Intrinsically Stable Spiking Neural Networks: Overcoming the Performance Barrier in the Absence of Batch Normalization
The Intrinsically Stable SNN (IS-SNN) architecture removes computationally expensive activation normalization layers from spiking neural networks to improve hardware efficiency. It achieves this by enforcing intrinsic signal stability through topology-aware weight standardization and modified residual connections. Experiments show IS-SNN maintains competitive accuracy on ImageNet while drastically reducing FPGA resource consumption.