Hierarchical Data Structures Through the Lenses of Diffusion Models

Learn about the relationship between hierarchical data structures and diffusion models in this research seminar presented at Harvard CMSA by Antonio Sclocchi from EPFL. Explore how diffusion models can be used to analyze and understand the hierarchical composition of data, particularly focusing on the distinct behaviors of high- and low-level features during noising-denoising processes. Discover the fascinating phenomenon where high-level features experience sharp transitions at specific noise levels, while low-level features recombine to form new data across different classes. Examine experimental validations using state-of-the-art diffusion models applied to both image and text data, revealing correlating changes in real-space variables and diverging correlation lengths at transition points. Gain valuable theoretical insights into how hierarchical models capture complex data structures and their implications for understanding generative AI systems.