Curriculum Learning in Humans and Neural Networks.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Machine Learning

Background:

• Training trial sequencing influences learning outcomes in humans and artificial neural networks.
• Previous research primarily focused on multi-task learning, leaving single-task curriculum learning under-explored.
• Understanding these similarities can validate neural networks as models for human learning.

Purpose of the Study:

• To investigate curriculum learning in a single perceptual decision-making task.
• To compare learning outcomes in humans and a parsimonious neural network under varied training curricula.
• To determine if neural network behavior under curriculum learning is replicable in human participants.

Main Methods:

• Human participants (n=200) and a parsimonious neural network were trained on a perceptual decision-making task.
• Training curricula involved progressively increasing difficulty, fixed difficulty, and random difficulty.
• Learning performance was assessed across different curriculum conditions.

Main Results:

• Progressively increasing task difficulty during training significantly facilitated learning in both humans and the neural network.
• A curriculum designed to impede neural network learning also impaired human learning.
• Qualitative similarities in learning patterns were observed between humans and the neural network.

Conclusions:

• Humans and neural networks exhibit comparable responses to curriculum learning in single perceptual decision-making tasks.
• The findings support the use of neural networks as computational models for human perceptual learning.
• Curriculum design is a critical factor influencing learning efficiency in both biological and artificial systems.