It might be cool to train neural network by minimizing error with assumption there's noise on both inputs and outputs.