The authors propose a new variational framework for performing inference in stochastic differential equations (SDEs) driven by fractional Brownian motion (fBM). SDEs are versatile tools for modeling real-world dynamic systems with inherent noise and randomness. The traditional application of SDEs assumes that the underlying noise processes are generated by standard Brownian motion (BM), which falls short of capturing the full complexity of observed real data. fBM extends BM to a more complex dependence structure, but existing methods for inferring fBM parameters are either computationally demanding or statistically inefficient. The authors address these challenges by using neural networks to learn the drift, diffusion, and control terms within their variational posterior. They also present a novel architecture for variational latent video prediction.
Publication date: 19 Oct 2023
Project Page: https://arxiv.org/abs/2310.12975
Paper: https://arxiv.org/pdf/2310.12975
