An Infinite-Feature Extension for Bayesian ReLU Nets That Fixes Their Asymptotic Overconfidence

While this may seem of mostly theoretical interest, in this work, we show that it can be used in practice to the benefit of BNNs.