Quantum Optical Neurons (QONs)

• Quantum Optical Neurons (QONs) are photonic neuromorphic elements that leverage quantum states and optical interference to replicate neural computation.
• They integrate mechanisms such as Kerr nonlinearity, photon subtraction, and light–matter interactions to implement weighted summation and nonlinear activations.
• Their architectures, from layered circuits to interferometric and continuous-variable designs, offer scalable and energy-efficient routes for quantum-enhanced machine learning.

Quantum optical neurons (QONs) are neuron-like computational primitives realized in photonic or quantum-optical hardware, with information encoded in quantum states of light and with weighted summation, nonlinearity, memory, or readout implemented by optical interference, non-Gaussian operations, measurement, or light–matter interaction. The term does not denote a single canonical device. In the literature it spans layered quantum optical neural networks (QONNs) built from beam splitters, phase shifters, and Kerr elements (Steinbrecher et al., 2018); continuous-variable architectures in which Gaussian transformations are followed by photon subtraction and homodyne detection (Krasimirov-Ivanov et al., 4 Dec 2025); interferometric neurons in which Hong–Ou–Mandel (HOM) or Mach–Zehnder (MZ) statistics directly encode overlaps between an input and a learned template (Minati et al., 30 Mar 2026); quantum-noise-limited stochastic optical neurons operated at the single-photon level (Ma et al., 2023); and all-optical activations based on atom–cavity systems, quantum emitters, waveguide QED, or excitable quantum-dot lasers (Zhu et al., 9 Nov 2025).

1. Terminological scope and defining features

In the discrete-variable circuit tradition, a QON is the local nonlinear optical element in a layered QONN: linear transformations are implemented optically, while the “activation” is a photon-number-dependent quantum operation, often Kerr-like (Steinbrecher et al., 2018). In continuous-variable formulations, the neuron analogue emerges from the combination of a Gaussian affine map and a non-Gaussian resource such as photon subtraction, with homodyne expectation values playing the role of outputs (Krasimirov-Ivanov et al., 4 Dec 2025). In interferometric proposals, the neuron is identified with a measurement primitive: the network pre-activation is extracted from coincidence or output-port probabilities determined by the overlap of two photonic states (Andrisani et al., 1 Sep 2025). In analogue and recurrent settings, the relevant unit is less a sitewise activation gate than a dynamical optical subsystem whose coherent, dissipative, or squeezed evolution provides state transition, memory, and nonlinear response (Labay-Mora et al., 2024).

Family	Physical substrate	Representative mechanism
Layered QONN	Fock-state photonics	Linear optics + Kerr nonlinearity
CV QON	Coherent/squeezed modes	Gaussian map + photon subtraction
Interferometric QON	Single-photon spatial modes	HOM or MZ overlap measurement
Stochastic QON	Few-photon optical ONN	SPD click statistics
All-optical QON	Atom–cavity, TLS, QD laser, waveguide QED	Saturability, transient Rabi, excitable dynamics

A recurrent misconception is that “quantum optical neuron” refers only to one of these families. The recent review literature instead treats QONNs as a heterogeneous class unified by the attempt to map the weighted-sum/nonlinearity pattern of neural computation into optics, and by the central role of non-Gaussian operations in achieving genuine nonlinear or quantum-advantaged behavior (Yu et al., 2024).

2. Computational primitives and mathematical structure

Across architectures, the optical analogue of the linear layer is well established. Inputs may be encoded as Fock states, dual-rail qubits, general $n$-photon $m$-mode Fock states, coherent states, squeezed states, or single-photon spatial states. The corresponding linear transformations are implemented by interferometers, beam splitters, phase shifters, Gaussian unitaries, or multiport devices (Steinbrecher et al., 2018). In the original layered QONN construction, a network with $N$ layers is written as

$$S(\vec{\Theta}) = \prod_{i=1}^{N} \Sigma(\phi)\cdot U(\vec{\theta}_i),$$

where $U(\vec{\theta}_i)$ is an $m$-mode unitary realized by linear optics and $\Sigma(\phi)$ is a single-mode Kerr activation (Steinbrecher et al., 2018). The latter is

$$\Sigma(\phi)=\sum_{n=1}^{\infty}|0\rangle\langle 0|+e^{i(n-1)\phi}|n\rangle\langle n|.$$

The central technical difficulty is the nonlinear stage. Direct Kerr interactions are one route, but the review literature emphasizes that non-Gaussian or measurement-induced operations are indispensable because all-Gaussian optical circuits are efficiently classically simulatable and lack computational universality (Yu et al., 2024). This has produced several distinct neuron models. In single-photon-detector neurons, the activation is the Bernoulli click statistic

$P_{\text{SPD}}(\lambda)=1-e^{-\lambda},$

with the output equal to $1$ with probability $m$0 and $m$1 otherwise (Ma et al., 2023). In continuous-variable photon-subtraction architectures, a single-mode quantum-optical activation is derived in closed form as

$m$2

which combines an affine term with a non-polynomial nonlinear contribution (Krasimirov-Ivanov et al., 4 Dec 2025).

