9ac1382fd8fc4b631594aa135d16ad75-AuthorFeedback.pdf

For ˆz = 2, boththenumericalresults(full)andourthird-orderprediction (dashed)) separate at the end of training. Power law expansion We thank the reviewers for highlighting3 this point, as it led to a better formulation of the expansion. Furthermore, when we19 truncate thechangesin connectivity, a rank 10 matrix is sufficient to achieve full performance. This is compared20 toarank 200 matrix when trying tocompress the full connectivity. Inparticular,thelearning rate is24 often chosen so high that learning dynamics become highly rugged.