Sg-snn: a self-organizing spiking neural network based on temporal information

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• The cerebral cortex self-organizes into functional networks that form attention maps based on input.
• Traditional network self-organization studies often neglect temporal dynamics in neuromorphic data.
• Dynamic neuromorphic data processing requires methods that capture temporal information.

Purpose of the Study:

• To propose a Temporal Self-Organizing (TSO) method for processing dynamic neuromorphic data using spiking neural networks.
• To develop a Self-organizing Glial Spiking Neural Network (SG-SNN) that incorporates glial cell dynamics.
• To generate hierarchical, coarse-to-fine attention topographies for event-based data.

Main Methods:

• Implemented a Temporal Self-Organizing (TSO) method integrating multi-time step information into Best Matching Unit (BMU) selection.
• Introduced a glial cell-mediated Glial-LIF (Leaky Integrate-and-fire) model to simulate neuronal dynamics.
• Optimized attention topological maps by adjusting multiple BMU levels and using a cognitive science-based heuristic.

Main Results:

• The SG-SNN successfully generated attention topographies for dynamic event data.
• Demonstrated improved accuracy on DVS128-Gesture (0.3%), CIFAR10-DVS (2.4%), and N-Caltech 101 (0.54%) neuromorphic datasets.
• Achieved state-of-the-art (SOTA) recognition accuracy of 99.3% on the DVS128-Gesture dataset.

Conclusions:

• The proposed SG-SNN effectively processes dynamic neuromorphic data by creating hierarchical attention maps.
• The TSO method and glial cell integration enhance the network's ability to capture temporal information.
• The SG-SNN represents a significant advancement in neuromorphic computing and event-based data recognition.