Abstract
Classification training typically involves viewing a series of category examples, making a classification response to each, and receiving corrective feedback regarding category membership. Objective feedback (i.e., based on actual category membership) suggests that perfect accuracy is possible even when it may not be (e.g., overlap exists between categories). Previous research has shown this type of feedback can be detrimental to learning an optimal (long-run reward-maximizing) decision criterion by fostering excessive attention to trial-to-trial accuracy. Some accuracy must be sacrificed to maximize long-run reward (Bohil, Wismer, Schiebel, & Williams, 2015; Bohil & Maddox, 2003). Thus, it is important to consider other types of feedback for training, such as using the responses of an "optimal" performer to create feedback that indicates that using even the optimal response criterion produces occasional response errors. In the current study, normal or cancer-containing mammograms were used to assess how feedback influences classification. Participants earned more points for correct "cancer" than correct "normal" responses. Feedback was given in one of two forms: objective (based on actual category membership) or based on a "best" classifier (i.e., the responses of the nearest-optimal performer from an earlier classification study). Critically, the performance of an optimal or "best" classifier indicates that even when using the best possible classification criterion, errors should be expected, as opposed to objective feedback which implies 100% accuracy may be possible when it is not. Signal detection analyses indicated decision criterion values that were closer to optimal in the best-classifier condition. Participants trained with best-classifier feedback also had higher point totals and a reduction (as predicted) in overall response accuracy compared to participants trained with objective feedback. This work replicates earlier research using simple artificial stimuli, and shows that feedback reflecting a more-attainable performance level supports more optimal decision criterion placement and performance.
Meeting abstract presented at VSS 2016