| dc.description |
Cognitive psychologists have developed many formal models of categorization, but they have been almost exclusively tested using artificial categories because deriving psychological representations of natural stimuli using traditional methods such as multidimensional scaling (MDS) has been an intractable task. In this dissertation, I show that deep convolutional neural networks (CNNs) may be used to solve this problem. First, I provide an overview of how CNNs work, and I review related work that has examined the relationship between the representations learned by CNNs and the psychological representations used by humans. I then demonstrate that CNNs can be trained to predict the MDS coordinates of rocks derived in previous work (Nosofsky, Sanders, Meagher, & Douglas, 2017). In Experiment 1, I conduct a conceptual replication of Nosofsky et al.’s (2017) methods and demonstrate that similar MDS dimensions emerge across different sets of rocks, and the CNNs are able to generalize from one set to the other. Then in Experiment 2, I conduct a categorization experiment and demonstrate that the CNN representations can be used in conjunction with a formal cognitive model to predict human behavior, indicating that CNNs can be used to automate MDS studies in the future. |
|