Empirical analysis of bifurcations in the full weights space of a two-neuron DTRNN

This is an empirical analysis of the dynamic behavior of Discrete-Time Recurrent Neural Networks (DTRNN) with two neurons based on the existing bifurcations on the full 4-dimensional synaptic weights space. We describe the existing bifurcation manifolds and the corresponding expected behavior in each region delimited by them in terms of the existing attractors. We found an unexpectedly rich variety of behaviors, however, finite and classifiable. We propose also an algebraic nomenclature that helps in the understanding of the underlying structure in the weights space.