On the implicit bias of SGD in deep learning - World Wide