Quantum-Trained SVMs Enhanced by Locality Techniques: Large-Scale Data Analysis

Researchers have proposed a local application of quantum-trained Support Vector Machines (SVMs) to overcome the training set size limits of current quantum annealers. The team, including researchers from the University of Trento, Italy, the Jülich Supercomputing Centre in Germany, and the University of Iceland, interfaced the Fast Local Kernel Support Vector Machine (FaLKSVM) method with two quantum-trained SVM models. The empirical evaluation, conducted using D-Wave’s quantum annealers and real-world datasets from the remote sensing domain, demonstrated the effectiveness, scalability, and practical applicability of the proposed approach in large-scale scenarios.

What are Quantum-Trained Support Vector Machines (SVMs)?

Support Vector Machines (SVMs) are supervised machine learning models designed for binary classification tasks. They aim to identify the optimal hyperplane that effectively separates data samples belonging to distinct classes. With the introduction of kernel functions, SVMs can go beyond linearly separable problems. Various formulations of the learning problem exist, and also extensions to multiclass classification and regression tasks.

In recent years, with the advent of quantum annealing machines produced by D-Wave, hybrid SVM models characterized by quantum training and classical execution have been proposed. These hybrid versions have been developed for binary classification, multiclass classification, and regression tasks. These models have been evaluated mainly in the remote sensing domain, showing comparable performance with respect to their classical counterparts.

However, due to the restricted connectivity of the available quantum annealers, they are limited in the training set size. Therefore, in order to leverage large datasets, a strategy is necessary.

How Can Locality Techniques Improve SVMs?

In the classical realm, reducing the number of input samples to a machine learning model through a locality technique such as the k-nearest neighbors (kNN) algorithm has proven to be successful, yielding performance improvements compared to the base model. For instance, Blanzieri and Melgani have proposed and empirically assessed the kNN-SVM classifier, a local binary SVM trained on data samples selected by a kNN model, achieving good results.

Moreover, local SVMs have been theoretically characterized by researchers like Hable and Meister and Steinwart. However, despite the accuracy improvement and reduced training time per model resulting from the lower number of samples employed for training, an SVM must be trained on the k-neighborhood of each test sample, which is a significant bottleneck in terms of execution time.

To address this issue, Segata and Blanzieri have developed the Fast Local Kernel Support Vector Machine (FaLKSVM), which relies on the usage of the cover tree data structure.

What is the Proposed Approach for Quantum-Trained SVMs?

In this work, the local application of quantum-trained SVM models is proposed and empirically evaluated. Indeed, local classically-trained binary SVMs have already demonstrated to be successful, and quantum-trained SVMs have exhibited similar performance to their classical counterparts. Moreover, the usage of local quantum-trained models as opposed to global ones represents a valid solution to the training set size limits imposed by the connectivity of the current quantum annealers.

In practice, FaLKSVM, the method for efficient local SVMs, has been interfaced with two quantum-trained SVM models: the quantum-trained SVM for binary classification (QBSVM) and the quantum-trained SVM for multiclass classification.

Who are the Key Researchers in this Field?

The research team includes Enrico Zardini, Amer Delilbasic, Enrico Blanzieri, Gabriele Cavallaro, and Davide Pastorello. Zardini, Blanzieri, and Pastorello are affiliated with the Department of Information Engineering and Computer Science at the University of Trento, Italy. Delilbasic is affiliated with the Jülich Supercomputing Centre in Germany, the University of Iceland, and ESA/ESRIN Φ-lab in Italy. Cavallaro is affiliated with the University of Iceland, the Jülich Supercomputing Centre, and AIDAS in Germany.

What are the Practical Applications of this Research?

The empirical evaluation of the proposed approach has been conducted using D-Wave’s quantum annealers and real-world datasets taken from the remote sensing domain. The results have shown the effectiveness and scalability of the proposed approach, but also its practical applicability in a real-world, large-scale scenario.

This research has potential applications in various fields that require large-scale data analysis and classification, such as remote sensing, earth observation, and other areas where machine learning models are used.

What are the Future Directions for this Research?

The research has been submitted to the IEEE for possible publication. The future directions of this research could involve further empirical evaluation of the proposed approach using different datasets and domains, as well as exploring other potential applications of quantum-trained SVMs. The researchers may also continue to refine and improve the efficiency of the FaLKSVM method and the quantum-trained SVM models.