In some recent works, it was shown that any algorithmic description might be mapped on a recurrent neural network. A neural oriented language called NETDEF, such that each program corresponds to a modular neural net that computes it, is the tool to achieve this. This article focuses on merging symbolic and subsymbolic computation. Adding high-order neurons to the network model allows learning integration into the NETDEF symbolic computing paradigm, since it is possible to execute learning algorithms in the same neural model that performs symbolic computation. It is shown how the model may process competitive learning methods inside this framework.