Training Recurrent Neural Nets with Backpropation through time.
For understanding backprop in matrix notation, look at this excellent post
By Simon. I will try to explain how we can extend it to Recurrent Networks and its implementation in TMVA::DNN.