The review perspective on quantum states of light identifies squeezing, coherence, entanglement, and quantum measurement as the principal resources. Squeezing can encode information nonlinearly even when the underlying evolution is Gaussian, can suppress readout noise, and can enhance memory in reservoir computing; measurement can itself play the role of activation; and entangling resources enable non-local correlations across modes (Labay-Mora et al., 2024). This suggests that the “neuron” in quantum optics is often better understood as a hardware-level nonlinear map on optical modes than as a direct transcription of a scalar classical activation function.

3. Layered circuit and interferometric realizations

The layered QONN introduced by Steinbrecher and collaborators maps classical neural-network structure into the quantum optical domain through alternating linear-optical layers and local Kerr nonlinearities (Steinbrecher et al., 2018). Inputs are encoded as Fock states, outputs are read by photon-number detectors, and the training objective is

$m$3

Within classical simulation, that architecture was trained for CNOT, Bell-state generation, GHZ-state generation, black-box Hamiltonian simulation, autoencoding of molecular hydrogen states, and reinforcement learning. The reported Bose–Hubbard simulation error of $m$4 with $m$5 layers, the $m$6 fidelity for the compressed H$m$7 reference state, and competitive cartpole performance all belong to this specific simulated QONN setting (Steinbrecher et al., 2018).

Subsequent work replaced the conventional “programmable linear / fixed nonlinear” pattern by meshes of nonlinear Mach–Zehnder interferometers with adjustable Kerr-like nonlinearities (Chernykh et al., 2024). There the variational resource is the nonlinearity strength $m$8, with each nonlinear sign-shift gate given by

$m$9

The main architectural claim is a reduction in adjustable parameters from $N$0 to $N$1 for $N$2 modes, with examples including a drop from $N$3 to $N$4 for $N$5-/$N$6-qubit maximally entangled states, Bell-state discrimination with one nonlinear layer and two tunable parameters instead of $N$7, and a $N$8-qubit Heisenberg VQE instance using $N$9 parameters instead of $$S(\vec{\Theta}) = \prod_{i=1}^{N} \Sigma(\phi)\cdot U(\vec{\theta}_i),$$0 (Chernykh et al., 2024).

A distinct line of work identifies the neuron directly with optical overlap estimation. In the quantum optical shallow-network protocol, both inputs and parameters are encoded into single-photon states, and the neuron output is extracted from HOM coincidences (Roncallo et al., 28 Jul 2025). For a hidden-state mixture,

$$S(\vec{\Theta}) = \prod_{i=1}^{N} \Sigma(\phi)\cdot U(\vec{\theta}_i),$$1

while the coincidence probability is

$$S(\vec{\Theta}) = \prod_{i=1}^{N} \Sigma(\phi)\cdot U(\vec{\theta}_i),$$2

That protocol is unusual in claiming that, once trained, inference requires constant optical resources regardless of the number of input features and neurons (Roncallo et al., 28 Jul 2025). A related “quantum optical model of an artificial neuron” encodes input and weight vectors as phase factors of single-photon multimode states and computes

$$S(\vec{\Theta}) = \prod_{i=1}^{N} \Sigma(\phi)\cdot U(\vec{\theta}_i),$$3

from detection at one output mode, with simulated reductions in circuit depth and width relative to the qubit-based version (Mehta et al., 23 Jul 2025).

Experimental interferometric image classifiers make this definition literal. In a camera-free classifier based on HOM interference, coincidence visibility

$$S(\vec{\Theta}) = \prod_{i=1}^{N} \Sigma(\phi)\cdot U(\vec{\theta}_i),$$4

directly reports the squared overlap between an input image mode and a learned probe mode (Minati et al., 30 Mar 2026). That experiment realized both a single-perceptron QON and a two-neuron shallow network, reported $$S(\vec{\Theta}) = \prod_{i=1}^{N} \Sigma(\phi)\cdot U(\vec{\theta}_i),$$5 test accuracy on the sampled binary MNIST task and $$S(\vec{\Theta}) = \prod_{i=1}^{N} \Sigma(\phi)\cdot U(\vec{\theta}_i),$$6 test accuracy for the two-neuron Fashion-MNIST task, and observed that performance remained insensitive to input resolution at fixed measurement budget (Minati et al., 30 Mar 2026). Software benchmarking of HOM- and MZ-based QONs further showed that HOM-based amplitude modulation and MZ-based phase-shifted modulation can be comparable to classical neurons, whereas intensity-based encodings are more sensitive to distributional shifts and training instabilities (Andrisani et al., 1 Sep 2025).

