Abstract
Previous work on visual category learning has demonstrated that the order of practice trials influences learning outcomes. For example, Carvalho and Goldstone (Memory and Cognition, 2014) provide evidence that a blocking policy results in better learning when the categories are relatively dissimilar, whereas an interleaving policy is superior when the categories are relatively similar. The primary objective of this work is to move beyond heuristic scheduling policies by using human-surrogate models to optimize training sequences. To that end, we must first identify a human-surrogate model that correctly predicts the effect of trial order. To discriminate among various candidate models, we used behavior data collected from a category learning experiment where subjects were randomly assigned to one of three heuristic scheduling policies: blocked, interleaved and nested (a hybrid between blocked and interleaved). The behavioral results of the experiment replicate previous findings demonstrating the importance of trial order. Each model was evaluated using cross-validation, revealing the extent to which the model could generalize to unseen data. Results of the model fitting procedure revealed that commonly used cognitive models are either insensitive to the order of practice trials or not flexible enough to model the empirical pattern of results. In contrast, using a long short-term memory (LSTM) recurrent neural network yields good fits. We further tested the generalization ability of the LSTM model by removing data belonging to one condition and testing it on unseen sequences from all conditions. The LSTM model correctly predicts the order of conditions, except when data from the blocked condition is removed. Taken together, these results suggest that an LSTM model trained on an appropriate span of behavioral data is a promising human-surrogate model.
Meeting abstract presented at VSS 2018