Implicit-perturbation zeroth-order optimization reduces perturbation energy to 0.46x-0.83x that of conventional explicit weight perturbation for spiking neural networks

IPZO reduces the perturbation energy to 0.46x-0.83x that of conventional explicit weight perturbation for a batch size of B=64 and T=4 time steps

http://arxiv.org/abs/2608.21223v1