For small training sets variance may contribute more to the overall error than bias. sometimes this is handled by reducing the complexity of the model, even if the model is too simple. why do you suppose this is the case? come up with your own example of this.