Abstract: Functional MRI brain map is characterized by precise positioning,but low signal-to-noise ratio and large volume of data.We analyse the hyper-parameter optimization of general regression model on fMRI brain data,based on the non-homogeneity validation method.To acknowledge the discrepancy in ability of different regression approaches（including ridge regression,support vector regression and Elman Recurrent Neural Network） we compare them over every feature rating of external environment.The original data is obtained from PBAIC2006.The experiment results show that this method has good stability and generalization;also it can be used in the field which lacks of knowledge of the relevant fields.At the same time,we find the linear method is simple,effective in the fMRI regression mission,and has higher predictive ability but lower calculating costs than other methods under many feature rating of external environment.