To get increased computational power from multiple neurons, need nonlinear activation functions. Examples:
Good: can represent more complex functions
Bad: harder to train