8 Supplementary material

These two terms can be directly combined to obtain Eq. (8). With this expression at hand, the only problem that remains is estimating these quantities from the outputs of all neural networks involved. We provide some information here, in addition to that in Sec. Although noisy, this estimator was found to be fast and suitable for training. Finally, note that in both cases the quantities computed correspond to a difference between the entropy of the average and the average of the entropies - which is the mutual information, a known part of the EFE.