Abstract
Classification image analysis is a psychophysical technique in which the noise components of signal+noise stimuli are analyzed to produce an image that reveals the critical features of a visual task. Here we examine classification images during the time course of a perceptual learning task to gain a greater understanding of what subjects learn through training on detection of oriented gratings in noise. To do this, we optimized standard classification image procedures by using designer (m-sequence) noise and a relatively low-dimensional stimulus space, so that we could achieve reasonable classification images within a single thousand-trial session. Subjects were trained across ten sessions to detect the orientation of a grating masked in noise, with an eleventh test session conducted using an oriented stimulus orthogonal to the trained stimuli. Subjects showed improvement in performance metrics such as reaction time and signal detection threshold. Clarity of the classification images and their correlation to an ideal target was also observed to improve across training sessions. These improvements showed only partial transfer to the orthogonal test session, indicating an orientation-specific learning effect. The main benefit of classification image techniques is that they allow for a variety of image-based analyses. We discuss how orientation tuning curves can be derived from the classification images and how these change through the time course of training, how individual correlation measures change for bright and dark components of the stimuli, and how per-trial feedback affects various metrics of learning. Our results shed insight on what is learned during orientation discrimination and demonstrate that classification image techniques are a promising method by which to study perceptual learning.