Maximizing neuron decisiveness for generalization

I like the idea in this paper
of maximizing neuron decisiveness on the training examples to get better generalization. I guess it would help put the examples in the middle of (nicely shaped) decision regions whereas other training methods would allow the examples to be at the boundary of (misshaped) decision regions, as long as they were just inside. I can try it with a type of associative memory I have by building up a pyramid structure.