Fraunhofer-GesellschaftIn Quantum Machine Learning, the goal is to adapt established methods for use on quantum computers in a way that preserves their theoretical properties while, in the best-case scenario, allowing for quantum advantages—thereby enabling, for example, high-performance yet significantly more data-efficient learning methods. This is necessary to capitalize on the theoretical findings of the past decades.
At the Fraunhofer Institute for Integrated Circuits IIS, research has focused in particular on whether quantum computing can improve or accelerate so-called reinforcement learning. In reinforcement learning, an agent (e.g., an algorithm) learns the best possible solution strategy through interaction with an environment (which describes the problem to be solved). As part of the Bavarian hub of the Fraunhofer Quantum Computing Competence Network BayQS (Bavarian Competence Center for Quantum Security and Data Science), Fraunhofer IIS conducted research into the fundamentals of quantum computing-based reinforcement learning and quantum computing-based time series forecasting. In the BMBF-funded QLindA project (Quantum Reinforcement Learning for Industrial Applications), the feasibility and applicability to industrial problems were investigated in collaboration with industry partners.
Quantum machine learning is one of the key areas of focus in the quantum computing activities of the Fraunhofer Institute for Manufacturing Engineering and Automation IPA. The emphasis here is on designing algorithms that go beyond current heuristic approaches, as well as on applying quantum mechanical models to wave functions. This is reflected in the “AQUAS” project of the Federal Ministry for Economic Affairs and Climate Action (BMWK) and in work conducted within the framework of the Baden-Württemberg Quantum Computing Competence Center. In addition, the application of QML is being tested in the H2Giga project (Federal Ministry for Economic Affairs and Climate Action, BMBF) in the context of hydrogen production. The Fraunhofer IPA places a high value on open-source developments for transfer to industry. For example, as part of the BMWK project AutoQML, QM activities were consolidated and the Python library sQUlearn was developed, which provides a low-threshold introduction to a wide range of common QML algorithms. The project itself aims to develop an automated framework for QML, with the goal of further lowering the barriers to entry. To promote wider adoption, the Fraunhofer ITWM, in collaboration with the Fraunhofer Institute for Open Communication Systems (FOKUS), the Fraunhofer Institute for Applied Information Technology (FIT), Fraunhofer IPA, and Fraunhofer IAIS, has developed a QML training program as part of the Fraunhofer Development Fund.
Drawing on its 30 years of expertise in the fields of artificial intelligence and machine learning, Fraunhofer IAIS is investigating the extent to which machine learning algorithms can be meaningfully transferred to quantum computers. These methods make it possible to transfer customers’ existing ML models—such as support vector machines, logistic regression, or Bayesian networks—directly to quantum computers without the need for re-modeling or re-training. A theoretical increase in efficiency can be demonstrated, which cannot yet be fully realized by current quantum hardware. These techniques are already being successfully deployed with industry partners and are continuously adapted to the latest developments in quantum hardware.
Machine learning is one of the standard mathematical methods for solving practical problems, and this is also the case at Fraunhofer ITWM. For this reason, the topic of QML is also being intensively pursued at ITWM in a wide variety of areas:
The AnQuC project developed a Quantum Architecture Search. This multi-criteria genetic algorithm uses the ZX-Calculus to generate a quantum circuit to solve a specific problem (e.g., option pricing), taking into account multiple factors such as solution quality, the specifications of the underlying hardware, and desirable circuit properties (e.g., circuit depth). It was developed specifically for generating QML circuits (e.g., quantum Monte Carlo integration), but can also be used for variational algorithms in optimization. Another focus is on the quantum Fourier transform (QFT), which is integrated into various QML models. The QFT enables an exponential acceleration of the classical Fourier transform, which represents a bottleneck in the creation of surrogate models for multiscale simulations of composite materials. Initial results have been achieved on IBM quantum computers. QAE was used to model a unit step function on a quantum computer. This is required, among other things, in the option pricing of financial derivatives. Another focus of our work was on quantum neural networks (QNNs). Here, we investigated the frequency spectrum of the underlying Fourier series and discovered properties that play an important role in the more efficient implementation of QNNs.
In addition, Fraunhofer ITWM is developing quantum AI pipelines—as part of the Quant²AI project, among other initiatives—for applications such as anomaly detection, clustering, classification, regression, and image recognition and segmentation. This project provides a comprehensive overview of relevant AI applications and the evaluation metrics of the benchmark. First, example problems and datasets were prepared for the benchmark, and a modular demonstrator was implemented that integrates all aspects of the pipeline and evaluates its performance. Finally, a visualization concept is being developed to clearly and visually represent the performance of the entire pipeline.
The implementation of quantum algorithms to accelerate and simplify machine learning (ML) approaches often proves to be highly application-specific and complex. The Automated Machine Learning (AutoML) approach already makes it possible to automate the implementation of ML and make it more user-friendly. Fraunhofer IAO is therefore working to extend these approaches to include quantum computing methods—particularly in the manufacturing and automotive sectors—thereby facilitating access to QML methods. This involves the development and integration of application-specific QML algorithms, as well as the development of quantum optimization algorithms for efficient (Q)ML selection using quantum neural networks (QNNs), quantum support vector machines (QSVMs), quantum computing (Convolutional) Neural Networks (QC(C)NNs), or quantum circuit learning.