Using Machine Learning to Overcome the Expert Blind Spot for Perceptual Fluency Trainings

摘要

Most STEM domains use multiple visual representations to illustrate complex concepts. While much research has focused on helping students make sense of visuals, students also have to become perceptually fluent at translating among visuals fast and effortlessly. Because perceptual fluency is acquired via implicit, nonverbal processes, perceptual fluency trainings provide simple classification tasks that vary visual features across numerous examples. Prior research shows that learning from such trainings is strongly affected by the sequence of the examples. Further, prior research shows that perceptual fluency trainings are most effective for high-performing students but may confuse low-performing students. We propose that a lack of benefits for low-performing students may result from a perceptual expert blind spot of instructors who typically develop perceptual fluency trainings: expert instructors may be unable to anticipate the needs of students who do not see meaningful information in the visuals. In prior work, we used a machine-learning approach to develop a sequence of example visuals of chemical molecules for low-performing students. This study tested the effectiveness of this sequence in comparison to an expert-generated sequence in a randomized experiment as part of an undergraduate chemistry course. We determined students' performance based on log data from an educational technology they used in the course. Results show that the machine-learned sequence was more effective for low-performing students. The expert sequence was more effective for high-performing students. Our results can inform the development of perceptual-fluency trainings for adaptive educational technologies.