Abstract
Previous work on visual category training has demonstrated that the sequence of training trials can influence learning outcomes. Studies have compared blocked presentations (multiple examples of the same category in a row) to interleaved presentation (switching between categories from trial to trial). For example, Carvalho and Goldstone (2014, Memory & Cognition) found that when categories are relatively dissimilar, blocking yields better generalization to novel exemplars, but when categories are relatively similar, interleaving is superior. Carvalho and Goldstone propose that learners focus on commonalities when two consecutive items are of the same category and on differences when two consecutive items belong to different categories. Leveraging their hypothesis, we develop a simple parameter-free probabilistic model of attention to stimulus dimensions. When presented with a sequence of experimental stimuli encoded as feature vectors, the model indicates the degree to which the discriminative features will be discovered. The model correctly predicts the empirical rank order of generalization performance across various experimental conditions. Using this model as a surrogate for a human learner, we search for stimulus sequences (schedules) that optimize learning. We consider schedules produced by a generalization of a Lévy flight—often used to model foraging—operating in the stimulus representational space. By explicitly utilizing category structure, our scheduling procedure is able to generate structure-sensitive sequences. Tuning the two free parameters of this scheduling procedure to optimize the performance of our surrogate, we show that the best schedule for dissimilar categories is blocking and for similar categories is interleaving. With this flexible framework, we can optimize training sequences for complex, hierarchical, and heterogeneous category structures. We report on experimental studies to validate the computational model and scheduling procedure.
Meeting abstract presented at VSS 2016