The length of attractors in asymmetric random neural networks with deterministic dynamics
K Nutzel · Journal of Physics A Mathematical and General · 1991
The author has developed a method to detect attractors of any length in large neural networks with up to 1024 neurons within a reasonable period of CPU-time. In networks with symmetric couplings only stable states and, in the case of parallel dynamics, cycles of length 2 exist. The presented simulations suggest that, in sufficiently large systems, this holds also for couplings up to a distinct value of asymmetry. Beyond this value extremely long cycles are detected and the average cycle length depends exponentially on system size.