Abstract
Perceptual learning is often slowed or eliminated when task variants are intermixed —so called task roving. We (Dosher et al., 2000) recently showed dramatically slower learning when four different orientation discrimination tasks were intermixed (+/-12deg from 4 separated axes), relative to training either with identical or sufficiently different axes in all four locations. A computational framework, the IRT, successfully modeled the pattern of results. A similar study (Xie & Yu, 2020) also intermixed training of four orientation discrimination tasks in twelve locations. Structurally, the two studies are similar—but here robust learning occurred, with transfer to an untrained task-and-location combination. We suggest that the key contrast reflects the use of a two-interval design vs the more commonly used single interval identification design. (The tasks also differ in using orientation difference thresholds versus contrast thresholds.). In this study, we create a new extension of the IRT for the two-interval design and provide an excellent fit to the four training conditions of Xie and Yu: learning where the locations/tasks shift predictably from trial to trial, where the locations/tasks are randomized over trials, where the training location/task is the same as in pre- and post- tests, and a control condition that measures only transfer without training, reproducing both the empirical learning and transfer effects. In the new IRT model extension, the activations in four mini-decision units (driven by connections from location-specific and location-invariant orientation tuned representations) from the first interval are stored and compared to those in the second interval. Capitalizing on these two-interval comparisons avoids the roving disruption; location-invariant learning from multiple locations enables transfer. We propose the new IRT as a computational framework for perceptual learning in two-interval designs that can be put to additional empirical tests to understand the differences or similarities in perceptual learning in single- and two-interval tasks.