Cartwright, M., Pardo, B. The Moving Target in Creative Interactive Machine Learning. In Proceedings of the Workshop on Human Centred Machine Learning (HCML) at the ACM Conference on Human Factors in Computing Systems (CHI), 2016.
It is common for the teacher’s understanding of a concept to evolve and change as they teach. This is especially common for creative tasks—preferred goals and methods can and should shift during the creative process. This can be problematic for interactively training a machine learning system to assist a creative task. Algorithms typically presume a constant goal and treat inconsistency in training data as unwanted noise. “Creative types" typically don’t understand the internals of learning algorithms and cannot compensate for the weakness of the algorithms. We must develop methods better able to handle training data that represents a shifting goal or concept. Ideally, these approaches should incorporate a training paradigm that even novice, non-technical users can use effectively.