Just curious, what prevents somebody to debug deep network in the same way? You potentially can check what features contributed to activation of each "neuron"..