Abstract
Methods of machine learning, especially from Statistical Learning Theory (SLT), yield excellent results. We wondered to what extent these theoretical tools are applicable to human Perceptual Learning (PL). Standard SLT (Vapnik, 2000), allows learning of the optimal decision rule, without learning the input information (data). We show two examples where the human brain can learn without learning the decision rule. First, we trained subjects in a 2AFC homogeneous-versus-non-homogeneous motion task with 34 segments. After extensive training, subjects couldn't perform the task. Hence, they never learned the optimal decision rule. However, training resulted in significant improvement for tasks with two segments. In the second example, subjects trained and learned a homogeneous-versus-non-homogeneous task with varying edge position across stimuli. Improvement in a Vernier task using a similar display was also seen. The two tasks have different optimal decision rules, that if learned should result in improvement of the Vernier task with further training, which is not the case.
In conclusion, mechanisms of PL in humans do not follow standard SLT. They involve forms of learning that are not limited to the optimization of the decision rule. Our data suggest that the brain may be learning the statistics of the input information, which allow improved performance in tasks different from the ones trained on. Alternatively, the brain may be tuning its noisy processing stages to deal with the inputs, regardless of the decisions. We have proposed a generalization of SLT that allows for these alternative forms of learning (Grzywacz and Padilla, 2006).