Learning algorithms generally need significant prior information on the task they are attempting to solve. This requirement limits their flexibility and forces engineers to provide appropriate priors at design time. A question then arises naturally: how can we reduce the amount of prior task information needed by the algorithm? Being able to answer this question could spark the development of algorithms with higher generalization potential, all the while reducing preliminary engineering efforts. If we want to discard any a priori knowledge about the task, we need the learning algorithm to efficiently explore and represent the space of possible outcomes it can achieve.