A new neural network architecture, ARTSTAR, is presented as a supervised modular extension to the ART2 network. ART2 suffers from deficiencies in terms of consistency and overall capability when applied to classification tasks. ARTSTAR uses a layer of INSTAR nodes to supervise and integrate multiple ART2 modules. Supervision takes the form of feedback to the ART2 output layer whenever a data pattern's true classification is known. This feedback technique may take a variety of forms and can model the supervision implemented in existing supervised extensions to ART networks. A more robust classification performance occurs when several ART2 networks are trained in a supervised manner, each under different conditions, and their outputs integrated during testing. These results are demonstrated in tests of ARTSTAR using handdrawn and computer generated digits. The general functionality of ARTSTAR is extensive, and several further modifications to it are discussed.