In this paper, Kohonen's self-organizing mapping (SOM) is used as a data-driven technique for analyzing functional magnetic resonance imaging (fMRI) data. Upon the completion of an SOM analysis, a cluster merging technique, based on examining the reproducibility of the fMRI data across epochs, is utilized to merge SOM nodes whose feature vectors are sufficiently similar to one another. The resulting 'super nodes' give time course templates of potential interest. These templates can be subsequently used in traditional template-based analysis methods, such as cross-correlation analysis, yielding statistical maps and activation patterns. This technique has been demonstrated on two fMRI datasets obtained from a visually-guided motor paradigm and a visual paradigm, respectively, showing satisfactory results.