Intrinsic stabilization of synaptic plasticity improves learning and robustness in artificial neural networks

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Synaptic plasticity is crucial for learning but can cause instability if unregulated.
• Biological systems use feedback mechanisms to control synaptic changes.
• Artificial neural networks often struggle with stable learning and generalization.

Purpose of the Study:

• To introduce a novel framework, intrinsic Top-Down Stabilization (iTDS), for stabilizing synaptic plasticity in artificial neural networks.
• To leverage top-down signals derived from network output to modulate synaptic updates.
• To enhance training efficiency, generalization, and noise resilience in neural networks.

Main Methods:

• Developed the intrinsic Top-Down Stabilization (iTDS) model, incorporating a slow, top-down signal that tracks network output.
• Applied iTDS to augment traditional supervised learning by modulating synaptic updates based on network activity.
• Evaluated iTDS across recurrent, feedforward, and reservoir networks on diverse tasks.

Main Results:

• iTDS significantly improved training efficiency and enhanced generalization capabilities.
• The framework demonstrated increased resilience to noise perturbations.
• Network activity was shown to influence the alignment of top-down signals with supervised learning objectives.

Conclusions:

• Intrinsic Top-Down Stabilization (iTDS) offers a biologically inspired mechanism for stable and efficient learning in artificial neural networks.
• The findings provide testable predictions for the role of feedback projections in stabilizing biological learning.
• This approach enhances network robustness and performance across various computational tasks.