Neural network ensembles and variational inference revisited

back to our research

Neural network ensembles and variational inference revisited

December 2nd 2018, AABI, Montreal, Canada

Authors: Marcin Tomczak (PROWLER.io), Siddharth Swaroop (University of Cambridge), Richard Turner (University of Cambridge)

Abstract: Ensembling methods and variational inference provide two orthogonal methods for obtain-ing reliable predictive uncertainty estimates for neural networks. In this work we compareand combine these approaches finding that: i) variational inference outperforms ensem-bles of neural networks, and ii) ensembled versions of variational inference bring furtherimprovements. The first finding appears at odds with previous work (Lakshminarayananet al., 2017), but we show that the previous results were due to an ambiguous experimentalprotocol in which the model and inference method were simultaneously changed.

Reinforcement Learning

Ensembles

Variational Inference

Approximate Inference

Uncertainty Estimation


See paper

Join us to make AI that will change the world

join our team