On the origins of suboptimality in human probabilistic inference

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Decision Theory

Background:

• Humans integrate noisy sensory data with prior experience, approximating Bayesian principles.
• Suboptimal performance arises with complex priors (e.g., skewed, bimodal), but the cause—imprecise representation or computational constraints—remains unclear.

Purpose of the Study:

• To investigate the sources of suboptimality in human probabilistic inference with complex priors.
• To differentiate between challenges in prior acquisition versus probabilistic computation.

Main Methods:

• A novel estimation task presenting explicit, trial-varying prior distributions (Gaussian, unimodal, bimodal).
• Subjects estimated target location using noisy cues and visual prior densities.
• Factorial model comparison of Bayesian observer models to identify noise sources.

Main Results:

• Performance qualitatively matched Bayesian Decision Theory but was suboptimal.
• Suboptimality varied with prior statistics but was independent of prior class or cue noise.
• Response variability primarily driven by noisy prior parameter estimation and stochastic posterior inference.

Conclusions:

• Suboptimal probabilistic inference with complex priors may be linked to difficulties in acquiring these priors, not necessarily in computation.
• Human limitations in probabilistic inference involve both noisy estimation of statistical parameters and stochasticity in the decision process.