Analysis and design of multivalued many-to-one associative memory driven by external inputs

This paper proposes a novel multivalued recurrent neural network model driven by external inputs, along with two innovative learning algorithms. By incorporating a multivalued activation function, the proposed model can achieve multivalued many-to-one associative memory, and the newly developed algorithms enable effective storage of many-to-one patterns in the coefficient matrix while maintaining the indispensability of inputs in many-to-one associative memory. The proposed learning algorithm addresses a critical limitation of existing models which fail to ensure completely erroneous outputs when facing partial input missing in many-to-one associative memory tasks. The methodology is rigorously derived through theoretical analysis, incorporating comprehensive verification of both the existence and global exponential stability of equilibrium points. Demonstrative examples are provided in the paper to show the effectiveness of the proposed theory.