In this study we compare five classification methods for detecting activation in fMRI data: Fisher linear discriminant, support vector machine, Gaussian nave Bayes, correlation analysis and k-nearest neighbor classifier. In order to enhance classifiers performance a variety of data preprocessing steps were employed. The results show that although kNN and linear SVM can classify active and nonactive voxels with less than 1.2% error, careful preprocessing of the data, including dimensionality reduction, outlier elimination, and denoising are important factors in overall classification.