This paper presents an active learning algorithm based on the Madaline network’s output sensitivity due to its input variation near a given sample. First, a portion of samples are used to train a Madaline network. Then, based on the sensitivity of the network, some other samples with higher sensitivity are actively selected and added into the training data to continue the training of the network. This process is repeated until the training requirement is met. Experiments verified the effectiveness and feasibility of the proposed active learning algorithm in treating discrete classification problems.