Composing recurrent spiking neural networks using locally-recurrent motifs and risk-mitigating architectural...

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Recurrent connectivity is vital for neural circuit function and stability.
• Existing recurrent spiking neural networks (RSNNs) lack systematic architectural optimization.
• Optimizing RSNN architecture is crucial for memory and learning but remains a challenge.

Purpose of the Study:

• To develop a scalable architecture for large RSNNs.
• To introduce an automated method for optimizing RSNN topology.
• To enhance RSNN performance and stability through systematic design.

Main Methods:

• Proposed a Sparsely-Connected Recurrent Motif Layer (SC-ML) architecture for scalability.
• Introduced Hybrid Risk-Mitigating Architectural Search (HRMAS) for topology optimization.
• Incorporated a biologically-inspired intrinsic plasticity mechanism for network self-adaptation.

Main Results:

• Achieved high accuracy on benchmark datasets: TI46-Alpha (96.44%), N-TIDIGITS (94.66%), DVS-Gesture (90.28%), and N-MNIST (98.72%).
• Demonstrated significant performance gains over manually designed RSNNs.
• Successfully enabled systematic architecture optimization for RSNNs.

Conclusions:

• The novel SC-ML architecture and HRMAS method enable scalable and optimized RSNNs.
• Automated architectural optimization with intrinsic plasticity enhances RSNN performance and stability.
• This work presents the first systematic architecture optimization for RSNNs.