An unified approach for integrating explicit knowledge and learning by example in recurrent networks

The authors propose a novel unified approach for integrating explicit knowledge and learning by example in recurrent networks. The hypothesis is that for a model to be effective, this integration should be as uniform as possible. The authors propose an architecture composed of two cooperating subnets. The first one is designed in order to inject the available explicit knowledge, whereas the second one is learned to allow management of uncertain information. Learning is conceived as a refinement process. The authors report preliminary results for a problem of isolated word recognition to evaluate the proposed model in practice.