Connection sparsity, with 91.29% accuracy.
In the paper’s DVS-Gesture S-TP experiment, the fully connected baseline achieved 89.02% accuracy: a 2.27 percentage-point difference.
Source: paper §4.2, Table 2.
ORIGINAL SCIENCE / SPIKING NEURAL NETWORKS
Sparse in time.
Sparse by design.
CH-SNN combines event-driven computation with dynamically learned sparse connectivity.
CONNECTIVITY THAT EVOLVES IN TRAINING
Where event-driven computation meets structural learning.
Spiking neural networks (SNNs) communicate through discrete spikes. CH-SNN combines correlation-informed topology initialization, sparsity-aware weight initialization, connection pruning and Cannistraci–Hebb link prediction to evolve connections during training.
Connection regrowth during training and spike events during inference are distinct processes.
PAPER RESULTS · ICLR 2026
Lower estimated synaptic-operation energy on average.
Average versus the fully connected network across four benchmark datasets, estimated from operation counts and the ANP-I processor’s measured 1.5 pJ per synaptic operation.
Source: paper §4.2, Table 2 and Figure 3.
RESEARCH OUTLOOK
New possibilities for intelligence
under resource constraints.
The research informs event-driven vision and neuromorphic computation under resource constraints, and future model–hardware co-design for edge and on-device intelligence.
Paper authors: Yuan Hua, Jilin Zhang, Yingtao Zhang, Leyi You, Baobo Xiong, Carlo Vittorio Cannistraci, Hong Chen.
Cannistraci-Hebb Training on Ultra-Sparse Spiking Neural Networks · ICLR 2026
Read the paper