Learning and inference using complex generative models in a spatial localization task

Area of Science:

• Cognitive science
• Neuroscience
• Computational modeling

Background:

• Human observers integrate uncertain sensory data with prior knowledge near-Bayes-optimally in simple tasks.
• Natural tasks often involve complex generative models with multiple causes, posing challenges for integration.

Purpose of the Study:

• To investigate if Bayes-optimal integration extends to complex, multi-cause generative models.
• To determine if humans use heuristics or near-optimal strategies in complex environments.

Main Methods:

• Participants localized a hidden target sampled from a bimodal generative model with varying variances.
• Repeated exposure allowed learning of the a priori bimodal model.
• Trial-by-trial analysis assessed integration of sensory information with learned priors.

Main Results:

• Participants learned the bimodal generative model's mode locations rapidly.
• Learning the relative variances of the modes occurred more slowly.
• Integration of sensory information with learned priors was consistent with Bayes-optimal predictions.

Conclusions:

• Human performance in complex localization tasks aligns with Bayes-optimal behavior.
• Learning complex generative models, like bimodal distributions, has a longer time-course than simpler models.
• Bayes-optimal integration is maintained even when dealing with more complex environmental structures.