Schlüsselworte: generalization theory risk distance sampling models theoretical framework bregman stability minimization problem network weights estimating data learning deep empirical approximation solution propagation regularization layer estimate training optimality parameters basic