A general framework for interpretable neural learning based on local information-theoretic goal functions.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Information Theory

Background:

• Understanding local learning dynamics in biological and artificial networks is challenging.
• Existing approaches lack a general, interpretable, and adaptable framework for local learning goals.

Purpose of the Study:

• To derive a parametric local learning rule based on Partial Information Decomposition (PID).
• To introduce and demonstrate the capabilities of novel "infomorphic" neural networks.

Main Methods:

• Formulated a local information processing goal for compartmental model neurons.
• Derived a local learning rule using advances in Partial Information Decomposition (PID).
• Introduced and tested "infomorphic" neural networks on diverse learning tasks.

Main Results:

• Successfully derived a parametric local learning rule.
• Demonstrated the versatility of infomorphic networks in supervised, unsupervised, and memory learning tasks.
• Showcased the interpretability of the PID framework for analyzing local learning.

Conclusions:

• Infomorphic neural networks provide a valuable, interpretable tool for studying local learning.
• This approach advances the understanding of how local dynamics contribute to network-level solutions.
• The framework is adaptable across various learning tasks and network structures.