Synaptic balancing: A biologically plausible local learning rule that provably increases neural network noise...

Area of Science:

• Computational Neuroscience
• Machine Learning
• Dynamical Systems

Background:

• Recurrent neural networks (RNNs) are powerful but sensitive to noise.
• Existing methods for noise robustness often sacrifice performance or task specificity.
• Understanding biologically plausible learning rules is crucial for brain-inspired AI.

Purpose of the Study:

• Introduce a novel, biologically plausible local learning rule for RNNs.
• Enhance neural dynamics' robustness to noise in nonlinear RNNs.
• Investigate the rule's impact on network performance and biological consistency.

Main Methods:

• Developed a local learning rule based on integrable dynamical systems (Lax dynamical systems).
• Analyzed the rule's effect on synaptic balancing (incoming and outgoing synapses).
• Proved noise robustness enhancement without task-specific knowledge.

Main Results:

• The learning rule provably increases noise robustness in nonlinear RNNs.
• Achieved higher noise robustness without performance degradation.
• Demonstrated synaptic balancing, consistent with heterosynaptic plasticity.
• Made experimentally testable predictions for synaptic plasticity.

Conclusions:

• The novel learning rule offers a practical method for creating noise-robust RNNs.
• The rule bridges neurobiology, AI engineering, and mathematical dynamical systems.
• Synaptic balancing is a key mechanism for noise robustness in neural networks.