Population codes enable learning from few examples by shaping inductive bias.

Area of Science:

• Computational Neuroscience
• Machine Learning Theory
• Systems Neuroscience

Background:

• Learning efficiently from limited data requires inductive biases.
• Neural codes are hypothesized to implement and shape these biases.
• Understanding this link is crucial for both biological and artificial learning systems.

Purpose of the Study:

• To develop a theory explaining how neural code structure influences sample-efficient learning.
• To investigate the role of task-code matching in generalization.
• To identify and characterize inductive biases in biological neural codes.

Main Methods:

• Analytical theory for predicting generalization error based on neural codes and data.
• Analysis of arbitrary stimulus-response maps with biologically-plausible readouts.
• Empirical validation using neural recordings from mouse primary visual cortex.
• Modeling of primary visual cortex to reproduce observed biases.

Main Results:

• A precise mathematical framework demonstrating how population code structure dictates inductive bias.
• Identification of an efficiency bias in mouse visual cortex favoring low-frequency discrimination and reconstruction tasks.
• Confirmation that code invariances impact learning performance.
• Observation that biological codes exhibit lower total activity for equivalent inductive biases.

Conclusions:

• Task-code matching is essential for sample-efficient learning.
• Biological neural codes possess specific inductive biases, such as favoring simpler explanations.
• The developed theory provides a method for elucidating brain's inductive biases and promotes sample-efficient learning as a normative principle.