A Bayesian observer model constrained by efficient coding can explain 'anti-Bayesian' percepts.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Psychophysics

Background:

• Bayesian observer models explain perception by assuming percepts are biased toward prior beliefs.
• Existing psychophysical data suggest this standard explanation may be overly simplistic.
• A gap exists in understanding how perception deviates from physical reality beyond simple prior biases.

Purpose of the Study:

• To propose a novel perceptual model based on efficient coding principles.
• To investigate if perception can be biased away from prior beliefs.
• To examine how stimulus uncertainty affects perceptual bias differently based on noise source.

Main Methods:

• Developed a new efficient coding model fully specified for natural stimulus distributions.
• Derived predictions regarding perceptual bias and the influence of stimulus uncertainty.
• Compared model predictions against existing psychophysical data for visual orientation and spatial frequency.

Main Results:

• The model predicts perception can be biased away from an observer's prior beliefs.
• It predicts differential effects of stimulus uncertainty on perceptual bias based on noise type (internal vs. external).
• These predictions align with reported perceptual biases in visual orientation and spatial frequency perception.

Conclusions:

• The proposed efficient coding model offers a more nuanced account of perceptual biases.
• It successfully explains psychophysical data not accounted for by standard Bayesian models.
• The model's generality suggests broad applicability to various perceptual variables and tasks.