Partially recurrent neural networks for production of temporal sequence

This paper proposes six partially recurrent neural network architectures to evaluate the roles played by interlayer and intralayer feedback connections in producing a temporal sequence of states. The models are divided in two groups according to number of interlayer feedback connections: the first three architectures have nontrainable one-to-one connections, while the last three models have adaptable all-to-all links. Each group has two options for intralayer connections location: either in the input or in hidden layer. The results suggest good performance for planning in different levels of complexity. However, the results suggest the models have poor generalization power