Applied Quantum Computing for Computer Vision

• Applicants: Dr.-Ing. Vladislav Golyanik; Professor Dr. Michael Möller; Professor Dr. Christof Wunderlich
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 534951134

Computer vision systems often have to process large data volumes to be able to mimic the capabilities of the human visual system. As a consequence, they require strongly increasing hardware resources to meet high data processing demands. Interstingly, only a few attempts to apply quantum computing in computer vision were undertaken so far, despite computer vision being one of the largest data-producing disciplines, with various challenging tasks and combinatorial formulations that can be potentially addressed on quantum computers. It is, thus, of utmost importance to investigate how computer vision can benefit from modern quantum hardware. While quantum computers still cannot be used to solve the largest real-world computer vision problems, we believe there is a huge mutual benefit for both disciplines, computer vision and quantum computing, in developing formulations for solving computer vision problems on quantum computers now. Thus, the aim of this proposal is to investigate how computer vision could benefit from the quantum computational paradigms. We will investigate 1) How quantum annealing can be applied in computer vision to develop approaches with new characteristics (such as higher accuracy compared to the existing state of the art), and 2) How quantum machine learning (QML) on circuit-based quantum computers can be used for auto-encoding of 3D data for improved compactness and expressivity of the learned 3D representations. This research proposal includes five work packages (WP). We are first planning to look into how to phrase the target problems in the form consumable by modern quantum annealers, i.e., quadratic unconstrained binary optimization (QUBO) objectives with binary solution encodings. As many computer vision problems naturally involve various high-level constraints, one work package is devoted to their embedding into QUBOs. Next, along with research questions relying on analytical QUBO derivation, a significant part of this proposal is devoted to learning scenario-specific costs for target problems and implicit binary representations. Moreover, we will study if iteratively defined QUBOs can be used to minimize the training costs of binary networks with the help of splitting-based optimization approaches. Last but not least, in anticipation of rapid advances in gate-based quantum hardware, we define one work package for this type of quantum computer and select a problem that is of broad interest in the community, is challenging enough, and remained largely unexplored so far, i.e., auto-encoding for 3D data. Quantum 3D data auto-encoders have the potential to improve the compactness of learned 3D representations. Questions associated with this work package are encoding of classical data into quantum states, interpreting the transformed quantum states as classical structures, the architecture or quantum circuits, and joint training of quantum/hybrid-quantum architectures with a variant of a back-propagation algorithm.

• DFG Programme: Research Grants