Figure A1: (a) word manifold capacity in word-CNN, (blue) after training, (black) before training, (b-d) character manifoldscomparedwithphonemes,words,POSmanifolds,(e)WERvswordcapacityonthetestset

We are grateful to the reviewers for their insightful and constructive comments. This is consistent with the methods of Refs. Hereweshareapreliminary resultinFigA1(b-d),22 showing that character-level manifolds also emerge across the layers of DS2. We also note that the network-level correspondence between the manifold capacity and the accuracy is40 showninthepriorworkinvision(Ref.