Beyond ELBOs: A Large-Scale Evaluation of Variational Methods for Sampling

• DOI: 10.48550/arxiv.2406.07423
• Publication Date: 2024-06-11
• Authors: Denis Blessing; Xiaogang Jia; Johannes Esslinger; Francisco Vargas; Gerhard Neumann
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Statistics - Machine Learning; Machine Learning (stat.ML); Machine Learning (cs.LG)

Monte Carlo methods, Variational Inference, and their combinations play a pivotal role in sampling from intractable probability distributions. However, current studies lack unified evaluation framework, relying on disparate performance measures limited method comparisons across diverse tasks, complicating the assessment of progress hindering decision-making practitioners. In response to these challenges, our work introduces benchmark that evaluates methods using standardized task suite broad range criteria. Moreover, we study existing metrics for quantifying mode collapse introduce novel this purpose. Our findings provide insights into strengths weaknesses serving as valuable reference future developments. The code is publicly available here.