Modeling Time to Visual Insight in Mooney Image Recognition with a Chaotic Recurrent Neural Network
Insight is often described as a sudden shift in or formation of a conceptual representation, enabling humans to restructure existing knowledge and solve problems beyond conventional analytical approaches. Although prior computational studies have modeled aspects of insight using deep neural networks (DNNs) or reinforcement learning, few have captured the dynamic emergence of insight through autonomous neural computation. Here, we present a neural network model that simulates the time required to reach visual insight in the Mooney image recognition task, a widely used paradigm for studying visual insight and perceptual reorganization. The model couples a DNN module for perceptual feature extraction with a recurrent neural network (RNN) that implements a chaotic search process for recognition. The RNN is formulated as a continuous-time dynamical system, autonomously explores internal states, and stabilizes when the missing visual features required for recognition are internally reconstructed. Using the same image set as in human psychophysical experiments, the model reproduces key statistical properties of human search times (STs), including (i) lognormal-like ST distributions across participants for each image, (ii) a proportional relationship between the log-scale mean and standard deviation estimated from lognormal fits across images, and (iii) discrete levels of the fitted log-scale mean across images (a proxy for image difficulty). Importantly, these properties emerge without assuming any lognormal distribution for participant-to-participant variability, whereas previous models reproduced similar signatures by positing lognormal-distributed individual differences. We further show that lognormal-like signatures can arise from exponential search dynamics when combined with both standard experimental preprocessing and finite observation windows, highlighting the need to distinguish generative mechanisms from measurement and analysis effects. Together, these results motivate a mechanistic link between chaotic neural dynamics and insight-related search and provide a computational framework for implementing insight in artificial systems.