A novel neural network for associative memory via dynamical systems

This paper proposes a novel neural network model for associative memory using dynamical systems. The proposed model is based on synthesizing the external input vector, which is different from the conventional approach where the design is based on synthesizing the connection matrix. It is shown that this new neural network (a) stores the desired prototype patterns as asymptotically stable equilibrium points, (b) has no spurious states, and (c) has learning and forgetting capabilities. Moreover, new learning and forgetting algorithms are also developed via a novel operation on the matrix space. Numerical examples are presented to illustrate the effectiveness of the proposed neural network for associative memory. Indeed, results of simulation experiments demonstrate that the neural network is effective and can be implemented easily.