Classification of data sets helps to understand their hidden nature. Among different existing classifiers, neural network is more frequently used, and despite its complexities, it has gained widespread applications. Due to significant capabilities of neural network, a good number of softwares have been developed to facilitate its usage. Unfortunately, unawareness of the logic of this technique, hidden in these softwares, leads to partially wrong interpretation of data sets. This paper, first present a new logic for joint set classification by neural network, and then attempt to discuss the uncertainties in performances of the MLP neural network classifier. In order to study the applicability and advantage of the new method in joint set classification, 8 sets of synthetic joints were developed, each with 4 defined characteristics (dip, dip direction, infilling type and infilling rate). The spatial distribution of joints were selected in a way that with two parameters (dip and dip direction), and traditional tools (rose diagram and stereonet), only a limited number of sets could be differentiated. A computer program, using neural network, was developed and the synthetic data was classified in four dimensional spaces. Present study showed that this new technique can easily differentiate and classify all the 8 synthetic sets.