The document discusses implementing machine learning incrementally using proprietary data. It presents two case studies: (1) a QA system that uses incremental training and stacking to answer natural language questions using a small database, and (2) single cell recognition that uses open set recognition models given limited labeled data. The choice of algorithm depends on factors like the business case, available data, architecture and need for rapid delivery when full labeled datasets are difficult to obtain. Incremental algorithms allow products to improve quality using small, proprietary datasets over time.