Transfer and Learning to Learn in Perceptual Learning

As there is considerable current interest in the training characteristics that produce nonspecific perceptual learning, we propose that it may be useful to differentiate between “transfer” and “learning to learn.” These two constructs emerge from learning at different levels of a hierarchical Bayesian model. At the lowest level of the hierarchy is the individual learning task, where the subject is typically asked to estimate the probability of a state given data (e.g., the probability that the answer is “clockwise” given an oriented gabor). By Bayes' rule, improving this estimate requires learning the probability of the data given the states (i.e. the likelihood). Tasks to which learning at this level should “transfer” are those that utilize the same likelihood as that which is learned during training (and thus requires that the tasks share a common state representation). Consistent with this viewpoint we have shown transfer of learning across orientation that is dependent on the state representation inherent in the training task. “Learning to learn” on the other hand requires that learning occur at levels above the individual task. By being exposed to multiple individual learning tasks, subjects can learn the manner in which likelihoods across tasks are generated. While subjects who have learned at this level may show weak immediate transfer effects, they should learn individual tasks more quickly. We have recently suggested that learning at this level is responsible for the broad range of tasks in which enhancements are noted as a result of action video game play