Balancing complexity, performance and plausibility to meta learn plasticity rules in recurrent spiking networks

Area of Science:

• Computational Neuroscience
• Machine Learning
• Systems Neuroscience

Background:

• Synaptic plasticity underlies the brain's learning and memory capabilities.
• The precise rules governing synaptic plasticity and their network-level consequences are not fully understood due to experimental constraints.

Purpose of the Study:

• To meta-learn local co-active plasticity rules in large recurrent spiking neural networks using evolutionary strategies (ES).
• To investigate the discovery of rules that stabilize network dynamics and enable complex computations like familiarity detection.

Main Methods:

• Employing evolutionary strategies (ES) for meta-learning plasticity rules in excitatory (E) and inhibitory (I) spiking neural networks.
• Systematically increasing the complexity of plasticity rule parameterizations.
• Analyzing the covariance matrix to understand parameter interdependencies.

Main Results:

• Successfully discovered plasticity rules that robustly stabilize network dynamics across all four synapse types (E-E, E-I, I-E, I-I).
• Demonstrated the ability to incorporate complex functions like familiarity detection into search constraints.
• Identified challenges in meta-learning complex co-active rules and the issue of degenerate solutions.

Conclusions:

• Machine learning, specifically ES, offers a viable approach for discovering synaptic plasticity rules in large spiking networks.
• Current meta-learning strategies face limitations with increasing rule complexity and require more sophisticated loss functions.
• Further development of search strategies is necessary to explore the degeneracy of solutions and fully understand network behavior.