Related Experiment Videos
Modeling time to visual insight in Mooney image recognition with a chaotic recurrent neural network
Misako Kimura1,2,3, Yuuki Matsushita4, Masayo Inoue5
1Center for Information and Neural Networks, Advanced ICT Research Institute, National Institute of Information and Communications Technology, 1-4 Yamadaoka, Suita, Osaka 565-0871 Japan.
Cognitive Neurodynamics
|June 8, 2026
Summary
This study introduces a novel neural network model that simulates visual insight emergence. The model successfully replicates human search time distributions, offering a new computational framework for understanding insight.
Area of Science:
- Cognitive Science
- Computational Neuroscience
- Artificial Intelligence
Background:
- Insight involves sudden conceptual shifts for problem-solving, often beyond analytical methods.
- Previous computational models of insight, including deep neural networks (DNNs) and reinforcement learning, have limitations in capturing autonomous neural computation dynamics.
- Few studies have modeled the dynamic emergence of insight through autonomous neural computation.
Purpose of the Study:
- To present a neural network model simulating the time to achieve visual insight in the Mooney image recognition task.
- To capture the dynamic emergence of insight through autonomous neural computation.
- To provide a computational framework for implementing insight in artificial systems.
Main Methods:
- Coupled a DNN for feature extraction with a recurrent neural network (RNN) implementing a chaotic search process.
- Formulated the RNN as a continuous-time dynamical system for autonomous state exploration and internal reconstruction of features.
- Utilized the same image set as human psychophysical experiments for model validation.
Main Results:
- The model reproduced key statistical properties of human search times (STs), including lognormal-like ST distributions, a proportional relationship between log-scale mean and standard deviation, and discrete levels of mean ST across images.
- These properties emerged without assuming lognormal distributions for participant variability, unlike previous models.
- Demonstrated that lognormal-like signatures can arise from exponential search dynamics combined with experimental preprocessing and finite observation windows.
Conclusions:
- The findings support a generative dynamical account linking intrinsic chaotic dynamics to insight-related search.
- The study provides a computational framework for insight generation in artificial systems.
- Highlights the importance of distinguishing generative processes from measurement and analysis effects in computational modeling.
