Epistemic Neural Networks

Effective decision, exploration, and adaptation often require an agent to know what it knows and, also, what it does not know. This capability relies on the quality of \textit{joint} predictions of labels assigned to multiple inputs. Conventional neural networks lack this capability and, since most research has focused on marginal predictions, this shortcoming has been largely overlooked. By assessing the quality of joint predictions it is possible to determine whether a neural network effectively distinguishes between epistemic uncertainty (that due to lack of knowledge) and aleatoric uncertainty (that due to chance). We introduce the \textit{epistemic neural network} (ENN) as a general interface for uncertainty modeling in deep learning. While prior approaches to uncertainty modeling can be viewed as ENNs, the 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 existing neural network, including pretrained models, and trained with modest incremental computation to represent 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 sequential decision problems. As part of this effort we open-source experiment code.