Perceptual learning and representational learning in humans and animals

Area of Science:

• Cognitive Neuroscience
• Comparative Psychology
• Machine Learning

Background:

• Perceptual learning (human) and classical conditioning (animal) are traditionally distinct research fields.
• Representational learning offers a unifying framework to explore commonalities.
• Understanding how internal representations form is key to linking these domains.

Purpose of the Study:

• To demonstrate common themes between human perceptual learning and animal classical conditioning.
• To explore the emergence of internal representations for complex visual patterns in humans.
• To propose a unified statistical theory for representational learning.

Main Methods:

• Experiments with human adults and infants on learning complex visual patterns.
• Analysis of learning through the lens of probabilistic inference, specifically Bayesian model averaging.
• Review of existing data from animal classical conditioning studies.

Main Results:

• Human learning of complex visual patterns is better explained by Bayesian model averaging than simple associative schemes.
• The brain appears to generate internal representations by chunking experiences into independent units.
• This generative model optimizes generalization to future stimuli.
• Similar representational schemes have successfully described animal classical conditioning data.

Conclusions:

• Human perceptual learning and animal classical conditioning can be unified under a representational learning framework.
• Bayesian model averaging provides a powerful model for understanding how the brain forms internal representations.
• Statistical theories of learning offer a coherent framework linking diverse learning phenomena.