Effective Human-in-the-Loop Learning in Structured Noisy Domain

Most machine learning methods rely on large amounts of data that is not noisy or uncertain in order to learn effective hypotheses. In many tasks such data is not available as it can be expensive to obtain. Furthermore, standard approaches use flat feature vectors to represent data while humans often think in terms of objects and the relationships among them. We present human-in-the-loop methods that are capable of learning from human experts to build robust models in the presence of noisy or uncertain, structured data.;Human experts have decades of experience that can be effectively leveraged during learning. However, many domain experts are not machine learning experts. Therefore, we close the loop by building systems that try to understand their own limitations and query the expert about the lesser known areas of the feature space. This could also improve the efficiency or effectiveness of machine learning experts. Taking advantage of the structure of the data, we effectively communicate with the expert in a more natural way. We investigate human-in-the-loop methods in sequential decision-making tasks and structured domains.