Researchers Pioneer Parameter-Efficient Anomaly Detection on Superconducting Processor
In a major advance in quantum machine learning (QML), researchers from Peking University and the Centre for Computational Science at University College London have successfully demonstrated a novel, highly parameter-efficient anomaly detection technique, Quantum Support Vector Data Description (QSVDD), on a superconducting quantum processor.
Given the constraints of present quantum technology, this work tackles the crucial challenge facing QML: can these sophisticated algorithms beat their conventional equivalents and handle real-world problems effectively? The effective implementation of QSVDD offers preliminary proof of QML’s applicability in the era of Noisy Intermediate-Scale Quantum (NISQ).
Rapid advancements in quantum computing have ushered in the NISQ era, in which computers incorporate dozens of noisy qubits. These systems have the potential to outperform traditional computing in some workloads, despite noise limiting their ability to execute fault-tolerant processing. Given these limitations, QML is one of the most promising options for achieving a useful quantum advantage.
Applying QML to intricate and varied classical datasets is difficult, especially when using noisy and constrained quantum equipment. Finding odd patterns or outliers in data is known as anomaly detection, and it’s an important field with important applications in areas like medical diagnostics, system intrusion detection, financial fraud protection, and industrial monitoring. Applications of QML to this job are still largely unexplored, despite the existence of several classical techniques.
You can also read The Superconducting Diode Effect in d-wave altermagnets
Introducing Quantum Support Vector Data Description (QSVDD)
With a focus on practical visual anomaly detection, QSVDD is a unique QML technique that aims to outperform classical models in terms of accuracy and parameter efficiency. Using a Quantum Neural Network (QNN) to convert raw data into a feature space where “normal data” is mapped onto a hypersphere is the main notion, conceptually presented. Anomalies are defined as data points that, during testing, fall outside of this learnt hypersphere.
An amplitude encoding component, a measurement component, a Variational Quantum Circuit (VQC) component, and a QSVDD post-processing component make up the network design. To better explore possible parameter efficiency, the researchers limited the amount of learnable parameters in QSVDD to a small portion, frequently only a few percent, of those utilized in conventional classical deep learning models. The study also looks at situations with little training data, which is representative of anomaly detection applications in the real world where a lot of labelled data is frequently unavailable.
Demonstrating Superior Performance in Emulation
The favorable recognition capabilities of QSVDD in comparison to classical baselines were validated by noiseless emulation findings. Using benchmarks with somewhat less trainable parameters, QSVDD demonstrated an average accuracy (as assessed by AUC) of over 90% when tested on the well-known picture datasets MNIST, Fashion MNIST, and CIFAR-10.
For example, QSVDD outperformed its Deep Support Vector Data Description (DSVDD) equivalent by 5.67% under comparable training conditions, achieving an AUC of 92.26% on the MNIST dataset with just 200 parameters and 300 training samples. With an equal number of parameters (200) employed by QSVDD and DSVDD, this relative improvement increased to 12.82%. In contrast to other deep neural networks such as DCAE, which require more than 6500 parameters to attain an AUC of 89.96% on MNIST, QSVDD showed better performance using significantly less processing power.
The stability of QSVDD was further demonstrated by the ablation research, which demonstrated strong performance even when the number of parameters and training epochs were changed. Significantly outperforming classical DSVDD despite the latter’s more trainable parameters, QSVDD showed a distinct advantage over classical deep learning when working with tiny datasets (less than 400 samples).
A theoretical examination of QSVDD’s expressivity revealed that, despite requiring a smaller number of parameters, it has an expressivity that is comparable to that of its classical counterparts. The practical viability of QSVDD is supported by the fact that the improved expressivity, which is attained through its post-processing step and effective design, does not hasten the onset of the troublesome Barren Plateaux phenomena.
You can also read Hermitian Codes Achieve Block Length With 1/3 ​Fewer Qubits
First Implementation on Quantum Hardware
The QSVDD method became the first quantum anomaly detection algorithm to be used to general picture datasets realized on a quantum device when it was implemented on a genuine quantum device. The Chinese Academy of Sciences (CAS) contributed the oneD12 12-qubit superconducting processor, which was used for the experiments.
The image data dimensionality was decreased to 16 in order to enable effective data encoding on the hardware, which had a four-qubit restriction. With just 16 parameters on the device, QSVDD achieved an accuracy of over 80% despite the noisy conditions common to NISQ devices.
The outcomes showed remarkable parameter efficiency: DSVDD required at least 300 parameters to achieve same recognition accuracy on traditional hardware, but QSVDD only required 4 qubits and 16 learnable parameters to reach this performance level. Because it lowers the complexity and difficulty of the training process, this notable reduction in parameter needs is an essential component of machine learning algorithms.
This successful implementation, which was accomplished without the use of any noise reduction or error mitigation approaches, demonstrates the resilience and versatility of QSVDD in true quantum computing settings, indicating that well-crafted QML techniques are ideally adapted for the NISQ age.
In conclusion
Combining parameter efficiency, scalability, and adaptability to general picture datasets, the creation and hardware validation of QSVDD contribute to the advancement of quantum anomaly detection. In order to further demonstrate the possibilities of QSVDD in useful, real-world problems, the researchers intend to improve the algorithm by adding error mitigation strategies and scaling tests to larger and other quantum processors.
You can also read Bosonic Binary Solver Advances Photonic Quantum Computing