4. Continuous-variable, reservoir, and associative-memory QONs

Continuous-variable quantum optics provides a more explicitly hardware-oriented analogue of the classical affine-plus-activation decomposition. In the CV-QONN framework, real-valued data are encoded as coherent states, each layer applies a general Gaussian transformation, nonlinearity is provided by multi-mode photon subtraction, and outputs are read by homodyne detection (Krasimirov-Ivanov et al., 4 Dec 2025). The affine map is written as

$$S(\vec{\Theta}) = \prod_{i=1}^{N} \Sigma(\phi)\cdot U(\vec{\theta}_i),$$7

with $$S(\vec{\Theta}) = \prod_{i=1}^{N} \Sigma(\phi)\cdot U(\vec{\theta}_i),$$8 the Bogoliubov matrix and $$S(\vec{\Theta}) = \prod_{i=1}^{N} \Sigma(\phi)\cdot U(\vec{\theta}_i),$$9 the displacement, while photon-subtracted modes define the effective quantum-optical neurons. The paper states that the architecture satisfies the Universal Approximation Theorem within a single layer, derives closed-form adaptive activations, and develops the QuaNNTO simulator to compute exact expectation values of non-Gaussian states without truncating the infinite-dimensional Hilbert space (Krasimirov-Ivanov et al., 4 Dec 2025).

The same review tradition frames quantum optical neurons more generally as analogue, recurrent, or reservoir subsystems rather than feedforward activation sites (Labay-Mora et al., 2024). In quantum reservoir computing, the optical reservoir is a fixed complex quantum system and only the readout is trained; squeezing phase can encode the classical input, homodyne moments such as $U(\vec{\theta}_i)$0 and $U(\vec{\theta}_i)$1 provide observables, and squeezing is reported to enhance expressivity, memory, and robustness to noise (Labay-Mora et al., 2024). Quantum associative memories are described through driven-dissipative nonlinear oscillators with multiple attractors in phase space, where coherent, amplitude-squeezed, or phase-squeezed states can serve as stored patterns (Labay-Mora et al., 2024).

Large-scale simulation of such bosonic neuromorphic systems is itself a major research problem. A phase-space positive-$U(\vec{\theta}_i)$2 framework represents the density matrix as a positive distribution over doubled phase-space variables and evolves stochastic differential equations rather than an exponentially large Hilbert-space object (Świerczewski et al., 10 Jul 2025). Applied to multimode quantum reservoirs, that approach shows that performance does not improve monotonically with the number of bosonic modes, but instead depends on the interplay of nonlinearity, reservoir size, and the average occupation of the input mode (Świerczewski et al., 10 Jul 2025). This is an important corrective to the common intuition that larger optical reservoirs are automatically better.

5. All-optical neurons from light–matter interaction

Several recent QON proposals shift the nonlinear primitive from interferometry to microscopic light–matter dynamics. In the atom–cavity QONN, each neuron is a two-level atom placed inside low-$U(\vec{\theta}_i)$3 and high-$U(\vec{\theta}_i)$4 cavities, with absorption, excitation transfer, and spontaneous emission furnishing the activation cycle (Zhu et al., 9 Nov 2025). After phase-locking, the activation is

$U(\vec{\theta}_i)$5

This construction is explicitly designed to avoid electronic nonlinearities in hidden layers. Simulations reported test accuracy $U(\vec{\theta}_i)$6 on MNIST at intermediate absorption duration $U(\vec{\theta}_i)$7, robustness to random detuning up to $U(\vec{\theta}_i)$8, about $U(\vec{\theta}_i)$9 MNIST accuracy at photon pass rate $m$0, and $m$1 accuracy on SAT-6 with a two-layer convolutional QONN (Zhu et al., 9 Nov 2025).

