Abstract
Perceptual learning (PL) is often due to observers becoming more sensitive to the diagnostic information in the specific stimuli used during training. Here, we asked how diagnostic information affects PL in a texture identification task. Specifically, we examined whether learning and stimulus generalization differed when diagnostic information was presented alone or embedded in an uninformative context. Texture stimuli were generated from low-frequency band-pass filtered white noise; to create stimuli with restricted diagnostic information, we applied a ±45 degree orientation filter centered on 0 (HORZ) or 90 (VERT) degrees. Explicit stimuli comprised HORZ and VERT orientation-filtered textures as well as an unfiltered FULL texture. Embedded stimuli comprised HORZ and VERT filtered textures summed with non-informative complementary orientation information from the average of all our texture stimuli. We measured accuracy in a 1-of-6 identification task: Observers were trained with either HORZ-explicit or HORZ-embedded stimuli, and tested with all stimulus types. When trained with HORZ-embedded stimuli, performance improvements could result from: i) increased reliance on the informative horizontal band, manifesting as improved accuracy for all conditions in which HORZ structure is informative, or ii) decreased reliance on the uninformative vertical band, manifesting as reduced accuracy for both VERT-embedded and VERT-explicit stimuli. When trained with HORZ-explicit stimuli, performance improvements can only result from increased sensitivity to horizontal structure. Our results reveal higher pre-training accuracy with explicit than embedded stimuli, demonstrating that uninformative structure impaired identification. Training with HORZ-embedded stimuli eliminated this effect, producing post-training accuracy similar to pre-training explicit stimuli. Further, accuracy for VERT stimuli was not decreased, suggesting that learning was not at a cost to the uninformative band. Finally, embedded and explicit learning were largely stimulus-specific, with minimal transfer to other conditions. Together, these results support an increased reliance on the trained orientation band, without suppression of non-informative orientations, but high context-specificity.
Meeting abstract presented at VSS 2015