AI-Generated Cognitive Models from Behavioral Data

This piece, by Onno Berkan, was published on 03/14/25. The original text, by (Kevin) Miller et al., was submitted to NeurIPS 2023

This UCL study introduces a method for creating cognitive models from behavioral data. The traditional approach requires researchers to manually propose and test different models of how the brain might work. This approach is limiting since researchers might miss better explanations they haven't considered.

The researchers solved this through a "disentangled RNN" (DisRNN) that can automatically discover simple, interpretable, and parsimonious (ie, cheap) models of cognitive processes from behavioral data. What makes this system unique is that it's designed to find the simplest possible explanation that accurately describes the observed behavior.

The DisRNN successfully recovered the correct cognitive mechanisms when tested on synthetic data where they knew the true underlying decision-making processes. For example, when analyzing data from simulated agents performing reward-learning tasks, the system identified these agents' exact learning rules. This validated that the method could reliably discover true cognitive mechanisms when they exist.

Afterwards, the researchers applied their system to real behavioral data from laboratory rats performing two tasks. In a reward-learning task, rats had to choose between two options with changing reward probabilities. The DisRNN discovered models that were just as good at explaining the rats' behavior as the best human-designed models. These automatically discovered models were relatively simple and easy for researchers to interpret. 

One exciting feature of DisRNN is its flexibility. Researchers can adjust how much the system prioritizes simplicity versus accuracy. This allows them to explore different possible explanations, from very simple models that capture the main patterns to more complex models that account for subtle details in the behavior. However, this requires considerable data, making this method not the most efficient. 

Want to submit a piece? Or trying to write a piece and struggling? Check out the guides here!

Thank you for reading. Reminder: Byte Sized is open to everyone! Feel free to submit your piece. Please read the guides first though.

All submissions to berkan@usc.edu with the header “Byte Sized Submission” in Word Doc format please. Thank you!