Learning of Periodic Attractors in Simple Dynamic Binary Neural Networks

This paper studies learning capability of simple dynamic binary neural networks characterized by signum activation function and ternary weighting parameters. Applying a correlation-based learning method, the network can store a class of periodic attractors. The network dynamics can be integrated into a simple return map on lattice points and we can grasp basic characteristics such as the number of periodic attractors and their domain of attraction. Performing basic numerical experiments, typical network dynamics learning examples are demonstrated.