ce016f59ecc2366a43e1c96a4774d167-AuthorFeedback.pdf

Empirically,alargeN issuitable for datasets with largeintra-class variance while asmall20 sub-graph (withr = 0.05)is optimal in most cases. Incontrast, our method only needs tocalculate and update the gradients ofd0.05 We will follow the suggestion to indicate the backbones used by other33 methods.