Abstract
Given appropriate training, humans will demonstrate improvement on virtually any perceptual task. However, the learning that occurs is typically highly specific to the training task and stimuli. In the language of machine learning, such specificity is indicative of what is known as policy learning. Policies can be thought of simply as lookup tables that map states onto actions (i.e.-“what to do”). Importantly, policies are specific to a given goal; if the goal is changed, knowing what was previously the right thing to do provides no information regarding what is currently the right thing to do. As an example, in a typical orientation discrimination task (“Was the gabor tilted clockwise or counterclockwise from a reference angle?”), the optimal policy relies on a discriminate. If the current “state” lies on one side of the discriminate, press ‘A’, otherwise, press ‘B’. Given this, it is clear why transfer is not observed – this policy completely inapplicable when the reference angle is rotated by 90°. If perceptual learning is analogous to policy learning, we hypothesized that in order to observe transfer, the training task must promote the development of a policy that can be extrapolated from and will be appropriate for new orientations. To this end, rather than a discrimination task, which promotes the development of an untransferable policy, an orientation estimation task was employed. Subjects were asked to indicate (by rotating a single line) the exact orientation of a quickly flashed gabor (+/-15° from 45°). The policy that should be learned in this task is a continuous function of orientation and thus it should be possible to extrapolate to previously unseen orientations. As predicted, full transfer was observed when the stimuli were rotated by 90°. These results and overall framework provide a novel way of approaching the field of perceptual learning.