Quantum-enhanced pattern recognition

Giuseppe Ortolano,Carmine Napoli,Cillian Harney,Stefano Pirandola,Giuseppe Leonetti,Pauline Boucher,Elena Losero,Marco Genovese,Ivano Ruo-Berchera

DOI: https://doi.org/10.1103/PhysRevApplied.20.024072

2023-04-12

Abstract:The challenge of pattern recognition is to invoke a strategy that can accurately extract features of a dataset and classify its samples. In realistic scenarios this dataset may be a physical system from which we want to retrieve information, such as in the readout of optical classical memories. The theoretical and experimental development of quantum reading has demonstrated that the readout of optical memories can be dramatically enhanced through the use of quantum resources (namely entangled input-states) over that of the best classical strategies. However, the practicality of this quantum advantage hinges upon the scalability of quantum reading, and up to now its experimental demonstration has been limited to individual cells. In this work, we demonstrate for the first time quantum advantage in the multi-cell problem of pattern recognition. Through experimental realizations of digits from the MNIST handwritten digit dataset, and the application of advanced classical post-processing, we report the use of entangled probe states and photon-counting to achieve quantum advantage in classification error over that achieved with classical resources, confirming that the advantage gained through quantum sensors can be sustained throughout pattern recognition and complex post-processing. This motivates future developments of quantum-enhanced pattern recognition of bosonic-loss within complex domains.

Quantum Physics,Optics

What problem does this paper attempt to address?

The paper aims to address challenges in pattern recognition, particularly by improving classification accuracy through quantum-enhanced techniques. Specifically, the research team utilized entangled photon states (TMSV states) and photon counting measurement techniques to demonstrate quantum advantage over classical methods in multi-unit pattern recognition tasks. The MNIST handwritten digit dataset was used as a test benchmark in the experiments, and the superiority of quantum technology in reducing classification error rates was validated through two different readout configurations—scanning readout and parallel multi-pixel readout.

Key findings of the research include:
1. **Experimental validation of quantum advantage**: Quantum-enhanced pattern recognition was achieved experimentally, confirming an advantage in classification error rates compared to the best classical strategies.
2. **Comparison of different readout methods**: The performance of quantum technology was studied under two configurations: scanning readout and single-shot multi-pixel readout. The former focused on demonstrating the magnitude of quantum advantage, while the latter was closer to practical applications.
3. **Applicability to complex post-processing**: It was shown that quantum-enhanced techniques are effective not only with simple k-nearest neighbor classifiers but also maintain their advantage in complex convolutional neural network (CNN) classifiers.

These results indicate that quantum sensors have significant advantages in pattern recognition tasks, especially in noisy environments, providing theoretical and experimental support for the future development of quantum-enhanced pattern recognition technologies.