Incremental knowledge acquisition in supervised learning networks

Acquiring new knowledge without interfering with old knowledge is a key issue in designing an incremental-learning system. The success of such a system hinges on an embedded incrementable information structure with improved performance over time. This paper describes an incremental-learning network for pattern recognition that uses a rule-based connectionist technique to represent general domain and case-specific knowledge, uses bounded weight modification to update its connection weights, and also performs structural learning. Specific strategies are developed for preventing overtraining and for incrementally growing and pruning the network. The soundness of this approach is demonstrated by empirical studies in two independent domains.