Imperfect Bayesian inference in visual perception

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Psychology

Background:

• Optimal Bayesian models successfully describe human perceptual decision-making.
• Concerns exist regarding model flexibility and the plausibility of optimal neural computation.

Purpose of the Study:

• To investigate human performance on a visual search task using constrained models and testing for computational imperfections.
• To evaluate whether human performance aligns with flawless or imperfect Bayesian inference.

Main Methods:

• Subjects performed a target detection task with varying external uncertainty (distribution overlap) and sensory noise (display time).
• Empirical performance was measured and compared against optimal Bayesian models and heuristic models.

Main Results:

• Human performance was suboptimal, falling 18.1% short of optimal Bayesian predictions on average.
• Suboptimality was not significantly affected by internal or external uncertainty levels.
• An imperfect Bayesian model provided a better account of the data than a flawless Bayesian model or heuristic models.

Conclusions:

• Human perception is likely based on Bayesian principles but implemented with computational imperfections.
• The imperfect Bayesian inference framework offers a balanced perspective between traditional and anti-Bayesian views of perception.