An architecture for running multiple ART2 neural networks in parallel is presented. Each of the networks can treat different sizes of clusters. As the result, these multiple networks can show higher performance for detecting new classes. A proposed method is demonstrated for the detection of model changes for time-series models.