A neural module to estimate context information for sequence learning

Presents an approach for building a modular neural system for sequence reproduction. The proposed modular neural system deals separately with two types of context information: the spatial context and the temporal context. The former is useful for a sequence identification, and is processed by a feedforward neural network. The latter is needed for learning and reproducing patterns with temporal dependencies, and is processed by a partially recurrent neural network. The achieved results show significant improvements in the generalization ability of the complete proposed modular neural system when compared to the recurrent neural network working alone