Sampling the "Inverse Set" of a Neuron: An Approach to Understanding Neural Nets