SPANNER: A Self-Repairing Spiking Neural Network Hardware Architecture.

Area of Science:

• Neuroscience and Neuromorphic Engineering
• Artificial Intelligence
• Computer Architecture

Background:

• The human brain possesses a self-repair mechanism involving astrocytes that modulate synaptic transmission probability (PR).
• Synaptic faults can lead to neuronal silence due to low PR, but astrocyte feedback can increase PR in healthy synapses.

Purpose of the Study:

• To propose a novel hardware architecture, Self-rePAiring spiking Neural NEtwoRk (SPANNER), that mimics the brain's self-repair capability.
• To demonstrate SPANNER's ability to self-detect and self-repair synaptic faults without conventional fault management components.

Main Methods:

• Development of a novel hardware architecture (SPANNER) inspired by astrocyte-mediated neural self-repair.
• Testing SPANNER's fault detection and self-repair capabilities under various synaptic fault densities.

Main Results:

• SPANNER successfully self-detects and self-repairs synaptic faults.
• The architecture maintains system performance with fault densities up to 40%.
• SPANNER exhibits only 20% performance degradation even at 80% fault density.

Conclusions:

• The proposed SPANNER hardware architecture effectively replicates biological self-repair mechanisms for neural networks.
• SPANNER demonstrates significant resilience and fault tolerance, outperforming conventional approaches in handling synaptic faults.