Mutual Information and Categorical Perception.

Area of Science:

• Cognitive Psychology
• Perception Science
• Machine Learning

Background:

• Categorical perception enhances sensitivity near category boundaries, particularly for informative dimensions.
• The precise reasons and dimensions driving this phenomenon remain unclear.
• Understanding this is key to explaining how the brain processes categories.

Purpose of the Study:

• To investigate which dimensions are treated as informative in category learning.
• To determine the underlying principles governing categorical perception.
• To test if perceptual discrimination is tuned to statistical environmental structure.

Main Methods:

• Subjects learned statistically defined categories in a novel 2D shape space.
• Perceptual discrimination of features was tested before and after category learning.
• Features varied in position and orientation relative to the most informative dimension.

Main Results:

• Improvement in perceptual discrimination was measured for each feature.
• The magnitude of improvement directly correlated with feature informativeness.
• A simple generalization emerged linking discrimination enhancement to mutual information.

Conclusions:

• Perceptual discrimination precision is rationally tuned to the statistical structure of the environment.
• Categorical perception is guided by the mutual information between features and categories.
• This provides a framework for understanding rational perception in novel environments.