In waveguide-QED architectures, the neural primitives are decomposed into three quantum-optical processes: programmable synaptic weights from phase-tunable nonlocal interference in a giant cavity, temporal summation by a critical-gain bad-cavity integrator, and nonlinear activation from transient Rabi dynamics of a two-level system (Cao et al., 18 May 2026). The output field of the activation stage obeys

$m$2

with the underlying Langevin equations producing a sigmoidal transient response. Full-physics simulations yielded $m$3 accuracy on MNIST and emphasized elimination of the optoelectronic activation bottleneck (Cao et al., 18 May 2026).

An even more strongly nonlinear route embeds quantum emitters in inverse-designed nanophotonic structures. There the activation is the saturable response of a two-level system,

$m$4

operating below $m$5 intensities (Zhou et al., 4 Jan 2026). The paper introduces a “growth factor” $m$6 linking nonlinearity to expressive power, reports $m$7 for the quantum-emitter QON versus $m$8 for silicon Kerr and $m$9 for graphene saturable absorption, and gives the scaling estimate $\Sigma(\phi)$0 for LLM inference, with $\Sigma(\phi)$1 even for trillion-parameter models (Zhou et al., 4 Jan 2026). These are system-level estimates, not a large-scale device demonstration.

Excitable dual-state quantum-dot lasers supply a different notion of photonic neuron: an all-or-none spiking element rather than an activation map on amplitudes (Dillane et al., 2022). The experimentally measured absolute and relative refractory times, $\Sigma(\phi)$2-$\Sigma(\phi)$3 ns and $\Sigma(\phi)$4-$\Sigma(\phi)$5 ns respectively, place these devices in the GHz regime for ultrafast analog computing (Dillane et al., 2022). A plausible implication is that some QON research is converging with photonic neuromorphic spiking hardware rather than only with variational quantum circuits.

6. Training, benchmarking, and unresolved issues

Training methodology is strongly architecture-dependent. Layered circuit QONNs have been optimized both in silico and in situ through task-specific cost functions based on output-state fidelity or measurement outcomes (Steinbrecher et al., 2018). Interferometric shallow networks use binary cross-entropy with gradient-based optimization under normalization and positivity constraints (Roncallo et al., 28 Jul 2025). The atom–cavity QONN was implemented in PyTorch and trained with stochastic gradient descent using an analytically available derivative of the activation function (Zhu et al., 9 Nov 2025). CV-QONNs rely on exact expectation-value engines such as QuaNNTO for differentiable simulation of non-Gaussian layers (Krasimirov-Ivanov et al., 4 Dec 2025).

At the quantum-noise floor, training itself must incorporate the physics of the neuron. In the stochastic single-photon-detector architecture, the hidden layer operates at signal-to-noise ratio $\Sigma(\phi)$6, the forward pass samples Bernoulli activations from the physical click statistics, and the backward pass uses the expectation value in a mean-field approximation (Ma et al., 2023). That experiment demonstrated $\Sigma(\phi)$7 MNIST test accuracy with a hidden layer in the single-photon regime, using $\Sigma(\phi)$8 photons per multiply-accumulate operation and $\Sigma(\phi)$9 attojoules per MAC in the hidden layer (Ma et al., 2023). This result is notable because the neuron is explicitly stochastic and binary, rather than a deterministic optical transfer function.

Two broad limitations recur across the literature. The first is the nonlinearity bottleneck. Reviews emphasize that scalable quantum-optical neural computation requires non-Gaussian operations, yet deterministic strong optical nonlinearities remain difficult; this motivates measurement-induced gates, photon subtraction, saturability, and transient quantum dynamics as alternative neuron mechanisms (Yu et al., 2024). The second is that headline resource claims are highly model-specific. Constant optical resources independent of input size are proven only for specific HOM-based shallow protocols (Roncallo et al., 28 Jul 2025), and resolution-insensitive performance at fixed measurement budget is demonstrated only for a particular camera-free overlap classifier (Minati et al., 30 Mar 2026). Conversely, large bosonic reservoirs do not exhibit monotonic scaling with mode count (Świerczewski et al., 10 Jul 2025), and some QON variants display optimization instability under distributional shift (Andrisani et al., 1 Sep 2025).

The field therefore does not yet support a single settled answer to what the “best” quantum optical neuron is. Instead, it offers a rapidly expanding taxonomy of physical primitives: Kerr activations for circuit QONNs, photon-subtraction neurons for continuous variables, HOM/MZ overlap neurons for direct inference, SPD neurons for few-quanta stochastic computing, and light–matter nonlinearities for all-optical hidden layers. What unifies them is not a common device geometry, but the attempt to make optical physics itself perform the neural operation.