UCLA

The Energy-Based Model (EBM) represents a class of probabilistic generative models that offers a robust and versatile framework for modeling arbitrary data distributions. In recent years, EBMs have increasingly captured the interest of both the academic and industrial sectors. Despite their potential, the training and practical application of EBMs pose significant challenges. This thesis conducts a systematic study of EBMs, beginning with a case study on 3D modeling. This example not only showcases the strengths of EBMs but also highlights the difficulties encountered during their training process. The discussion then progresses to the intricacies of EBM training, particularly emphasizing the challenge of the time-consuming sampling process, which may not always yield beneficial samples for updating EBMs. To address this issue, we introduce a strategy to amortize the sampling process using specially designed initializer models. We developed algorithms to facilitate cooperative training between the EBM and the initializer models. The resulting algorithms, CoopFlow and CDRL, demonstrate competitive performance across a variety of tasks, showcasing their efficacy in overcoming traditional EBM training hurdles