Sebastian Loeschke

_______________________________________________________

MSc in Computer Science
PhD @ IT-University of Copenhagen

Abstract

Tensor Networks for Efficient Image Rendering

Efficient and realistic image rendering (IR) has long been a focus of research. Machine learning (ML) techniques for IR have enabled the creation of complex and photorealistic images. Despite recent advances, these techniques are often slow and memory-intensive, limiting their practical use.

This Ph.D. proposal aims to explore the potential of quantum-inspired tensor network (TN) methods for IR tasks, with the goal of reducing memory and computational costs. TNs are versatile and powerful scientific simulation tools that have been successful in simulating strongly correlated quantum many-body systems and quantum circuits. TNs have also been used to compress deep neural networks, leading to significant memory savings in ML applications. However, TNs have not been utilized as extensively as neural networks in ML, and the development of tools and techniques for training them has been limited.

This project will develop novel algorithms and techniques that leverage TNs’ full capabilities in an ML and IR setting to achieve real-time or animated 3D IR at high precision. The project will identify promising TN embeddings for images and scenes, and develop efficient learning algorithms for constructing them. Specific projects include exploring discrete vs. continuous TN embeddings, upsampling methods, and incorporating TNs into normalizing flows and diffusion models to improve representational power and inference time.

This project has the potential to significantly contribute to the fields of ML, IR, quantum computation, and life sciences, which heavily rely on the analysis of large datasets. By developing efficient IR techniques, this project aims to make IR more practical and accessible, benefiting fields such as medical imaging, gaming, virtual and augmented reality, and robotics. Additionally, TN methods have the potential to significantly reduce the carbon footprint of ML applications by developing more efficient algorithms that can process large datasets with fewer computational resources. This will not only benefit the environment but also democratize ML by making it more accessible to a wider range of individuals. In addition, the use of TNs allows for better explainability compared to deep learning models. Lastly, this project will contribute to the collaboration between the quantum and ML communities and also help map out the landscape where TN algorithms provide an advantage, paving the way for future advancements in quantum-native algorithms.