Reproduce “Cannistraci-Hebb Training on Ultra-Sparse Spiking Neural Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuan Hua, Jilin Zhang, Yingtao Zhang, Wenqi Gu, Leyi You, Baobo Xiong, Carlo Vittorio Cannistraci, Hong Chen
The Cannistraci-Hebb Spiking Neural Network (CH-SNN) framework enables ultra-high structural sparsity in Spiking Neural Networks (SNNs), reaching up to 98.84% while consistently maintaining or improving accuracy across various datasets. This approach significantly reduces synaptic operations and energy consumption, demonstrating enhanced efficiency on neuromorphic hardware.
Key insights
- Leveraging brain-inspired Cannistraci-Hebb learning theory and principles from complex network science enables the robust and dynamic formation of ultra-sparse network topologies.
- SNN-specific initialization methods (SSCTI and SSWI) are critical for maintaining training stability and high performance, particularly at extreme sparsity levels.
- The combination of dynamic link pruning with neuron-level chain removal effectively eliminates redundant connections and inactive neurons, leading to improved information flow and higher overall sparsity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.