Epistemic Neural Networks

Abstract

We introduce the epistemic neural network (ENN) as an interface for uncertainty modeling in deep learning. All existing approaches to uncertainty modeling can be expressed as ENNs, and any ENN can be identified with a Bayesian neural network. However, this new perspective provides several promising directions for future research. Where prior work has developed probabilistic inference tools for neural networks; we ask instead, 'which neural networks are suitable as tools for probabilistic inference?'. We propose a clear and simple metric for progress in ENNs: the KL-divergence with respect to a target distribution. We develop a computational testbed based on inference in a neural network Gaussian process and release our code as a benchmark at https://github.com/deepmind/enn. We evaluate several canonical approaches to uncertainty modeling in deep learning, and find they vary greatly in their performance. We provide insight to the sensitivity of these results and show that our metric is highly correlated with performance in sequential decision problems. Finally, we provide indications that new ENN architectures can improve performance in both the statistical quality and computational cost.1 Epistemic uncertainty relates to knowledge (ancient Greek episteme↔knowledge), as opposed to aleatoric uncertainty relating to chance (Latin alea↔dice) (Kendall and Gal, 2017).Preprint. Under review.

Uncertainty in Contrastive Learning: On the Predictability of Downstream Performance

“…We show that this goal can be achieved by directly estimating the distribution of the training data in the embedding space and accounting for the local consistency of the representations. Our experiments show that this notion of uncertainty for an embedding vector often strongly correlates with its downstream accuracy.1 Aleatoric uncertainty relates to chance (Latin: alea ↔ dice) and epistemic uncertainty relates to knowledge (ancient Greek: episteme ↔ knowledge) Osband et al [2021]…”

“…1 Aleatoric uncertainty relates to chance (Latin: alea ↔ dice) and epistemic uncertainty relates to knowledge (ancient Greek: episteme ↔ knowledge) Osband et al [2021]…”

Evaluating High-Order Predictive Distributions in Deep Learning

“…Figure 9 provides some insight to the robustness of Algorithm 1 under varying number of agent samples. We make use of the epistemic neural network notation introduced by Osband et al [2021]. We can see that these monte carlo estimates converge empirically as we increase the number of samples.…”

Contextual Information-Directed Sampling

“…t } are the independent samples of contexts. We use the Epistemic Neural Networks (ENN) (Osband et al, 2021a) to quantify the posterior uncertainty of the value network. For each given s (i) t , following the procedure in Section 6.3.3 and 6.3.4 of Lu et al (2021), we can approximate the one-step expected regret and the information gain efficiently using the samples outputted by ENN.…”