Feed-back neural networks with discrete weights
We use the Monte Carlo Adaptation learning algorithm to design feed-back neural networks with discrete weights. The dynamic properties of these types of neural networks are investigated as a function of the states of weights. The numerical results of these networks show three phases: the"chaos phase", the "pure memory phase"and the "mixture phase" in the parameter space. The maximum storage ratio for the "pure memory phase" increases with the increasing of the states of the weights, which is favorable for practical applications.