generalization of SGD in SCO is well-established, and we are left with the question of how well can we account for generalization through an investigation of its bias