Your reward system is hard coded. So you pay attention to the sensory data correlates that result in the highest preference satisfaction. This gets compounded until you've bootstrapped up to having a reasonable causal model of the things that affect you the most.
