Identifying Accurate and Inaccurate Stimulus Relations: Human and Computer Learning

Abstract

In three experiments, we explore human and simulated participants’ potential for deriving and merging analogous forms of stimulus relations. In the first experiment, five human participants were exposed to compound stimuli (stimulus pairs) by way of an automated yes–no protocol. Participants received discrimination training focusing on four three-member stimulus classes, where only two of the four classes were correctly related algebraic expressions. Training was intended to establish generalized identification of novel correct stimulus pairs and generalized identification of novel incorrect stimulus pairs. In Experiment 2, we employed a three-layer connectionist model (CM) of a yes–no protocol aimed at training and testing an analogous set of stimulus relations. Our procedures were aimed at assessing a neural network’s ability to simulate derived stimulus relations consistent with the human performances observed in Experiment 1. In Experiment 3, we employed a four-layer CM to compute the number of training epochs required to attain mastery. As with our human participants, our neural network required specific training procedures to become proficient in identifying stimuli as being members or nonmembers of specific classes. Outcomes from Experiment 3 suggest that the number of training epochs required to attain mastery for our simulated participants corresponded closely with the number of training trials required of our human participants during Experiment 1. Moreover, generalization tests revealed that human and simulated participants exhibited analogous response patterns. We discuss the evolving potential for CMs to emulate and predict human training requirements for deriving and merging complex stimulus relations during generalization tests.