Normalized Cut

Normalized cut is a graph-based clustering technique aiming to partition data points (e.g., image pixels, network nodes) into groups that minimize the connections between clusters while maximizing internal connectivity. Current research focuses on improving the efficiency and scalability of normalized cut algorithms, including the development of novel solvers based on coordinate descent and expander decompositions, and their application in diverse areas like image segmentation and graph neural network training. These advancements are leading to improved performance in various applications, particularly in unsupervised learning tasks such as object discovery and zero-shot semantic segmentation, where normalized cut is proving highly effective when combined with deep learning models like diffusion UNets and self-supervised transformers.