Feature Reduction Improves Quantum Extreme Learning Machines

Annalisa De Lorenzis detailed a new approach to machine learning in her thesis, bridging quantum computing and the analysis of data from water Cherenkov detectors used in neutrino physics. Her thesis explores Quantum Extreme Learning Machines, or QELMs, a hybrid framework encoding classical data into quantum states before processing them with a classical readout layer. Within this system, De Lorenzis analyzed the interplay between quantum dynamics, expressivity, and entanglement, and how easily these quantum processes can be replicated by conventional computers. The work also demonstrates the successful application of convolutional architectures, including residual networks, for classifying complex events within simulated neutrino detector data, highlighting the potential of machine learning, in both quantum and classical forms, for fundamental physics.

A paper on QELMs in Physical Review Applied was published in April 2025 and selected for inclusion in Quantum Frontiers, a curated collection by Physical Review Applied. De Lorenzis and colleagues also co-authored a paper outlining Hyper-Kamiokande’s contribution to the European Strategy for Particle Physics, submitted to [hep-ex] in June 2025. A separate Hyper-Kamiokande sensitivity paper was published in Eur.Phys.J.C 86 (2026) 2, 170 in February 2026.

Recent work demonstrates a synergy between quantum computing and classical machine learning, achieving promising results in image classification tasks. This is not simply about harnessing quantum power, but about strategically blending the strengths of both computational paradigms. The core of this investigation, detailed in a paper published in Physical Review Applied, centers on understanding how different components within a QELM influence performance. The researchers tested various encoding schemes, including angle, dense-angle, and amplitude encoding, alongside classical feature-reduction techniques like Principal Component Analysis and convolutional autoencoders. Results indicate that “nonlinear latent representations produced by autoencoders consistently outperform PCA,” even when data compression is substantial, a critical factor for near-term quantum devices. This work extends beyond theoretical exploration; the QELM architecture was tested on standard image datasets like MNIST and Fashion-MNIST, and later CIFAR-10.

Entanglement and Classical Simulability in QELMs

A surprising approach to quantum machine learning is gaining traction beyond the established path of variational quantum circuits: Quantum Extreme Learning Machines, or QELMs. This research focuses on understanding where, and if, quantum mechanics offers a genuine advantage within this specific architecture, rather than simply achieving quantum supremacy in machine learning. A key focus of the research lies in understanding the limits of classical emulation. Researchers explored how effectively classical computers can mimic the behavior of the quantum layer within a QELM. The analysis reveals that the choice of data encoding and feature reduction techniques significantly impacts the degree to which classical simulation becomes computationally prohibitive. “The preprocessing stage is already decisive in the strongly compressed regime relevant to near-term quantum models,” the study finds, highlighting that effective classical compression can, paradoxically, enhance the potential for quantum advantage by pushing the problem beyond the reach of straightforward classical algorithms.

The findings demonstrate a clear preference for autoencoders in the preprocessing stage, suggesting that the quantum layer isn’t necessarily about performing complex quantum computations, but rather about efficiently processing and transforming data that has already been effectively pre-processed by classical methods. This fixed nature allows researchers to isolate and analyze the impact of quantum dynamics on the overall learning process, disentangling it from the complexities of optimizing quantum circuits. The study’s detailed analysis of entanglement, expressivity, and simulability provides a valuable framework for assessing the potential of QELMs and guiding the development of future hybrid quantum-classical machine learning models.

Deep Learning for Neutrino Detector Image Analysis

Researchers at Physical Review Applied are increasingly leveraging the power of convolutional neural networks to interpret data from water Cherenkov detectors, a critical step in unraveling the mysteries of neutrinos. Annalisa De Lorenzis’s work details the development of these architectures, specifically residual networks, for classifying complex events within simulated detector datasets, demonstrating an application of advanced deep learning beyond typical image recognition tasks. This shift reflects a growing trend: applying techniques honed on everyday images to the far more subtle signals produced by fundamental particles. Distinguishing genuine neutrino events from background noise requires sophisticated analysis, traditionally reliant on hand-crafted algorithms.

De Lorenzis’s team bypassed this approach, instead training deep learning models to automatically extract relevant information directly from detector data. The success of these convolutional architectures highlights their ability to discern subtle features indicative of neutrino interactions, a task previously demanding significant human expertise. The team developed these networks for the specific demands of neutrino physics.

QELM Framework: Encoding, Dynamics, and Measurement

In her recent doctoral thesis, Annalisa De Lorenzis details a hybrid approach called Quantum Extreme Learning Machines (QELMs) that seeks to leverage the strengths of both classical and quantum computation. A core focus of De Lorenzis’s work is understanding how quantum properties impact the performance of QELMs, moving beyond simply asking if a quantum advantage exists. Central to this is the challenge of efficiently embedding high-dimensional classical data into the limited number of qubits available on near-term quantum devices. De Lorenzis’s team developed convolutional architectures for this purpose, allowing for a controlled evaluation of the impact of different quantum and classical components. This rigorous testing regime allowed for a controlled evaluation of the impact of different quantum and classical components.

The work extends beyond theoretical exploration, demonstrating the potential of QELMs for real-world applications, while also outlining the challenges that remain in harnessing the power of quantum computation for complex data analysis. The thesis concludes by emphasizing the need for effective representations of complex, high-dimensional data, a common thread linking both quantum machine learning and the analysis of data from neutrino detectors.

Hyper-Kamiokande Experiment Neutrino Oscillation Parameters

Understanding neutrino behavior often feels counterintuitive; these ghostly particles routinely defy expectations, shifting between defined “flavors” in a process known as oscillation. While current experiments like T2K have begun to map this phenomenon, the next generation of detectors promises a far more detailed picture. De Lorenzis’s research extends beyond simply applying machine learning; it focuses on bridging the gap between quantum computing concepts and the practical demands of analyzing data from water Cherenkov detectors. These detectors, crucial to neutrino physics, generate complex images of particle interactions.

To interpret these images, De Lorenzis and colleagues developed convolutional architectures, demonstrating that such models “can effectively extract relevant information from detector data.” This is not merely pattern recognition; it’s about distilling meaningful signals from the noise inherent in these massive experiments. The team’s work builds on the Hyper-Kamiokande Collaboration’s ongoing efforts, as evidenced by their co-authorship on a paper detailing the experiment’s sensitivity to neutrino oscillation parameters, published in Eur. Phys. J.C in February 2026. This paper, alongside another submitted to [hep-ex] in June 2025 outlining Hyper-Kamiokande’s contribution to the European Strategy for Particle Physics, underscores the experiment’s central role in the field. The thesis also explores the potential of Quantum Extreme Learning Machines (QELMs), a hybrid quantum-classical approach. The results, published in Physical Review Applied in April 2025 and subsequently selected for inclusion in Quantum Frontiers, showcase the potential of these techniques for real-world applications, extending beyond theoretical exploration.