Twin neural network regression

We introduce twin neural network (TNN) regression. This method predicts differences between the target values of two different data points rather than targets themselves. The solution a traditional regression problem is then obtained by averaging over an ensemble all predicted unseen point and training points. Whereas ensembles are normally costly to produce, TNN intrinsically creates predictions twice size set while only single network. Since have been shown be more accurate models this property naturally transfers show that TNNs able compete or yield for sets, compared other state-of-the-art methods. Furthermore, constrained self-consistency conditions. find violation these conditions provides estimate prediction uncertainty.