Source-linked AI summary
Quantum Convolutional Neural Networks for High Energy Physics Data Analysis
Samuel Yen-Chi Chen, Tzu-Chieh Wei, Chao Zhang, Haiwang Yu, Shinjae Yoo
TL;DR
The paper addresses limited progress in robust quantum representation learning for HEP data. It proposes and evaluates a QCNN on simulated DUNE data, finding faster learning and higher test accuracy than comparable CNNs, motivating further study of quantum machine learning in HEP and other scientific fields.
Problem
Quantum computing has made limited progress toward robust representation-learning methods for HEP data, despite the need to analyze growing data volumes.
Method
The paper develops a hybrid QCNN using variational quantum circuits and tests it on simulated HEP events from a neutrino experiment.
Results
With comparable parameter counts, the QCNN learns faster or achieves better test accuracy with fewer training epochs than classical CNNs.
Takeaways & Limitations
The results support studying QCNNs and other quantum machine-learning models for HEP and additional scientific applications.
Abstract
from arXiv · showhide
This work presents a quantum convolutional neural network (QCNN) for the classification of high energy physics events. The proposed model is tested using a simulated dataset from the Deep Underground Neutrino Experiment. The proposed architecture demonstrates the quantum advantage of learning faster than the classical convolutional neural networks (CNNs) under a similar number of parameters. In addition to faster convergence, the QCNN achieves greater test accuracy compared to CNNs. Based on experimental results, it is a promising direction to study the application of QCNN and other quantum machine learning models in high energy physics and additional scientific fields.
I. INTRODUCTION
The paper introduces QCNNs for HEP event classification, addressing limited quantum progress in robust representation learning. Using simulated neutrino-detector data, it compares QCNNs with similarly parameterized CNNs and reports faster learning and higher testing accuracy.
- Motivation and contribution: QCNNs are proposed to bring robust quantum representation learning to HEP event classification, where such methods have seen limited progress.The framework is presented as a hybrid quantum convolutional neural network for comparison with classical algorithms.
- Data and task: The study uses simulated DUNE neutrino-experiment data to develop and test QCNN algorithms for HEP events.The data are simulated for neutrino experiments and used to compare quantum and classical classifiers.
- Results: With similar parameter counts, QCNNs can learn faster or reach better test accuracy with fewer training epochs than classical CNNs.The introduction frames this as an empirical comparison of learning speed and testing accuracy.
- Data and task: The simulated detector images contain particle-specific spatial patterns, making them suitable for CNN-type pattern recognition.The dataset includes four simulated particle types and images generated through LArTPC detector-response and signal-processing steps.
III. VARIATIONAL QUANTUM CIRCUITS
Variational quantum circuits encode classical inputs into quantum states, apply trainable parameterized operations, and return measurement-derived features for further processing.
- VQC structure: A variational quantum circuit uses tunable parameters optimized through classical iterative methods such as gradient descent.The parameters in the variational block are optimized with gradient-based methods, including Adam and RMSProp.
- VQC structure: The F(x) block encodes classical data x into a quantum state and is designed for the research problem rather than optimized.The V(θ) block supplies the learnable variational component.
- Measurement: Quantum measurements produce stochastic bit strings, so repeated circuit executions estimate expectation values for each qubit.The paper illustrates this estimation using repeated measurements and observed frequencies.
- Measurement: The resulting expectation values form an N-dimensional vector that can be processed by classical or quantum neural networks.The paper notes that Pauli-Z expectation values are used at the end of its VQC.
IV. QUANTUM CNN
The QCNN replaces classical convolutional filters with variational quantum circuits that sweep across image pixels, measure local inputs, and produce lower-dimensional representations.
- Motivation: The framework seeks possible quantum acceleration and advantage by comparing QCNN performance with classical CNNs.The motivation is connected to the expressive power attributed to quantum entanglement in the variational kernels.
- Quantum convolution: QCNNs use variational quantum circuits to perform convolutional operations instead of classical neural-network convolutional filters.The architecture is presented as a quantum implementation of convolution for image-based HEP analysis.
- Quantum convolution: Quantum convolutional kernels sweep through input image pixels and transform them into lower-dimensional representation vectors through measurements.This operation is the core quantum analogue of convolution described for the architecture.
- Quantum convolution: A stack of VQCs captures features at varied length scales across the input.The stacked architecture is described as capturing features of different spatial scales.
A. Quantum Convolutional Filters
The quantum convolutional filter encodes each n×n image patch into rotation angles, processes it with a quantum circuit, and produces measurements for subsequent processing.
- Data encoding: Pixel inputs are transformed into rotation angles before being encoded into a quantum state.The paper applies an arc tangent transformation because pixel values are not generally restricted to [−1, 1].
- Quantum state: A general N-qubit state represents amplitudes for computational-basis states whose probabilities sum to one.The amplitudes are complex, and their squared magnitudes determine measurement probabilities.
- Data encoding: The filter receives an n×n matrix, where n is the filter or kernel size, and flattens it before quantum processing.The flattened pixel values are transformed into gate-rotation angles.
- Data encoding: Each of the n×n input values generates two rotation angles, producing 2 × n × n angles for single-qubit rotations.This specifies the angle expansion used in the encoding layer.
2. Variational Layer
The QCNN encodes input values, applies variational quantum transformations, and measures expectation values while using quantum filters to extract spatial features through convolutional layers.
- Variational Layer: Input values are encoded into a quantum state before passing through variational transformations.The variational layer follows state preparation and unitary transformations.
- Variational Layer: The variational layer combines CNOT-based entanglement with single-qubit rotations parameterized by α_i, β_i, and γ_i.These parameters are updated during optimization.
- Variational Layer: Quantum measurements transform VQC outputs into expectation values, with repeated measurements required on real quantum computers to obtain sufficient statistics.Simulation software can calculate the values deterministically.
