Using a Deep Learning-Based Visual Computational Model to Identify Cognitive Strategies in Matrix Reasoning

Abstract

Constructive matching and response elimination strategies are two primarily used cognitive strategies in Raven’s Advanced Progressive Matrices (APM), a valid measurement instrument of general intelligence. Identifying strategies is necessary for conducting studies on the relationship between cognitive strategy and other cognitive factors and for cognitive strategy training. However, the strategy identification method used in research is either subjective, or the information in the behavior data is not fully utilized, or it is limited by the size of the sample and cannot be widely used. Therefore, this study trained a convolutional neural network-based visual computational model (CVC) for cognitive strategy identification based on eye movement images. Focusing on the APM, the trained CVC can be used for strategy identification by learning and mining the pattern information in the eye movement images with predefined training labels from a psychometric model. An empirical study was conducted to illustrate the training and application of the CVC. Utilizing the trained CVC and a developed graphical user interface application, the primary finding of the study reveals a high level of agreement in strategy identification between the CVC and the psychometric model, as well as between the CVC and expert judgment. This implies that, akin to the psychometric model, the CVC can be used to identify the two cognitive strategies of constructive matching and response elimination. Overall, the proposed deep learning-based model follows the data-driven perspective and provides a new way of studying cognitive strategy in the APM by presenting objective and quantitative identification results.