The curriculum effect in visual learning: The role of readout dimensionality

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Generalization of visual perceptual learning (VPL) varies across tasks and is influenced by training curriculum.
• A theoretical explanation for curriculum effects on VPL generalization is currently lacking.
• Understanding these effects is crucial for optimizing learning processes in both humans and artificial systems.

Purpose of the Study:

• To propose a theory explaining visual learning generalization and curriculum effects.
• To compare human VPL with artificial neural network (ANN) models.
• To investigate the role of training curriculum in VPL generalization.

Main Methods:

• Developed an artificial neural network (ANN) model of visual perceptual learning (VPL).
• Compared human VPL performance with ANN models under different training curricula (sequential vs. interleaved).
• Analyzed the relationship between neural population dimensionality and generalization in ANN models.

Main Results:

• Easy-to-hard sequential training improved VPL generalization in both humans and ANNs.
• Interleaved training led to poorer generalization in ANNs but maintained performance in humans, with high inter-individual variability.
• ANN models trained with curricula emphasizing early exposure to easier tasks developed lower-dimensional neural population representations and showed superior generalization.

Conclusions:

• Training curriculum significantly influences VPL generalization, with sequential easy-to-hard training being more effective for robust generalization.
• Neural population dimensionality is a key factor in VPL generalization, modulated by the training curriculum.
• The proposed theory provides a mechanistic link between curriculum design, neural population dimensionality, and VPL generalization.