Chaotic Neural Networks with Non-monotonic Functions and its Adaptation to Memory Search Systems

In past studies, neurons in chaotic neural networks have sigmoid function as an activated function. This paper proposes a new chaotic neural networks using a non-monotonic activated function. This network generate chaotic dynamics by transforming a shape of the function. We apply this network to the memory search systems. In that case, we introduce an original control term to the networks. It leads neurons to give stronger signal outputs with characteristic condition agreements, and weaker signal for disagreements. Adding constraints to the state transitions of the network, the output of network becomes more changeable to the state where the condition is satisfied. Due to its effect, recalling on a target pattern in fewer steps is achieved on average. Performance of the memory search system has been also greatly improved with the number of memory patterns where the conventional methods with sigmoid functions hardly recalled. Furthermore, our memory search system shows a great improvement in the case that each stored pattern has high degree of correlation.