Tag
The author describes building a spiking neural network engine that initially failed the NARMA-10 benchmark, but by applying heterogeneous wire delays from neuroscience, it achieved usable memory depth and a 15x computational efficiency advantage over continuous nets on a recognition task.