In Optica Quantum, Okinawa Institute of Science and Technology (OIST) researchers propose the first practical application of boson sampling for image identification, which is critical in a variety of domains, including forensic science and medical diagnostics.
In their simulated system, image data is first simplified using a process called principal component analysis (PCA), which reduces the amount of information while preserving key features. A complex photonic state is generated, onto which this data is encoded, before being processed in the quantum reservoir, where interference between photons produces rich, complex patterns used for image recognition. This system requires training only at the final stage (a simple linear classifier), making the overall approach both efficient and effective for accurate image recognition. Image Credit: Sakurai et al., 2025
For more than a decade, academics have seen boson sampling, a quantum computing technique employing light particles, as a critical step toward showing the superiority of quantum approaches over classical computing. However, because earlier tests have shown that traditional computers struggle to replicate boson sampling, practical uses have so far remained elusive.
Their technique requires only three photons and a linear optical network, marking a significant step toward low-energy quantum AI systems.
Harnessing Quantum Complexity
When bosons (particles like photons that obey Bose-Einstein statistics) move through certain optical circuits, they create complex interference patterns. In boson sampling, scientists send individual photons into a circuit and then analyze the resulting output probability distribution after the photons have interacted.
To understand how such sampling works, consider marbles on a pegboard. When the marbles are dropped, the probability distribution of where they land is sampled, forming a bell curve. However, when this experiment is repeated with single photons, the findings are radically different.
They exhibit wave-like behavior, which allows them to interfere with each other and interact with their environment in ways that differ significantly from how larger, macroscopic objects behave. This means that they have extremely complicated probability distributions that are difficult for traditional computing approaches to predict.
From Quantum Reservoir to Image Recognition
In this study, the researchers created a novel quantum AI approach for picture identification using boson sampling. In their simulated experiment, they started by creating a complicated photonic quantum state on which simplified picture data was encoded.
The researchers imported grey-scale images using three separate data sets. Since each pixel is in greyscale, the information is easily represented mathematically and might be compressed using principal component analysis (PCA) to preserve significant properties. This reduced data was encoded in the quantum system by changing the characteristics of individual photons.
The photons were then sent via a quantum reservoir, a sophisticated optical network, where interference resulted in rich, high-dimensional patterns. Detectors captured photon positions, and repeated sampling resulted in a probability distribution for bosons.
This quantum output was mixed with the original image data and analyzed using a simple linear classifier. This hybrid strategy maintained information while outperforming all similarly sized machine learning algorithms assessed by the researchers, resulting in extremely accurate image identification across all data sets.
Although the system may sound complex, it’s actually much simpler to use that most quantum machine learning models. Only the final step—a straightforward linear classifier—needs to be trained. In contrast, traditional quantum machine learning models typically require optimization across multiple quantum layers.
Dr. Akitada Sakurai, Study First Author and Member, Quantum Information Science and Technology Unit
“What’s particularly striking is that this method works across a variety of image datasets without any need to alter the quantum reservoir. That’s quite different from most conventional approaches, which often must be tailored to each specific type of data,” added Professor William J Munro, co-author and head of the Quantum Engineering and Design Unit.
Unlocking New Frontiers in Image Recognition
Image recognition is important in a variety of real-world applications, from interpreting handwriting at a crime scene to recognizing malignancies in MRI images. The intriguing findings of this study revealed that this quantum methodology detected photos with more accuracy than comparably sized machine learning approaches, opening up new options in quantum AI.
This system isn’t universal— it can’t solve every computational problem we give it. But it is a significant step forward in quantum machine learning, and we’re excited to explore its potential with more complex images in the future.
Kae Nemoto, Study Co-Author and Head, Quantum Information Science and Technology Unit
This study is financed in part by the MEXT Quantum Leap Flagship Program (MEXT Q-LEAP) under Grant No. JPMXS0118069605.
Journal Reference:
Sakurai, A., et al. (2025) Quantum optical reservoir computing powered by boson sampling. Optica Quantum. doi.org/10.1364/OPTICAQ.541432