Abstract
Visual perceptual learning has, with few exceptions, been investigated in the context of two-alternative forced (2AFC) choice tasks. In this study, we trained observers with an 8AFC orientation task in external noise with two forms of feedback: response feedback (RF) in which feedback indicates the correct response and accuracy feedback (AF) in which feedback indicates whether the response is accurate. Perceptual learning in such tasks may more closely approximate the complexity of real world tasks. We elaborated and extended a computational model of nAFC tasks based on the Integrated Reweighting Theory (IRT, Dosher et al., 2013) by using multiple decision units and a max rule. Generally, the elaborated IRT predicts better performance for RF than AF feedback under a range of conditions. Performance was measured over the course of practice at three contrast levels (0.3, 0.6, 1.0) for Gabors of eight orientations separated by 22.5°. Performance measures included percent correct, weighted k (partial credit for adjacent responses), and confusion matrices. Four of five observers receiving RF and three of five observers receiving AF showed significant learning (percent correct, k) over eight sessions of 960 trials each. Two additional AF observers who appeared not to learn dropped after a few sessions. Subsequent training of a few observers without external noise provided an estimate of asymptotic accuracy. Two AF observers who failed to learn showed lower asymptotic noiseless accuracies, suggesting either broader bandwidth of orientation tuning or higher levels of internal noise. The confusion matrices provide further diagnostic information about stereotyped response biases from broad errors clustering around the correct diagonal, and also show how the weights on evidence grow during learning. These results are consistent with the predictions from the nAFC IRT model that performance is improved more with RF than AF.
Meeting abstract presented at VSS 2017