During learning, both the sources and the mapping of the network evolve together. When the network is presented new data, it is necessary to find the estimates of the sources corresponding to the new data. This can be difficult using similar update process as was used in learning because it is possible that during learning the network develops local minima which make later inversion difficult.
Experiments have shown that it is possible to learn an auxiliary MLP network which will estimate the mapping from observations to sources and can thus be used to initialise the sources given new data. The resulting system with two MLP networks resembles auto-associative MLP network. As was argued before, learning only the generative model is faster than learning a deeper auto-associative MLP with both the generative model and its inverse. Initial experiments have also shown that updates of the sources after the initialisation with the auxiliary MLP network lead to better estimates of the sources.