Memory capacity and bandwidth are becoming major constraints in neural-network computing, motivating parameter-efficient representations that reduce model storage and data movement. Tensor-network methods already enable substantial compression of neural networks, but their structured operations are often poorly matched to conventional processors optimized for dense linear algebra. Such representations can be implemented in photonic hardware, although this typically requires optical architectures tailored to a particular factorization. Here we propose a different approach based on programmable photonic meshes. Compact electronic parameters determine the interferometer phases, while optical interference produces a dense, generally high-rank transformation. Since the parameterization remains electronic, its dimensionality and structure can be changed without modifying the optical circuit. Numerical experiments in end-to-end training demonstrate that the proposed photonic layers match or outperform strongly compressed tensor-train and brickwork layers, often with fewer trainable parameters. Experiments in parameter-efficient fine-tuning of a language model also show advantages over established digital adapters. The reduced trainable state lowers persistent memory requirements and offers the potential to reduce data movement and bandwidth demands.
Memory capacity and bandwidth are becoming major constraints in neural-network computing, motivating parameter-efficient representations that reduce model storage and data movement. Tensor-network methods already enable substantial compression of neural networks, but their structured operations are often poorly matched to conventional processors optimized for dense linear algebra. Such representations can be implemented in photonic hardware, although this typically requires optical architectures tailored to a particular factorization. Here we propose a different approach based on programmable photonic meshes. Compact electronic parameters determine the interferometer phases, while optical interference produces a dense, generally high-rank transformation. Since the parameterization remains electronic, its dimensionality and structure can be changed without modifying the optical circuit. Numerical experiments in end-to-end training demonstrate that the proposed photonic layers match or outperform strongly compressed tensor-train and brickwork layers, often with fewer trainable parameters. Experiments in parameter-efficient fine-tuning of a language model also show advantages over established digital adapters. The reduced trainable state lowers persistent memory requirements and offers the potential to reduce data movement and bandwidth demands.
There has been an escalating demand for dynamically reconfigurable photonic structures to implement any linear optical transformation between the input and output ports. Such devices are often constructed as a mesh of reconfigurable Mach-Zehnder interferometers(MZIs). Consequently, the programmed MZIs can redistribute the light to implementarbitrary unitary transformations from an input vector of optical signals to an outputvector of coherent optical modes. These structures have been widely used as a promising hardware platform for fast and energy-efficient matrix multiplications in different applications, such as information processing, machine learning algorithms, neural networks(NNs), etc. Matrix multiplication is an essential but time- and energy-consuming operation in machine learning computations. The computational complexity of NNs necessitates well designed and implemented hardware accelerators for fast and energy-efficient data processing, which allows for scaling the structure to larger NNs in current and upcoming applications. Over the last few years, the considerable progress in silicon photonics has made it possible to fabricate multiport optical interferometric structures with small ootprints and high-power efficiency, which can be exploited as hardware accelerators, thanks to the inherent parallelism of optics. This Ph.D. thesis presents the theoretical and experimental analysis of multiport reconfigurable MZI-based optical processors. The optical processors are meshes of 2×2 reconfigurable MZIs. It is demonstrated how the optical transformation matrix of the optical processors is theoretically decomposed for a given application. The Reck-mesh-based optical processor is experimentally calibrated before programming the device for a given application. Additionally, the thesis includes the theoretical and measurement results of a single layer NN implemented by a 4×4 Reck-mesh-based optical processor. Finally, a novel phase-error- and loss-tolerant multiport field-programmable MZI-based optical processor for optical NNs is proposed. The new structure with a diamond mesh shape makes use of a set of additional MZIs compared to the commonly used Reck (triangular) mesh. The additional MZIs in the Diamond mesh make its topology more symmetric yielding higher phase error- and loss-tolerance, and also provide extra degrees of freedom in the optimization of ONN weight matrix during training. Furthermore, the better classification performance of the Diamond mesh compared to the Reck mesh is associated with the fact that the Diamond mesh is capable of excluding the light intensity that could have degraded the classification accuracy of the implemented ONNs. This is done by means of the tapered-out wave guides of the additional MZIs in the Diamond mesh, while the optimal light intensity is sent to its outputs for better classification. The demonstration of these analyses is significant as it can be used to implement any universal field-programmable photonic device as an optical processor, i.e., optical computational accelerator, which can be experimentally programmed to implement arbitrary linear optical transformation matrix for a given application. Additionally, the presented analytical results in this thesis can be used to enhance the classification performance of ONNs implemented by such structures. As a result of these analyses, the thesis proposes the Diamond-mesh-based optical processor as a promising candidate for large scale ONNs which outperforms the Reck-mesh-based ones in the presence of experimental imperfections
Parametrized quantum circuits are essential components of variational quantum algorithms. Until now, optical implementations of these circuits have relied solely on adjustable linear optical units. In this study, we demonstrate that using programmable nonlinearities, rather than linear optics, offers a more efficient method for constructing quantum optical circuits -- especially quantum neural networks. This approach significantly reduces the number of adjustable parameters and the circuit depth needed to achieve high-fidelity operation. Specifically, we explored a quantum optical neural network (QONN) architecture composed of meshes of two-mode interferometers programmable by adjustable Kerr-like nonlinearities. We assessed the capabilities of our quantum optical neural network architecture and compared its performance to previously studied architectures that use multimode linear optics units. Additionally, we suggest future research directions for improving programmable quantum optical circuits.
