Efficient Neural Codes That Minimize L_p Reconstruction Error.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Information Theory

Background:

• The efficient coding hypothesis posits that sensory systems evolve neural codes optimized for environmental stimuli.
• Existing models primarily rely on information-theoretic measures, with some exploring downstream decoding performance.
• A unified framework for evaluating diverse optimality criteria in neural coding is lacking.

Purpose of the Study:

• To systematically evaluate different optimality criteria for neural coding using a parametric approach.
• To analytically derive optimal neural tuning curves for single neurons and populations.
• To test the framework's predictions against biological data from early visual systems.

Main Methods:

• Developed a parametric formulation of the efficient coding problem based on maximum likelihood decoder reconstruction error.
• Analytically derived optimal single-neuron tuning curves for arbitrary input distributions.
• Introduced the concept of a meta-tuning curve to generalize findings to neural populations.

Main Results:

• The parametric framework encompasses both information maximization and squared decoding error as special cases.
• Optimal tuning curves were derived, providing a theoretical basis for neural representations.
• Predictions align with empirical data from early visual pathways across various animal models.

Conclusions:

• Neural representations in early visual systems appear to optimize criteria close to information maximization for natural stimuli.
• The proposed framework offers a unified approach to understanding efficient coding principles.
• Findings suggest conserved optimization strategies in biological sensory processing.