A Kohonen network was utilized in order to analyze single unit recordings taken in the inferior temporal cortex of the monkey while it performed a categorization task of distinguishing tree from non-tree images. In particular, the network was used to respond to the question of whether or not the recordings in the area IT were sufficient in order to account for the behavioral success in categorization (above 95%). Due to experimental difficulties, the response of each neuron could only be recorded for some images. The Kohonen network was therefore designed to ignore the missing entries of the incomplete response matrix. With the use of the neuronal responses as input, the network was found to be able to correctly categorize the responses as tree or non-tree responses with up to 88% success. The importance of a particular type of neuron to the success of categorization, was judged by the degree of success in categorization that could be achieved once these neurons were eliminated from the neuronal response matrix. Input from the narrowly tuned, category specific neurons were found unimportant to the degree of success in categorization. By contrast, the elimination of comparable numbers of broadly tuned cells was found to greatly deteriorate the performance of the network. Further examination of the activities of the broadly tuned cells without the Kohonen network revealed that some ‘simple’ algebraic operations among the broadly tuned cells was able to correctly classify the neuronal responses with an 80% success rate.