Computational modelling is increasingly expected to match experimental data quantitatively, not just qualitatively. Yet the actual biology is messy and complicated, hence realistic models typically are nonlinear and have a multitude of weakly constrained parameters. Advances in non-invasive neuroimaging have renewed the interest in neural population models (NPMs), since they provide manageable descriptions at the experimentally available spatial resolution. NPMs consider the collective activity of large groups of neurones as the effective source of (macroscopically observable) brain dynamics. Such a coarse-grained view allows simulations of the activity of entire brains even with currently available computers. Using them as forward models can provide mechanistic insights into the effects of psychoactive drugs, and I will focus particularly on general anaesthesia and burst suppression as examples. However, in order to properly engage with available experimental data it is necessary to deal with the problem of inversion, i.e., adjusting the model state and parameters to match the given observations. I will report on several attempts we have made using data assimilation, fitting of power spectral densities, and MCMC methods. This will include some comments of likely general technical interest concerning the importance of equilibria when attempting fits in quasi-linear regimes.