Learning with Prior Domain Knowledge and Insufficient Annotated Data

Machine learning exploits data to learn, but when not enough data is available (often due to increasingly complex models) or the quality of the data is insufficient, then prior domain knowledge from experts can be incorporated to guide the learner. Prior knowledge typically employed in machine learning tends to be concise, single statements. But for many problems, knowledge is much more messy requiring in-depth discussions with domain experts to extract and often takes many iterations of model development and feedback from experts to collect all the relevant knowledge. In the Bayesian learning paradigm, we learn which hypotheses are most likely given the data as evidence. How can we refine this model when new feedback is given by domain experts? We are working with domain experts on a problem where data is expensive, but we also have prior knowledge. This research has two objectives: (1) automatically refine models using prior knowledge, and (2) handle various forms of prior knowledge elicited from experts in a unified framework.