Summary: Motivated by the potential field of static electricity, a binary potential function classifier views each training sample as an electrical charge, positive or negative according to its class label. The resulting potential field divides the feature space into two decision regions based on the polarity of the potential. In this paper, we revisit potential function classifiers in their original form and reveal their connections with other well-known results in the literature. We derive a bound on the generalization performance of multiclass potential function classifiers based on the observed margin distribution of the training data. A new model selection criterion using a normalized margin distribution is then proposed to learn “good" potential function classifiers in practice.