Magnetoencephalography (MEG) aims at reconstructing the unknown neuroelectric activity in the brain from non-invasive measurements of the magnetic field produced by neural sources in the outer space. The solution of this ill-posed and ill-conditioned inverse problem is dealt with regularization techniques that are often time-consuming, computationally and memory storage demanding. So far, MEG data are pre-processed and analyzed off-line, but currently there is a lot of interest in the development of tools for real-time processing and data analysis to make MEG effective in applications like brain-computer interface training or neurofeedback rehabilitation. In this talk we present a slim procedure, the RAndoM Sampling mEThod (RAMSET), that allows us to considerably reduce the computational cost of an inversion algorithm. We test RAMSET on both synthetic data and real MEG measurements and we investigate how the random sampling affects the localization accuracy of well established inversion methods. This is a joint work with C. Campi (Univ. Padova) and A. Pascarella (IAC-CNR).