Biologically plausible learning in neural networks with modulatory feedback

Area of Science:

• Computational Neuroscience
• Neural Plasticity
• Machine Learning

Background:

• Hebbian learning is a cornerstone of neural plasticity but struggles with modulatory feedback connections.
• Modulatory feedback constitutes a significant portion of neural connections.
• Existing models fail to accurately capture learning in these complex networks.

Purpose of the Study:

• To develop a new, physiologically plausible learning rule for modulatory feedback connections.
• To demonstrate the rule's efficacy in learning complex neural models.
• To address limitations of Hebbian learning in specific neural network architectures.

Main Methods:

• Proposed a novel learning rule based on three simple, physiologically grounded concepts.
• Utilized border ownership as a test case to compare the new rule against Hebbian learning.
• Applied the rule to model orientation selectivity networks, including those without modulatory connections.

Main Results:

• The proposed rule successfully learned a stimulus-driven border ownership model, unlike standard Hebbian learning.
• This marks the first successful learning of a border ownership network.
• The rule served as a viable replacement for Hebbian learning in orientation selectivity models.

Conclusions:

• The developed learning mechanisms are crucial for biological learning of modulatory connections.
• Modulatory connections likely exhibit a strong reliance on inhibitory processes.
• This work advances computational models of neural plasticity and learning.