A0948
Title: Efficient non-Gaussian variational inference for continuous functions using sparse autoregressive normalizing flows
Authors: Paul Wiemann - The Ohio State University (United States) [presenting]
Matthias Katzfuss - University of Wisconsin-Madison (United States)
Abstract: An innovative framework is proposed for efficient and flexible variational inference aimed at non-Gaussian posteriors of latent continuous functions or fields. The framework employs sparse autoregressive structures represented by nearest-neighbor-directed acyclic graphs for both the prior and variational families. The conditional distributions within the variational family are modeled using normalizing flows, providing high flexibility. An algorithm is introduced for doubly stochastic variational optimization, achieving polylogarithmic time complexity per iteration. Preliminary numerical comparisons suggest that the proposed method may surpass the accuracy of Gaussian variational families while maintaining comparable computational complexity.
