A　GMDH-type　neural　network　and　its　modified　training　algorithm　were　presented　in　this　paper　to　improve　the　classifying　accuracy　of　EEG　with　different　mental　tasks.　The　network　was　formed　through　evolution,　the　classification　rules　were　described　by　a　concise　set　of　polynomials　and　the　training　algorithm　was　able　to　prevent　overfitting　effectively.　Experimental　results　showed　the　GMDH-type　nearal　could　classify　the　EEG　of　math　or　relaxtasks　with　accuracy　of　84.　5　%　.　It　was　indicated　that　GMDH-type　neural　network　exhibited　higher　classifying　accuracy　compared　to　the　feedforward　neural　network　(FNN).