We are developing a machine learning system that modifies existing knowledge about specific types of biological sequences. It does this by considering sample members and nonmembers of the sequence motif being learned. Using this information (which we call a ``domain theory``), our learning algorithm produces a more accurate representation of the knowledge needed to categorize future sequences. Specifically, the KBANN algorithm maps inference rules, such as consensus sequences, into a neural (connectionist) network. Neural network training techniques then use the training examples of refine these inference rules. We have been applying this approach to several problems in DNA sequence analysis and have also been extending the capabilities of our learning system along several dimensions.