Laser interferometry as a robust neuromorphic platform for machine learning

Amanuel Anteneh, Kyungeun Kim, J. M. Schwarz, Israel Klich, Olivier Pfister · 2026

Read Laser interferometry as a robust neuromorphic platform for machine learning by Amanuel Anteneh, Kyungeun Kim, J. M. Schwarz, Israel Klich, Olivier Pfister (2026) 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-2601-18047.txt.

Introduction

Amanuel Anteneh, Kyungeun Kim, J. M. Schwarz, Israel Klich, and Olivier Pfister turn laser interferometry into a learning machine. Their 2026 preprint presents an optical neural network built only from linear tools. The tools are field displacement and interferometry applied to coherent states of light. A neural network needs a nonlinear step to learn, and optics is mostly linear. The team gets the needed nonlinearity from how the input is encoded into phase.

Encoding tricks let plain optics do a smart machine's work. Light-based learning promises speed and low energy use. It may also resist some flaws of electronic hardware. The work adds a robust platform to the race for physical machine learning.

Catalog record

CREATORAmanuel Anteneh, Kyungeun Kim, J. M. Schwarz, Israel Klich, Olivier Pfister
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE44 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-18047

Browse the full library · About the Tower of Babel

Full text

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

Laser interferometry as a robust neuromorphic platform for machine learning Preprint: APS/123-QED Amanuel Anteneh, Kyungeun Kim, J. M. Schwarz, Israel Klich, Olivier Pfister Abstract We present a method for implementing an optical neural network using only linear optical resources, namely field displacement and interferometry applied to coherent states of light. The nonlinearity required for learning in a neural network is realized via an encoding of the input into phase shifts allowing for far more straightforward experimental implementation compared to previous proposals for, and demonstrations of, in situ inference. Beyond in situ inference, the method enables in situ training by utilizing established techniques like parameter shift rules or physical backpropagation to extract gradients directly from measurements of the linear optical circuit. We also investigate the effect of photon losses and find the model to be very resilient to these. I Introduction Modern deep learning models have reached unprecedented sizes with recent large language models having billions of trainable parameters and taking months of time for pretraining on clusters of multiple graphics processing units (GPUs) 13 ; 50 ; 60 . The training and inference phases for these models require increasingly large amounts of energy due to the energy-efficiency limitations imposed by the modern silicon-based digital computers on which they are implemented, a phenomenon referred to as the ‘von Neumann bottleneck’. This bottleneck is a consequence of the separation of computing and memory into distinct units in the current von-Neuman computer architecture which necessitates costly data transfer between these units during computation 40 ; 17 . This challenge has motivated the development of neuromorphic computing, a field that seeks to construct hardware architectures inspired by the remarkable energy efficiency of the human brain 40 ; 17 . Consequently, the intersection of machine learning (ML) and neuromorphic computing has lead to advances in the development of physical neural networks (PNNs) 43 . These neural networks seek to harness the capabilities of various analog systems to improve both the speed and resource efficiency with which these models can be trained and deployed for inference. A promising area being explored for neuromorphic computing, and thus for the implementation of PNNs, is the field of photonics 17 ; 43 . Neuromorphic computing implementations based on photonics have several advantages such as higher energy efficiency, throughput, speed and parallelization capability 17 . However, a limitation of current approaches based on photonics is scalability and difficulty of implementing low power optical nonlinearities 17 . On the scalability front quantum photonics, particularly with continuous variable (CV) cluster states, can provide a remedy as much work has been done on scaling these systems to large sizes for measurement-based quantum computing schemes 15 ; 71 ; 34 ; 4 . The original proposal for a CV quantum neural network 31 , as well as derivative work 6 , utilized optical Kerr nonlinearities which are difficult to implement experimentally. An alternative proposal utilized repeat-until-success photon number resolving measurements on ancillary modes to introduce nonlinearity into the network 8 . However, the implementation of quantum optical nonlinearities is still a challenging task as in the classical optics case. One solution that has been explored in the classical optics approach is to utilize hybrid architectures based on optical-digital conversion 57 ; 62 ; 46 or opto-electronic