Two complementary paths to efficiency
Spiking networks transmit information through events; sparse structure determines which connections participate. CH-SNN combines correlation-guided initialization, connection removal and CH link prediction to explore learning at high sparsity.
Testing structure on defined tasks
In the paper’s S-TP experiment on DVS-Gesture, connection sparsity is 98.84% and accuracy is 91.29%, compared with 89.02% for the dense baseline. These results apply to the specified task and configuration.
Understanding energy results
Across four benchmarks, the paper reports an average 55-fold reduction in estimated energy for synaptic operations. The estimate is based on operation counts and assumed energy per operation, rather than measurements of whole-device energy use or battery life.