Fast yet Simple Natural-Gradient Descent for Variational Inference in Complex Models
This work addresses the problem of efficient Bayesian inference for machine learning practitioners, offering a method that is incremental but enhances existing variational inference techniques.
The paper tackles the computational challenges of Bayesian inference in complex models like deep neural networks by proposing a fast and simple natural-gradient descent method for variational inference, which improves convergence by exploiting information geometry and provides accurate local approximations for model components.
Bayesian inference plays an important role in advancing machine learning, but faces computational challenges when applied to complex models such as deep neural networks. Variational inference circumvents these challenges by formulating Bayesian inference as an optimization problem and solving it using gradient-based optimization. In this paper, we argue in favor of natural-gradient approaches which, unlike their gradient-based counterparts, can improve convergence by exploiting the information geometry of the solutions. We show how to derive fast yet simple natural-gradient updates by using a duality associated with exponential-family distributions. An attractive feature of these methods is that, by using natural-gradients, they are able to extract accurate local approximations for individual model components. We summarize recent results for Bayesian deep learning showing the superiority of natural-gradient approaches over their gradient counterparts.