A²SG adaptive software-only learning maintains high spiking neural network accuracy at 1/6th computational overhead/ memory usage

https://www.eurekalert.org/news-releases/1138818

"neuromorphic stochastic spiking networks have high energy efficiency but suffer from high computational overhead during training due to non-differentiable spiking functions... overcome replacing complex surrogate gradients during backward propagation while dynamically updating synaptic weights based on operational errors... optimization requires no specialized physical hardware modifications... scalable deployment in low-power edge AI devices, wearable sensors, and autonomous systems"