Bayesian Deep Ensembles via the Neural Tangent Kernel

Previous work has shown that even in the infinite width limit, when NNs become GPs, there is no GP posterior interpretation to a deep ensemble trained with squared error loss.