We introduce a distributed quantum-classical framework that synergizes photonic quantum neural networks (QNNs) with matrix product state (MPS) mapping to achieve parameter-efficient training of classical neural networks. By leveraging universal linear-optical decompositions ofM-mode interferometers and photon-counting measurement statistics, our architecture generates neural parameters through a hybrid quantum-classical workflow: photonic QNNs withM(M+1)/2trainable parameters produce high-dimensional probability distributions that are mapped to classical network weights via an MPS model with bond dimensionχ. Empirical validation on MNIST classification demonstrates that photonic QT achieves an accuracy of95.50%±0.84%using 3,292 parameters (χ=10), compared to96.89%±0.31%for classical baselines with6,690parameters. Moreover, a ten-fold compression ratio is achieved atχ=4, with a relative accuracy loss of less than 3 %. The framework outperforms classical compression techniques (weight sharing/pruning) by6−12%absolute accuracy while eliminating quantum hardware requirements during inference through classical deployment of compressed parameters. Simulations incorporating realistic photonic noise demonstrate the framework's robustness to near-term hardware imperfections. Ablation studies confirm quantum necessity - replacing photonic QNNs with random inputs collapses accuracy to chance level(10.0%±0.5%). Photonic quantum computing room-temperature operation, inherent scalability through spatial mode multiplexing, and HPC-integrated architecture establish a practical pathway for distributed quantum machine learning, combining the expressivity of photonic Hilbert spaces with the deployability of classical neural networks.
We report on a gate-based variational quantum classifier implemented with single photons and probabilistic gates, to emulate the standard quantum circuit model framework. We evaluate the expressive power of two deployable quantum neural networks (QNNs) by computing their effective dimension, a capacity measure grounded in a proven generalization-error bound, and compare them with classical artificial neural networks (ANNs) of equivalent trainable-parameter count. Supervised binary classification tasks are used to benchmark performance across photonic and superconducting QNNs, both of which exhibit superior converged (lower) cross-entropy loss and (higher) prediction accuracy relative to matched-parameter ANNs. For a nonlinearly separable task, our photonic QNN with a single pair of trainable parameters successfully converged (loss 0.04 and accuracy 100%), whereas the equivalent ANN failed to learn the decision boundary, saturating at random-guessing performance. We simulate photonic quantum circuits, training them on the XOR problem and a two-class Iris subset using gradient-free optimization, and assess their robustness to sampling errors under realistic noise processes including photon loss and phase-shifter imperfections. Circuits with comparatively high effective dimension were deployed remotely on a six-qubit photonic quantum processor, achieving classification accuracies of up to 100% in both online and offline learning settings. Notably, even the simplest QNN deployed, with just two trainable parameters, successfully solved tasks that classically require ANNs with at least quadruple the number of parameters, suggesting an emergent algorithmic advantage. Overall, these results demonstrate a clear proof-of-principle that gate-based QNNs can be realized and trained effectively on current photonic hardware, providing proof of algorithmic advantage on a gate-based photonic QNN.
General-purpose programmable photonic processors offer a flexible foundation for integrating various functionalities within a single chip. A two-dimensional hexagonal waveguide mesh of Mach Zehnder interferometers has been shown to have great potential in the field of microwave photonics. Additionally, they are a promising platform for the creation of unitary linear transformations, which are key elements in photonic neural networks, In this article, we expand the portfolio of available applications for recirculating bricks mesh architecture to quantum technologies. We will show that a single programmable optical system is capable of performing various functions depending on the requirements. In particular, we will focus in this work on boson sampling, a task that best demonstrates quantum advantage, as well as on tasks that enable the determination of photon indistinguishability, which plays a key role in photonic quantum technologies. We will also show that, in addition to spatial modes, the same optical system can be equally well-suited for work on temporal modes through the implementation of an appropriate number of loops.
Reconfigurable photonics have rapidly become an invaluable tool for information processing. Light-based computing accelerators are promising for boosting neural network learning and inference and optical interconnects are foreseen as a solution to the information transfer bottleneck in high-performance computing. In this study, we demonstrate the successful programming of a transformation implemented using a reconfigurable photonic circuit with a non-conventional architecture. The core of most photonic processors is an MZI-based architecture that establishes an analytical connection between the controllable parameters and circuit transformation. However, several architectures that are substantially more difficult to program have improved robustness to fabrication defects. We use two algorithms that rely on different initial datasets to reconstruct the circuit model of a complex interferometer, and then program the required unitary transformation. Both methods performed accurate circuit programming with an average fidelity greater than 99% and 97%, respectively. Our results provide a strong foundation for the introduction of non-conventional interferometric architectures for photonic information processing.