Neural representational geometry underlies few-shot concept learning.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• Human cognition excels at learning new concepts from few examples.
• The neural underpinnings of this few-shot learning capability remain largely unknown.

Purpose of the Study:

• Propose a biologically plausible neural mechanism for few-shot concept learning.
• Develop a mathematical theory linking neural geometry to learning performance.
• Investigate the role of neural manifold geometry in few-shot learning.

Main Methods:

• Modeling neural representations in higher-order sensory areas.
• Simulating a downstream readout neuron with a simple plasticity rule.
• Analyzing neural manifold geometry in primate and deep neural network (DNN) models.

Main Results:

• Achieved high few-shot learning accuracy on natural visual concepts.
• Demonstrated learning from linguistic descriptors alone.
• Developed a predictive mathematical theory of few-shot learning.
• Identified high-dimensional manifolds as beneficial for learning.

Conclusions:

• The proposed neural mechanism effectively supports few-shot learning.
• Neural representation geometry is a critical predictor of learning performance.
• Discrepancies exist between primate and DNN manifold geometries.