Digital computing through randomness and order in neural networks.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Information Theory

Background:

• The brain's mechanisms for coding and decoding information remain incompletely understood.
• Existing models often struggle with the efficiency and scalability observed in neural processing.
• The role of randomness and specific coding schemes in neural computation requires further elucidation.

Purpose of the Study:

• To propose and validate a digital computation model for neural coding and decoding.
• To demonstrate the sufficiency of relative ordinal coding, random connections, and belief voting for error-free sequence reconstruction.
• To derive information-theoretic limits and explore implications for artificial neural networks and neurobiology.

Main Methods:

• Development of a computational model based on three principles: relative ordinal coding, random neural connections, and belief voting.
• Mathematical analysis to determine neuron requirements for coding exponentially growing input repertoires.
• Derivation of Shannon equations for neural information capacity and generalization to artificial neural networks.

Main Results:

• The proposed model achieves error-free reconstruction of sequences despite coarse relative codes and randomization.
• Neuron count scales linearly with input repertoire size, which grows exponentially, demonstrating high efficiency.
• Random connections were shown to decorrelate redundant information, leading to more compact neural codes and increased information capacity.

Conclusions:

• A digital computation framework effectively explains neural coding and decoding, aligning with efficient coding principles.
• The model provides a theoretical basis for understanding neural information processing and designing efficient artificial neural networks.
• This work contributes to developing a unified neural information theory applicable to both biological and artificial systems.