Seminars

Home Centers & Programs AI and Natural Sciences Seminars

ABSTRACT: Hierarchical generative modeling is of vital importance in a wide range of real-world applications as it can provide realistic content generation as well as compact multi-layer data representations. However, prevailing hierarchical learning approaches have proven inadequate,characterized by limited modeling of contextual dependencies in a hierarchical latent space and ineffective learning of complex latent structures. In this talk, I will present new hierarchical generative frameworks, enhanced by the context-aware latent space energy-based models, that could surpass the current capacity of generative modeling. The proposed frameworks seamlessly integrate two key families of probabilistic models: one being the top-down generative model (for hierarchical structural modeling), and the other being the bottom-up energy-based model (for latent contextual modeling). Such frameworks have enhanced model expressivity in generating high-quality images and capturing meaningful hierarchical representations.