Human and animal learning shows many perplexing behaviors that appear (from some perspectives) to be downright irrational. I will describe how many of these behaviors can be explained as a rational solution to the problem of learning quickly. The explanation relies on the assumption that attention is selectively re-allocated such that error is reduced. These shifts of attention reduce error and accelerate learning, but the shifts of attention also distort the encoding of the stimuli and consequently generate "distorted" behaviors later. When the principle of attention shifting is rigorously implemented in connectionist architectures, detailed quantitative fits to human performance are achieved. I will describe a new comprehensive architecture that unifies several previous models.