Leveraging the Potentials of Dedicated Collaborative Interactive Learning: Conceptual Foundations to Overcome Uncertainty by Human-Machine Collaboration

摘要

When a learning system learns from data that was previously assigned to categories, we say that the learning system learns in a supervised way. By "supervised", we mean that a higher entity, for example a human, has arranged the data into categories. Fully categorizing the data is cost intensive and time consuming. Moreover, the categories (labels) provided by humans might be subject to uncertainty, as humans are prone to error. This is where dedicate collaborative interactive learning (D-CIL) comes together: The learning system can decide from which data it learns, copes with uncertainty regarding the categories, and does not require a fully labeled dataset. Against this background, we create the foundations of two central challenges in this early development stage of D-CIL: task complexity and uncertainty. We present an approach to "crowdsourcing traffic sign labels with self-assessment" that will support leveraging the potentials of D-CIL.