Visual statistical learning is facilitated in Zipfian distributions.

Area of Science:

• Cognitive Psychology
• Developmental Psychology
• Machine Learning

Background:

• Statistical learning (SL) explains how humans learn environmental regularities.
• Previous studies often used uniform frequency distributions, where unit appearance is unpredictable.
• Real-world environments, like language and visual scenes, exhibit non-uniform (skewed) distributions.

Purpose of the Study:

• To investigate if visual statistical learning is facilitated by skewed (Zipfian) distributions, similar to findings in auditory learning.
• To examine the domain-generality and age-appropriateness of this effect.
• To determine if uniform distributions underestimate learning performance.

Main Methods:

• Analyzed an existing database to confirm skewed distributions in children's object co-occurrence data.
• Designed experiments presenting visual triplets in both uniform and skewed (Zipfian) distributions.
• Assessed learning performance in children and adults across different distribution types and frequency levels.

Main Results:

• Object combinations in children's environments naturally follow a skewed distribution.
• Both children and adults demonstrated superior learning performance in the Zipfian distribution compared to the uniform distribution.
• Learning benefits were particularly pronounced for low-frequency visual triplets in the skewed condition.

Conclusions:

• Skewed (Zipfian) distributions facilitate statistical learning across different modalities (visual, auditory) and age groups.
• The use of uniform distributions in research may lead to an underestimation of learning capabilities.
• Real-world environments with skewed distributions may offer a learnability advantage.