Serial Dependence Predicts Generalization in Perceptual Learning

Recent visual experiences systematically bias current perceptual reports, a phenomenon known as serial dependence. This study examines how serial dependence interacts with perceptual learning, specifically focusing on its role in learning generalization. We reanalyzed data from 50 observers who practiced the texture discrimination task (TDT) under three conditions designed to modulate learning generalization: consistent target location, alternating two target locations, and mixed with targetless trials (Harris et al., 2012); over 200,000 trials. Our analysis revealed substantial report biases toward the orientation of targets from previous trials, extending up to 10 trials back, much longer than previously reported. These biases, when accumulated across consistent trial sequences, nearly overshadowed responses to the current target. Remarkably, although TDT thresholds improved by 40% over the 8 days of training, the biases persisted despite the randomization of target orientations, with trial history being irrelevant to the current trial. The effect was most pronounced when current targets had low visibility, prior targets were highly visible, and prior targets appeared at the same location as the current target. Training conditions that facilitated learning generalization exhibited stronger and longer serial dependence effects from distant trial histories compared to conditions promoting specific learning. Across observers, stronger serial dependence from distant history was associated with greater learning transfer, suggesting that distant memory traces are essential for efficient learning. We propose that these extended memory traces integrate information across multiple trials to create a generalized learning template, preventing overfitting to specific stimuli thus enhancing learning transfer. In contrast, strong adaptation in the consistent target location condition disrupts these traces, thereby limiting generalization.