Ironically, you could use the sigmoid function instead of horses. The training stimulus slowly builds over multiple iteration and then suddenly, flip: the wrong prediction reverses.