Amortized Posterior on Latent Variables in Gaussian Process

Qing Sun

Keywords: meta-learning, uncertainty estimation

Abstract: Deep neural networks have achieved impressive performance on a variety of domains. However, performing tasks in partially observed, dynamic environments is still an open problem. Gaussian Process (GPs) is well-known for capturing uncertainty in model parameters. However, it simply assumes a fixed Gaussian prior on latent variables. Thus, agents are not able to update their beliefs about latent variables as observing data points. Instead, in this paper, we propose to replace the prior with an amortized posterior, which enables quick adaptation, especially to abrupt changes. Experiments show that our proposed method can adjust behaviors on the fly (e.g., blind “Predator” take 56% more chance to approach “Prey”), correct mistakes to escape bad situations (e.g., 25% ↑ on avoiding repeating to hit objects with negative rewards), and update beliefs quickly (e.g., 9% faster convergence rate on learning new concepts).