Bayesian filtering, parallel hypotheses and uncertainty: a new, combined model for human learning