conversion 56 ; 66 ; 24 ; 5 ; 16 ; 69 ; 7 . Another approach uses the intrinsic nonlinearity within coupled semiconductor nanolaser arrays 29 . However, learning nonlinearity should not be equated with physics nonlinearity. Indeed, recent breakthroughs in optical implementations 41 ; 63 ; 70 ; 68 stemmed from the realization that optical nonlinearities in the field variables are not needed to implement learning 41 if one doesn’t encode the data in them exclusively and, rather, uses the experimental parameters for that purpose: a simple optical phase shift \theta imparts field transformations \cos\theta,\sin\theta that are nonlinear in \theta (though linear in the fields). This has recently been proposed 63 and demonstrated 70 ; 68 using classical linear wave scattering. In this paper, we propose a different linear optics encoding: a multimode laser interferometer with added amplitude and phase displacements, see Fig. 1 . This approach is considerably simpler than the previous ones in terms of experimental implementation, does not utilize digital linear layers for postprocessing, and is also compatible with the current state of integrated photonics 11 . It is interesting—in particular to set the stage for future extensions to the quantum regime—to deliberately adopt a quantum optical description of the optical circuit, even though it sits firmly within the classical border in this paper. Indeed, three classes of optical circuits can be defined: (i) , classical linear optics (laser interferometers with photodetectors of dark current well above the single-photon level), (ii) , second-order nonlinear optics yielding only two-wave mixing (quantum squeezers, e.g. undepleted-pump optical parametric oscillators) and, (iii) , true three-wave mixing as well as third- and higher order nonlinear optics. While (ii) and (iii) both enter the realm of quantum optics, only (iii) involves nonlinear quantum evolution, a.k.a. non-Gaussian Wigner function gates and/or states of the quantum fields. This is why such operations—be they implemented via third-order nonlinearities or Fock state projection—were included in previous works on quantum neural networks 31 ; 8 ; 6 . Again, the gist of linear optics implementations 41 ; 63 ; 70 ; 68 is to eschew relying solely on field encoding, and hence the need for optical nonlinearities, in favor of optical parameter encoding. Figure 1 : Schematic diagram of linear optical circuit for nonlinear processing with an input \vec{x}=[x_{1},x_{2},...,x_{M}]^{T}\in\mathbb{R}^{M} . The nonlinearity is realized by encoding the input into quadrature phase shifts \mathbf{R}(x) while affine transformations are carried out using an M\times M interferometer \mathbf{U}(\vec{\theta}) and displacements d(\alpha) with the model output being extracted using homodyne measurements of the Q quadrature. Note that \vec{\theta} denotes the set of transmittivity angles for the M(M-1)/2 beamsplitters that comprise the interferometer. II Laser interferometer neural network II.1 Gaussian quantum formalism As mentioned above, we adopt here the more general quantum optical formalism of qumodes in continuous-variable quantum information 12 ; 65 ; 49 , for both the sake of generality and in anticipation of future extensions beyond the classical context of this work. Consider an M -qumode system 1 given by the vector of quadrature operators Q_{j}=(a_{j}+a_{j}^{\dagger})/\sqrt{2} and P_{j}=i(a_{j}^{\dagger}-a_{j})/\sqrt{2} of each qumode j , a_{j} being the photon annihilation operator, \displaystyle\vec{R}=\begin{pmatrix}Q_{1},P_{1},\cdots,Q_{M},P_{M}\end{pmatrix}^{T} (1) the vector of quantum expectation values \vec{r}\in\mathbb{R}^{2M} is defined as \displaystyle\vec{r}=\langle\vec{R}\rangle, (2) and the covariance matrix \mathbf{\Sigma}\in\mathbb{R}^{2M\times 2M} is given by \displaystyle\Sigma_{ij}=\langle\vec{R}_{i}\vec{R}_{j}+\vec{R}_{j}\vec{R}_{i}\rangle-2\langle\vec{R}_{i}\rangle\langle\vec{R}_{j}\rangle. (3) These two mathematical objects provide a complete description of the quantum system in the Gaussian Wigner function case for all gates and states but only the first moments will be relevant for our work. The unitary evolution of the system under Gaussian unitaries is given by a symplectic matrix \mathbf{S} which acts on the vector of means and covariance matrix as 1 \displaystyle\vec{r} \displaystyle\mapsto\mathbf{S}\vec{r} (4) \displaystyle\mathbf{\Sigma} \displaystyle\mapsto\mathbf{S}\mathbf{\Sigma}\mathbf{S}^{T}. (5) The symplectic matrix representations of the single-mode phase shift \mathbf{R}(\phi) and two-mode beamsplitter \mathbf{B}(\theta) unitaries are respectively given by 65 \displaystyle\mathbf{R}(\phi)=\begin{pmatrix}\cos\phi&\sin\phi\\ -\sin\phi&\cos\phi\end{pmatrix} (6) and \displaystyle\mathbf{B}(\theta)=\begin{pmatrix}\cos\theta\times\mathbb{I}_{2}&\sin\theta\times\mathbb{I}_{2}\\ -\sin\theta\times\mathbb{I}_{2}&\cos\theta\times\mathbb{I}_{2}\end{pmatrix} (7) where \mathbb{I}_{2} is the 2\times 2 identity matrix. The special case of displacements by M complex amplitudes

Read the complete text