AnAlternativeView:WhenDoesSGDEscapeLocalMinima?RobertKleinberg1YuanzhiLi2YangYuan1Abstractmoreiterationstoconverge,butfewergradientevaluationsperiteration.Therefore,forthestandardempiricalriskmin-S...
HowtoEscapeSaddlePointsEfficientlyChiJin1RongGe2PraneethNetrapalli3ShamM.Kakade4MichaelI.Jordan1Abstractapointwithsmallgradientisindependentofthedimension(“dimension-free”).Moreprecisely,forafunc...
