| Recent advances in Magnetoencephalography (MEG) provide a significant new approach to study neural activity in humans. In many MEG studies different brain states are considered, which differ in the effect of an interested mental process only. These brain states are compared to determine significant neural activity associated with the interested mental process, which is referred to as task-related neural activity.;To investigate the task-related activity using MEG, we should first check whether this activity is detectable in MEG recordings. We put forward that MEG decoding can be used as an instrument to address this problem. If high decoding accuracy is achieved, strong task-related activity exists. However, low accuracy might because of the poor performance of decoding algorithms. Thus it is important to develop high performance decoding algorithms to efficiently extract the discriminant information between brain states. In this thesis, we propose a clustering linear discriminant analysis (CLDA) algorithm for high performance MEG decoding based on small training sets.;MEG decoding tells whether the task-related activity is detectable. However, it does not uncover the neural activity in the brain space. To localize the cortical regions that produce task-related activity, we propose a discriminant pattern source localization (DPSL) algorithm. DPSL consists of two major steps. First, discriminant analysis is applied to find a filter to optimally distinguish different brain states. Next, the gain of the filter is computed at each source location to reveal the activation map of task-related sources. Since the discriminant analysis algorithms in the first step are particularly designed to be robust to noise, DPSL can efficiently reduce the impact of noise and accurately identify the task-related sources.;As human mental processes are controlled by distributed cortical networks, the task-related activity often appears in multiple cortical regions. However, a number of new MEG applications have recently emerged and suggest a need to decode different brain states by signals arising from a specific cortical region. Towards this goal, we propose a region-of-interest-constrained discriminant analysis (RDA) algorithm, which integrates linear classification and beamspace transformation into a unified framework by formulating a constrained optimization problem. |
