Probabilistic Deep Learning with Generalised Variational Inference
We study probabilistic Deep Learning methods through the lens of Approximate Bayesian Inference. In particular, we examine Bayesian Neural Networks (BNNs), which usually suffer from multiple ill-posed assumptions such as prior and likelihood misspecification. In this direction, we investigate a recently proposed approximate inference framework called Generalised Variational Inference (GVI) in comparison to state-of-the-art methods including standard Variational Inference, Monte-Carlo Dropout, Stochastic gradient Langevin dynamics and Deep Ensembles. Also, we expand the original research around GVI by exploring a broader set of model architectures and mathematical settings on both real and synthetic data. Our experiments demonstrate that approximate posterior distributions derived from such a method offer attractive properties with respect to uncertainty quantification, prior specification robustness and predictive performance, especially in the case of BNNs.
PDF Abstract