Bayesian estimation yields anti-Weber variability

Area of Science:

• Cognitive Psychology
• Psychophysics
• Computational Neuroscience

Background:

• Human perception often exhibits Weber behavior, where response variability increases with stimulus magnitude.
• The prominent Bayesian paradigm in perception has not extensively focused on Weber behavior.
• Previous models often assumed linear encoding, failing to capture certain perceptual patterns.

Purpose of the Study:

• To investigate Bayesian observer variability in comparison to human subjects.
• To examine how manipulating prior distributions and reward functions affect perceptual variability.
• To reconcile observed perceptual patterns with established psychophysical principles and Bayesian frameworks.

Main Methods:

• Conducted two preregistered experiments involving a numerosity-estimation task.
• Manipulated the prior probability distribution of numerosities.
• Altered the reward function associated with accurate estimations.
• Compared human subject responses to predictions from a Bayesian observer model.

Main Results:

• Bayesian observers exhibited "anti-Weber behavior" (less variability for larger magnitudes) when large numerosities were more frequent or rewarding.
• Human subjects demonstrated a similar anti-Weber behavior, contradicting classic psychophysical findings.
• Subject responses were best explained by a logarithmic encoding of magnitudes, aligning with Fechner's proposals.

Conclusions:

• Human perception can exhibit anti-Weber behavior, challenging long-standing psychophysical results.
• Bayesian models, under specific conditions (e.g., skewed priors), can replicate this anti-Weber behavior.
• Logarithmic encoding, previously associated with Weber behavior, is compatible with anti-Weber behavior in this context.
• The skewness of natural priors is suggested as a primary driver for increased variability in perception.