Implicit Bias of the Step Size in Linear Diagonal Neural Networks