Machine Learning Framework Systematically Benchmarks 18 Quantum Models for Practical Gains

Scientists are increasingly focused on identifying practical advantages of quantum machine learning (QML) models, necessitating a shift from isolated algorithmic development towards systematic empirical investigation across diverse models, datasets and hardware limitations. Addressing this need, Cassandre Notton and Benjamin Stott from Quandela Quantique Inc., alongside Philippe Schoeb from DIRO & Mila, Université de Montréal, Anthony Walsh from Quandela, Grégoire Leboucher from ENS Paris-Saclay, Vincent Espitalier and Vassilis Apostolou from Quandela, and Louis-Félix Vigneux, Alexia Salavrakos, and Jean Senellart, and colleagues, present MerLin, an open-source framework designed as a discovery engine for photonic and hybrid machine learning. This collaborative effort integrates optimised linear optical circuits into standard machine learning workflows, enabling end-to-end differentiable training of layers and establishing a shared experimental baseline through the reproduction of eighteen state-of-the-art QML works. By embedding photonic models within established ecosystems, MerLin facilitates ablation studies, cross-modality comparisons, and hybrid classical workflows, positioning it as a future-proof tool for co-designing algorithms, benchmarks, and hardware.

This new platform integrates optimised simulations of linear optical circuits directly into standard machine learning workflows using PyTorch and scikit-learn, enabling fully differentiable training of quantum layers.

MerLin addresses a critical need for systematic benchmarking and reproducibility within the rapidly evolving field of quantum machine learning, moving beyond isolated algorithmic proposals toward empirical exploration across diverse models, datasets, and hardware. The framework’s architecture allows researchers to leverage existing machine learning tools for detailed analysis, comparisons, and the creation of hybrid classical-quantum systems.

By embedding photonic quantum models within established artificial intelligence ecosystems, MerLin facilitates ablation studies and cross-modality comparisons previously hindered by fragmented software landscapes. The research team has initially reproduced eighteen state-of-the-art photonic and hybrid quantum machine learning works, encompassing kernel methods, reservoir computing, convolutional and recurrent architectures, generative models, and modern training paradigms.

These reproductions are released as reusable, modular experiments, establishing a shared experimental baseline consistent with empirical benchmarking methodologies widely adopted in modern artificial intelligence. This approach ensures that results are verifiable and adaptable, fostering collaboration and accelerating progress.

MerLin’s hardware-aware features allow for testing on available quantum hardware while simultaneously enabling exploration beyond current capabilities, positioning it as a future-proof co-design tool. The framework links algorithms, benchmarks, and hardware, addressing the challenge of translating theoretical advances into practical applications.

This capability is particularly relevant for photonic quantum computing, a promising platform due to its scalability, robustness, and compatibility with existing optical communication technologies. By providing a unified platform for implementation, training, and evaluation, MerLin promises to unlock the full potential of photonic quantum machine learning and drive innovation in the field.

MerLin framework validation, computational complexity and scalability limitations

Reproductions within the MerLin framework have successfully validated eighteen state-of-the-art photonic and hybrid quantum machine learning works, establishing a consistent baseline for empirical benchmarking. These reproductions demonstrate MerLin’s correctness and provide a foundation for future contributions to the field.

The framework’s core relies on strong linear optical simulation, achieving a time complexity of O(n m+n−1 n), where ‘n’ represents the number of photons and ‘m’ the number of modes. This simulation method efficiently computes quantum states, avoiding redundant calculations of matrix permanents and offering a significant speedup for practical applications.

MerLin’s implementation necessitates a memory footprint of O(n m+n−1 n), currently limiting practical simulations to approximately 20 photons on standard hardware. However, this constraint positions the framework as a valuable co-design tool, linking algorithms, benchmarks, and hardware capabilities for future exploration.

By embedding photonic models within established machine learning ecosystems like PyTorch and scikit-learn, MerLin facilitates ablation studies, cross-modality comparisons, and hybrid classical workflows. Independent searches confirm that photonic contributions currently account for approximately 6% of all quantum machine learning publications, a figure consistent with the research team’s findings.

Code availability remains a challenge within the broader QML landscape, hovering around 27% for gate-based approaches and 43% for photonic QML papers, highlighting the importance of MerLin’s focus on reproducibility. The framework’s design prioritises not only identifying performance gains but also understanding their origins, disentangling improvements stemming from data preprocessing, model engineering, or genuinely novel computational mechanisms.

Reproducing established quantum machine learning via differentiable photonic simulation

MerLin, an open-source framework, underpins this work as a discovery engine for photonic and hybrid quantum machine learning. The core of MerLin integrates optimised strong simulation of linear optical circuits directly into established machine learning workflows utilising PyTorch and scikit-learn. This integration facilitates end-to-end differentiable training of quantum layers, a crucial step towards seamless hybrid quantum-classical model development.

By embedding photonic models within these standard ecosystems, researchers can readily apply existing tools for detailed ablation studies, cross-modality comparisons, and the creation of complex hybrid classical-quantum workflows. The study systematically reproduced eighteen state-of-the-art photonic and hybrid QML works, encompassing diverse architectures including kernel methods, reservoir computing, convolutional and recurrent networks, generative models, and contemporary training paradigms.

These reproductions were deliberately designed as reusable, modular experiments, establishing a shared experimental baseline aligned with empirical benchmarking methodologies common in modern artificial intelligence. This modularity allows for direct extension and adaptation of existing experiments, fostering collaborative research and accelerating the pace of discovery.

MerLin’s design prioritises systematic benchmarking and reproducibility, addressing a critical need within the rapidly evolving field of quantum machine learning. The framework implements hardware-aware features, enabling tests on currently available quantum hardware while simultaneously facilitating exploration of algorithms beyond current hardware limitations.

This forward-looking approach positions MerLin as a co-design tool, linking algorithms, benchmarks, and hardware in a cohesive and future-proof manner. The choice of linear optical circuits stems from their potential scalability, robustness, and compatibility with existing optical communication technologies, offering a promising pathway towards practical quantum computation.

The Bigger Picture

The relentless pursuit of quantum machine learning has, until now, often felt like chasing a mirage. Demonstrations of potential speedups have frequently been overshadowed by the sheer difficulty of building practical, scalable systems. This new framework, MerLin, represents a significant shift by focusing not on a single quantum algorithm, but on providing the tools to systematically explore a wide range of possibilities.

It’s a move away from isolated proofs-of-concept and towards a more engineering-led approach, acknowledging that the path to useful quantum computation will likely be paved with hybrid classical-quantum solutions. MerLin’s strength lies in its integration with existing machine learning ecosystems. By embedding photonic models within established workflows like PyTorch and scikit-learn, it allows researchers to leverage familiar tools for crucial tasks like ablation studies and cross-modality comparisons.

This lowers the barrier to entry, enabling a broader community to contribute to the field and accelerating the process of identifying genuinely promising applications. Reproducing eighteen existing photonic and hybrid QML works is no small feat, and establishing a shared experimental baseline is vital for meaningful benchmarking.

However, the framework’s current focus on photonic systems represents a limitation. While photonics offers advantages in terms of room-temperature operation and connectivity, it is not the only viable quantum hardware platform. Furthermore, the true test will be demonstrating performance gains on real-world datasets and problems, not just on carefully curated benchmarks.

The next step must involve tackling more complex, messy data and evaluating the framework’s ability to handle the noise and imperfections inherent in current quantum hardware. Ultimately, MerLin’s success will depend on its ability to bridge the gap between theoretical potential and practical utility, guiding the field towards a future where quantum and classical computation work in concert.