Using deep learning to predict human decisions and using cognitive models to explain deep learning models.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) offer high capacity for modeling complex data.
• The opaque nature of DNNs limits their explanatory power in scientific research.
• Understanding human decision-making is crucial for cognitive science.

Purpose of the Study:

• To propose and validate a method for using DNNs as exploratory tools in cognitive science.
• To characterize DNN behavior using explicit, theory-driven models.
• To enhance the interpretability of DNNs for scientific investigation.

Main Methods:

• Trained an exploratory DNN to predict human decisions in a four-armed bandit task.
• Compared DNN accuracy against explicit reward-oriented and reward-oblivious models.
• Used experimental simulations to characterize the exploratory DNN using explicit models.

Main Results:

• The exploratory DNN model demonstrated higher accuracy than explicit models in predicting human decisions.
• The DNN model's predictions aligned with reward-oriented models when options were clearly differentiated.
• The DNN model exhibited pattern-based exploration similar to reward-oblivious models.

Conclusions:

• Predictable, non-reward-oriented patterns likely influence human decision-making.
• Theory-driven models can effectively characterize DNN operations, enhancing their utility.
• This hybrid approach makes DNNs valuable explanatory tools for cognitive processes.