C2-LSM: A Storm-NoC Based Neuromorphic Processor for High-Accuracy Liquid State Machine With Cube-Cluster Topology

Area of Science:

• Neuromorphic Engineering
• Spiking Neural Networks
• Reservoir Computing

Background:

• Liquid State Machines (LSMs), a variant of Spiking Neural Networks (SNNs), are known for their low training complexity.
• Biological brains exhibit
• small-world
• network structures that inspire efficient neural processing.

Purpose of the Study:

• To propose C2-LSM, a neuromorphic processor developed through algorithm-hardware co-design.
• To enhance accuracy and efficiency for diverse spatiotemporal tasks using LSMs.

Main Methods:

• Algorithm-level design: Introduced a novel reservoir layer with a cubecluster topology inspired by biological neural networks.
• Hardware implementation: Developed a customized C2-LSM processor on an AMD Virtex UltraScale+ VCU129 FPGA with runtime configurability.
• Network-on-Chip (NoC): Integrated a Storm routing algorithm to optimize spike event transmission.

Main Results:

• Achieved high classification accuracies: 98.02% on MNIST, 94.26% on N-MNIST, and 93.00% on FSDD.
• Demonstrated superior performance compared to recently benchmarked LSM neuromorphic processors.
• Achieved 1155 FPS inference and 1154 FPS learning speeds on MNIST with 103 GSOPS/W power efficiency.

Conclusions:

• C2-LSM processor achieves state-of-the-art accuracy and efficiency for spatiotemporal tasks.
• Algorithm-hardware co-design is effective for developing high-performance neuromorphic systems.
• The cubecluster topology and optimized NoC contribute to the enhanced performance of C2-LSM.