Demystifying unsupervised learning: how it helps and hurts

Area of Science:

• Cognitive Science
• Machine Learning
• Neuroscience

Background:

• Humans and machines learn without explicit supervision.
• Unsupervised learning is crucial for machine success.
• Human learning outcomes with unsupervised data are inconsistent.

Purpose of the Study:

• Investigate why unsupervised learning yields mixed results in humans.
• Propose a framework explaining human self-reinforcement in learning.
• Clarify conditions under which unsupervised learning benefits or harms human learning.

Main Methods:

• Synthesized empirical results across diverse learning domains.
• Analyzed the role of self-reinforcement in human prediction.
• Framework development based on alignment between predictions and tasks.

Main Results:

• Mixed results in human unsupervised learning stem from self-reinforcement.
• Self-reinforcement can be beneficial or detrimental.
• Learning outcomes depend on the alignment of internal predictions with external task demands.

Conclusions:

• Unsupervised learning's impact on humans is contingent on prediction-task alignment.
• This framework reconciles conflicting findings in human learning research.
• Provides insights for optimizing instruction and lifelong learning strategies.