Tag: Learning with Non-Convex Truncated Losses by SGD