On the reliability of the nervous (Nv) nets

This paper investigates the reliability of a particular class of neural networks, the Nervous Nets (Nv). This is the class of nonsymmetric ring oscillator networks of inverters coupled through variable delays. They have been successfully applied to controlling walking robots, while many other applications will shortly be mentioned. The authors will then explain the robustness of Nv nets in the sense of their highly reliable functioning--which has been observed through many experiments. For doing that the authors will show that although the Nv net has an exponential number of periodic points, only a small (still exponential) part are stable, while all the others are saddle points. The ratio between the number of stable and periodic points quickly vanishes to zero as the number of nodes is increased, as opposed to classical finite state machines--where this ratio is relatively constant. These show that the Nv net will always converge quickly to a stable oscillatory state--a fact not true in general for finite state machines.