NSPDI-SNN: An efficient lightweight SNN based on nonlinear synaptic pruning and dendritic integration.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) are gaining traction in AI for their biological plausibility.
• Existing SNNs often lack the complex dendritic structures found in biological neurons, limiting their processing capabilities.
• Biological dendrites exhibit nonlinear processing and sparse properties crucial for efficient computation.

Purpose of the Study:

• To propose an efficient and lightweight SNN method that incorporates nonlinear dendritic integration and synaptic pruning.
• To enhance the spatiotemporal information representation in SNN neurons.
• To achieve high sparsity in SNNs while maintaining performance.

Main Methods:

• Introduced Nonlinear Dendritic Integration (NDI) to enhance neuron information representation.
• Implemented heterogeneous state transition ratios for dendritic spines.
• Developed a flexible Nonlinear Synaptic Pruning (NSP) method for high SNN sparsity.
• Conducted experiments on benchmark datasets (DVS128 Gesture, CIFAR10-DVS, CIFAR10) and complex tasks (speech recognition, maze navigation).

Main Results:

• The proposed NSPDI-SNN method achieved high sparsity with minimal performance degradation across all tested tasks.
• NSPDI-SNN demonstrated superior performance on event stream datasets compared to existing methods.
• Analysis confirmed that NSPDI significantly improved synaptic information transfer efficiency as sparsity increased.

Conclusions:

• The nonlinear structure and computation of neuronal dendrites offer a promising avenue for developing efficient SNNs.
• NSPDI-SNN presents an effective approach for creating lightweight and high-performing SNNs.
• This research highlights the potential of bio-inspired dendritic computing for advancing artificial intelligence.