Causal Model Comparison Shows That Human Representation Learning Is Not Bayesian.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Machine Learning

Background:

• Understanding how humans learn relevant environmental features for tasks is crucial for cognitive science.
• Representation learning, the process of identifying salient information, remains computationally complex.
• Existing models often struggle to capture the nuances of human learning strategies.

Purpose of the Study:

• To propose and validate a novel causal model comparison method for studying representation learning.
• To investigate the computational mechanisms underlying human feature selection in a probabilistic task.
• To compare the predictive and perturbational power of different computational models of human learning.

Main Methods:

• A probabilistic learning task was designed where participants identified relevant features.
• Two computational models, a reinforcement-learning model and a Bayesian ideal observer model, were run in parallel with participants.
• Participant learning was perturbed in real-time using model predictions to assess model validity.

Main Results:

• The reinforcement-learning model successfully predicted and influenced participants' learning, both facilitating and hindering it.
• The Bayesian ideal observer model, while theoretically optimal, failed to perturb or accurately predict human learning.
• The causal model comparison method demonstrated sensitivity in distinguishing between models of human representation learning.

Conclusions:

• Human representation learning appears to rely on computationally tractable, albeit suboptimal, mechanisms like reinforcement learning.
• The proposed causal model comparison method offers a sensitive tool for evaluating cognitive models by directly manipulating learning.
• This approach provides new insights into the dynamic and adaptive nature of human learning processes.