if the total intensity scales sublinearly with $m$,” and the same conclusion extends to homodyne, heterodyne, photon-counting, and attenuated settings. Specifically, barren plateaus arise when $E=O(m)$, whereas no barren plateau appears when $E=o(m)$ but not exponentially small in $m$ [2008.09173].

Parameter scaling has become another axis of comparison. In the programmable-nonlinearity QONN, each new nonlinear layer adds only $M$ adjustable parameters, giving $P_{\rm NL}\approx D\times M$, whereas a purely linear-optics QONN requires $P_{\rm L}\approx(D+1)M(M-1)\sim D\,M^2$ [2410.07868]. The paper reports concrete reductions such as $P_{\rm NL}=62\ll P_{\rm L}=168$ for $4$-qubit GHZ generation and $P_{\rm NL}=2$ versus $P_{\rm L}=24$ for a deterministic $4$-state Bell analyzer at depth $D=1$ [2410.07868]. This suggests a design principle in which programmable nonlinear elements compensate for reduced global interferometric freedom.

Because bosonic Hilbert spaces grow rapidly, simulation methods have diversified. Large bosonic reservoirs have been studied with the positive-P representation, which yields stochastic equations for doubled complex variables and avoids density-matrix truncation. In the reservoir framework, one trajectory costs $O(N\times T/\Delta t)$ and $S$ trajectories cost $O(NST)$, while sampling error scales as $1/\sqrt S$ [2507.07684]. The same work reports a non-monotonic dependence of performance on reservoir size: in quantum state classification at $U/\gamma=0.02$, the test accuracy rises to a maximum $\sim0.77$ at $N\approx7$ and then declines toward $0.33$ for large $N$ [2507.07684].

For hardware-inspired continuous-variable QONNs with Gaussian maps and photon subtraction, exact classical simulation has been developed through the QuaNNTO library using Bogoliubov commuting rules and Wick–Isserlis expansion, “without truncating the infinite-dimensional Hilbert space” [2512.05204]. That framework also derives closed-form adaptive activations and argues that a single layer with sufficiently many subtraction channels satisfies the Universal Approximation Theorem [2512.05204].

## 6. Limitations, misconceptions, and research directions

A persistent misconception is that QONNs are a single mature hardware platform. The literature instead spans idealized Kerr-layer circuits, all-optical feed-forward systems, bosonic reservoirs, HOM-based neurons, phase-space simulators, quantum-emitter activations, and differentiable quantum-inspired modules [1808.10047], [2507.07684], [2509.01784]. The common denominator is optical implementation of weighted mode mixing plus a nonlinear optical or measurement layer, not a unique circuit topology.

The principal bottleneck remains nonlinearity. Reviews emphasize that deterministic, high-fidelity optical Kerr or cubic phase gates remain beyond current bulk nonlinear materials, and that measurement-induced non-Gaussian elements are probabilistic and therefore introduce overhead [2409.02533]. Experimental work identifies additional constraints: partial temporal or spectral distinguishability reduces HOM visibility; losses and detector inefficiency lower count rates; dark counts and multi-photon events add noise; large interferometer meshes demand calibration and phase stability; and integrating thousands of cavity-QED sites with uniform coupling and precise detuning control remains challenging [2507.21036], [2511.06167], [2603.28879].

At the same time, several directions recur across the literature. One is tighter integration: room-temperature vapor cells or integrated nonlinear waveguides for all-optical tomography, on-chip photonic implementations for quantum-enhanced attention, inverse-designed nanophotonic activations based on saturable quantum emitters, and integrated photonic platforms with on-chip squeezers, modulators, and homodyne arrays [2103.06457], [2411.13520], [2410.17702], [2601.01690]. Another is hybridization: optical forward passes combined with classical gradient updates, FPGA or thermo-optic controllers, or hardware-in-the-loop training [2507.17349], [2509.01784], [2605.17752].

A further implication is that “QONN” research now occupies two scales simultaneously. At the quantum-information scale, QONNs are variational photonic circuits for state preparation, tomography, simulation, and entanglement-sensitive processing [1808.10047], [2103.06457], [2410.07868]. At the photonic-machine-learning scale, they are neuromorphic or quantum-inspired optical substrates whose primitive operation is an interferometric overlap, a few-photon detection event, or a saturable light–matter response [2307.15712], [2603.28879], [2601.01690]. The convergence of these scales suggests that future QONN work will continue to be defined less by a single formal model than by a shared program: realizing neural computation directly in quantum-optical hardware, with non-Gaussianity as the decisive resource.

Source: https://www.emergentmind.com/topics/quantum-optical-neural-networks-qonns