- Variational Layer: The quantum convolutional filter outputs one value per sweep step by measuring the first qubit.This produces the filter response used by the QCNN.
- Quantum Convolutional Operations: Convolutional output dimensions depend on input dimension, filter size, padding, and stride according to the stated convolution rule.The supplied variable descriptions define W_out, W_in, F, and P, while the rule also includes S.
- Quantum Convolutional Operations: QCNN filters sweep across input pixels, transform an n × n vector into a single value, and can be stacked or multiplied to extract multiple features.Each filter captures a single kind of feature.
C. Classical Post-processing
The QCNN combines quantum convolutional processing with classical post-processing, gradient-based optimization, and regularization-oriented design choices.
- Classical Post-processing: The final quantum convolutional output is flattened, passed through a fully connected classical layer, and converted into class probabilities with softmax.This forms the classical output stage of the hybrid model.
- Loss Function: Cross-entropy loss evaluates classification predictions using class indicators and predicted probabilities.The supplied passages identify M as the number of classes and define the logarithm and probability terms.
- Optimization: The parameter-shift rule is used to calculate gradients of measured quantum functions.Quantum measurements are stochastic, so gradient calculation differs from classical neural-network layer operations.
- Quantum Processing: Quantum convolutional filters encode input values, apply variational gates, and retrieve Pauli-Z expectation values.Input values are mapped to Ry and Rz rotation angles before measurement.
- Quantum Processing: The QCNN uses a VQC as its convolutional kernel, with qubit count equal to the square of the filter size and measurement restricted to the first qubit.The number of qubits and measurements can be adjusted for the task.
E. Dropout
The paper motivates dropout as a way to reduce overfitting, including for QCNNs trained on relatively small datasets, by randomly zeroing layer outputs during training.
- Dropout: Overfitting occurs when a model learns statistical noise in the training data, reducing performance on unseen data.The paper describes this as poor generalizability.
- Dropout: Small training sets can create overfitting difficulties for classical DNNs, and QCNNs are no exception.The same generalization concern is stated for QCNN training.
- Dropout: Dropout approximates averaging predictions from many neural-network architectures without training all of them separately.Training every possible architecture would require unlimited computational resources.
- Dropout: During training, dropout independently sets selected layer outputs to zero with probability p, but it is not applied during testing.The zeroing occurs on each forward pass.
V. EXPERIMENTS AND RESULTS
The experiments simulate QCNN classification of 30 × 30-pixel HEP images and compare quantum and classical CNNs using similar parameter counts.
- Experiments and Results: The experiments numerically simulate QCNN classification on different high-energy-physics events.The supplied section introduces the numerical simulation task.
- Experimental Design: The quantum and classical CNNs are arranged with similar numbers of parameters to support a fair comparison.Parameter comparability is used to examine possible quantum advantage.
- Classical CNN: The classical CNN uses four first-layer channels, two second-layer channels, and 498 total parameters.Its fully connected layer contributes 198 parameters.
- QCNN: The QCNN uses 3 × 3 and 2 × 2 quantum filters followed by a classical fully connected layer with 394 parameters.The architecture has one channel in each quantum convolutional layer.
- Implementation: The experiments use PyTorch, PennyLane, and Qulacs for the hybrid quantum-classical implementation.These software packages support the reported simulation workflow.
A. Muon versus Electron
The QCNN is evaluated on simulated binary particle-classification tasks against a classical CNN with comparable parameter counts. It matches CNN accuracy for muon–electron classification and outperforms it for muon–proton classification in both accuracy and learning speed.
- Muon versus Electron: 92.5% QCNN test accuracy and 95% CNN test accuracy are comparable for muon–electron classification.The comparison uses a comparable number of parameters.
- Muon versus Electron: The QCNN converges to its optimal accuracy much faster than the CNN for muon–electron classification.The passage attributes the faster convergence to requiring fewer epochs.
- Muon versus Proton: Muon–proton classification is more difficult than muon–electron classification, with the CNN reaching 80% test accuracy.Proton tracks are described as higher-intensity and straighter than muon tracks because of greater mass and lower multiple Coulomb scattering.
C. Muon versus Charged Pion
The QCNN is tested on the difficult task of distinguishing muons from charged pions, whose detector signatures are similar except for additional pion nuclear interactions. It achieves higher accuracy and faster learning than the classical CNN while supporting larger input dimensions than cited prior architectures.
- Muon versus Charged Pion: Muon–charged-pion classification is difficult because the particles share similar energy loss, multiple Coulomb scattering, and decay behavior.Charged pions additionally undergo nuclear interactions that can produce a trajectory kink.
- Related Work and Architecture: Unlike cited approaches using randomly sampled quantum circuits or QRAM, this work trains its quantum and classical components end to end on classical input data.The related-work discussion identifies QRAM as difficult to implement on physical devices in the near term.
VII. CONCLUSION AND OUTLOOK
The paper proposes a QCNN framework for HEP-event learning and reports faster learning and higher testing accuracy than a comparably parameterized classical CNN. It identifies broader applications while noting simulation, input-size, and circuit-evaluation constraints.
- The proposed QCNN framework demonstrates significant learning capacity for high energy physics event classification.
- The QCNN outperforms a classical CNN in learning speed and testing accuracy with a comparable number of parameters.
- The framework is expected to support applications in NISQ-era quantum computing, additional HEP experiments, and other scientific fields.
- The experiments use 30×30 inputs and a single channel, while higher-dimensional inputs and multiple channels remain future directions.
- The study is a noise-free proof of concept, and parameter-shift gradients require circuit evaluations that scale with circuit width and depth.
- The general QCNN architecture is not limited to image classification and can extend to domains involving temporal or sequential dependencies.