Fully memristive spiking neural network for energy-efficient graph learning.

Area of Science:

• Computer Science
• Materials Science
• Neuroscience

Background:

• Shortest path searching on large graphs is computationally intensive and energy-consuming with conventional sequential methods.
• Existing approaches struggle with the demands of large-scale and real-time graph processing.

Purpose of the Study:

• To develop a parallel and energy-efficient method for shortest path searching and graph learning.
• To leverage memristor spiking neural networks (SNNs) through algorithm-device codesign for enhanced graph computation.

Main Methods:

• A highly parallel memristor SNN approach utilizing simultaneous spike traveling for natural shortest path identification.
• Implementation of a nonlinear weight mapping strategy to address neuron nonlinearity and ensure accuracy for large graphs.
• Experimental validation of memristor hardware in unsupervised and supervised classification tasks.

Main Results:

• The proposed method achieves parallel shortest path discovery with extremely low time and space complexity.
• Demonstrated accuracy in classification tasks using memristor-based SNNs.
• Achieved an estimated energy efficiency of 517.82 giga-traversal edges per second per watt, outperforming FPGAs by 3-4 orders of magnitude.

Conclusions:

• The memristor SNN approach offers a significant advancement in energy-efficient graph computing.
• This work provides a viable pathway towards developing highly efficient hardware for graph-based machine learning and computation.
• The algorithm-device codesign strategy is effective for overcoming hardware limitations in neuromorphic computing.