The study of human brain functions has dramatically increased greatly due to the advent of Functional Magnetic Resonance Imaging (fMRI). Recently, it has been noted that the use of machine learning classifiers for decoding cognitive states directly from fMRI data is a powerful technique that enables researchers to make predictions about the mental state of a subject. In this paper we explore some of the fundamental questions fMRI-decoding raises by applying and comparing different machine learning techniques and feature selection methods to the problem of classifying the instantaneous cognitive state of a person based on fMRI data. In particular, we present successful case studies of induced classifiers which accurately discriminate between cognitive states produced by different stimuli. We show how classifiers can be used as confirmatory tools allowing the testing of competing hypothesis about the structure in the data, and show that it is possible to train successful classifiers without prior anatomical knowledge and using only a very small number of features.