Exploring Machine Teaching with Children

Abstract

Recent government initiatives call for an AI curriculum as early as the first five years of schooling. Initial efforts in AI for K-12 typically emphasize learning activities that require some programming skills, e.g. Scratch. However, informal activities that introduce AI concepts without requiring children to know programming hold promise for broadening participation in STEM fields. In this thesis, we hypothesize that solutions that support machine teaching can help achieve this goal. With a focus on supervised learning, we explore how we can leverage machine teaching to create “low-floor” and “wide-walls” activities that enable AI experimentation by positioning the children as the “teachers” and an image classification algorithm as the “learner”. We propose a series of participatory design studies with university-based multi-generational design teams of children (7-12 years old) and adults. In the first study, we explore how children use a machine teaching system like Google Teachable Machines. Observations from this study indicate that (1) making metrics such as confidence scores visible and dynamic can enable experimentation; (2) engaging children in collaborative activities where they compare training sets and strategies with others can support reasoning; and (3) employing accessible modalities (e.g. images for sighted children) can enable quick data inspection and training adjustment. We use these insights to design and develop a novel teachable application, called IAMGroot, which enables children to display their 3D artwork via augmented reality in a museum exhibit. Observations and preliminary analysis of children’s datasets from a deployment in the Renwick Gallery indicate that children feel empowered to (1) use IAMGroot to display their art in a public museum and (2) explore machine learning concepts like diversity in their datasets. In a third study, we invited children to engage in problem formulation activities where they imagine their own machine teaching applications. Our findings indicate that many of children’s designs require advanced intelligence, e.g. emotion detection and understanding of a user’s social relationships. Children envision that the underlying models could be trained under multiple modalities and any errors would be fixed by adding more data or by anticipating negative examples. Their ideas emphasize more human involvement in AI, although some prefer autonomous systems. Their designs prioritize values like capability, logic, helpfulness, responsibility, and obedience, and a preference for a comfortable life, family security, inner harmony, and excitement as end-states. In the final study, we explored how to translate the above machine teaching activities in the context of a game. The goal is to enable similar AI experimentation without adult guidance (present in our previous studies) or specific settings (like the museum). Through this game, we aim to engage children in playful activity with actionable feedback and promote AI experimentation via carefully engineered problems ordered by difficulty (i.e., “leveling up” in the game). We co-designed a game, Photo Fighters, with children as a low-barrier-to-entry means for children to engage playfully with machine learning concepts. Children ages 7-12 can interact and experiment directly with machine learning and teaching concepts by taking pictures and training an image classifier. “Low-floor approaches” like this are supported by research indicating that machine teaching methods, where the children are the teachers and the model is the student, are practical and accessible for this age group. The game features multiple levels that introduce various challenges, such as increasing data diversity and managing numerous classes in a classifier. Children are rewarded with victory and unlocking new customization options to enhance their character’s visual attributes. Both battling and customization stages are intended to help children understand and navigate the complexities of training image classifiers. Every battle level unlocks new powers and customization options, with 3 battle levels and 8 customization options leading to 16 possible avatar configurations. All of these studies aim to broaden children’s participation in a broader variety of meaningful experiences that expose them to essential machine learning and AI concepts.