Efficient and Accurate Likelihood Estimation via Learning Amortized Adaptive Proposal Distributions

TMLR Paper5673 Authors

Abstract: Recent advancements in probabilistic modeling have driven significant progress in deep learning, particularly through the development of generative models based on variational inference. These models optimize a tractable lower bound of the log-likelihood, rather than the log-likelihood itself. However, they often encounter trade-offs between approximation accuracy and computational efficiency. To address these limitations, we propose a novel generative model grounded in importance sampling. Central to our approach is the Amortized Adaptive Proposal Distribution (AAPD), which simultaneously serves as both the proposal distribution for importance sampling and an approximation to the posterior. Extensive evaluations on both synthetic and real-world datasets demonstrate the superior performance and versatility of our method in latent variable modeling. Additionally, we extend our model to mixed-effects settings, effectively addressing some limitations of traditional statistical approaches.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Tom_Rainforth1

Submission Number: 5673