Neuroimaging datasets often have a very large number of voxels and a very small number of training cases, which means that overfitting of models for this data can become a very serious problem. Working with a set of fMRI images from a study on stroke recovery, we consider a classification task for which logistic regression performs poorly, even when L1- or L2- regularised. We show that much better discrimination can be achieved by fitting a generative model to each separate condition and then seeing which model is most likely to have generated the data. We use discriminative fitting of exactly the same set of models to demonstrate that the superior discrimination performance is caused by the generative fitting rather than the type of model. We used restricted Boltzmann machines as our generative models, but our results suggest that many other generative models should be tried for discriminating different conditions in neuroimaging data.