3a01fc0853ebeba94fde4d1cc6fb842a-AuthorFeedback.pdf
–Neural Information Processing Systems
Note that allour experiments areexamples when parametric42 SGD has been stuck at local optimal, but splitting allows us to further decrease the loss (by escaping a"functional"43 saddlepoint).
Neural Information Processing Systems
Feb-19-2026, 11:59:15 GMT