Epistemic Neural Networks

Intelligence relies on an agent's knowledge of what it does not know. This capability can be assessed based on the quality of \textit{joint} predictions of labels across multiple inputs. Conventional neural networks lack this capability and, since most research has focused on marginal predictions, this shortcoming has been largely overlooked. We introduce the \textit{epistemic neural network} (ENN) as an interface for models that represent uncertainty as required to generate useful joint predictions. While prior approaches to uncertainty modeling such as Bayesian neural networks can be expressed as ENNs, this new interface facilitates comparison of joint predictions and the design of novel architectures and algorithms. In particular, we introduce the \textit{epinet}: an architecture that can supplement any conventional neural network, including large pretrained models, and can be trained with modest incremental computation to estimate uncertainty. With an epinet, conventional neural networks outperform very large ensembles, consisting of hundreds or more particles, with orders of magnitude less computation. We demonstrate this efficacy across synthetic data, ImageNet, and some reinforcement learning tasks. As part of this effort we open-source experiment code.