781877bda0783aac5f1cf765c128b437-AuthorFeedback.pdf

The former requires training multiple NNs and performs similarly to deep ensembles. These will be added to the Appendix. Expressive power: DUNs trade off expressivity and explanation diversity automatically; Earlier layers19 obtain lowaccuracy's(seeb))and are assigned lowposterior probabilities (see 3.1).