Introducing children to machine learning concepts through hands-on experience

Teaching Machine Learning in K–12 Classroom: Pedagogical and Technological Trajectories for Artificial Intelligence Education

“…Instead of relying on explicitly hand-coded rules that a computer will follow on different inputs, many ML initiatives guide students to train ML models by giving the system a lot of data to learn from [21]. Studies have used drawings, poses, speech, and video [Authors2019b, Authors2020a, Au-thors2020b], as well as data from, for instance, tracking sports activities [36], [120], webcam [92], [93] gestures [35], web searches [112], and cartoon pictures about kids and mock data about them [88]. As a result, how to curate, create, clean, label, and feed the training data has become a central learning objective for many machine learning initiatives in school computing education [35], [46], [54], [88], [112] [Authors2020c].…”

“…As a result, how to curate, create, clean, label, and feed the training data has become a central learning objective for many machine learning initiatives in school computing education [35], [46], [54], [88], [112] [Authors2020c]. The concerns in making those kinds of ML systems work well in the classroom are much more about the quality of data than about choosing the right rules; for instance, one study looked at learners understanding the impact of sample size, sample versatility, and negative examples on ML model [36]. That computers can learn from data has been seen as one of the key lessons of teaching AI in K-12 [22], [75], [99].…”

“…Many ML in K-12 initiatives have found that even though the feature extraction methods in popular K-12 ML systems may work reasonably reliably and be capable of capturing the essential information from input, the student-produced data has always noise and unwanted features in it-most likely much more so than data sets produced by domain experts working with traditional ML systems has [Authors2020a, 2020b] [22], [35], [36]. For instance, photos taken by the children are often out of focus, audio input can pick up plenty of unwanted background noise from the classroom [Authors2020a,2019b], and one app that children designed to tell whether the speaker is a boy or a girl was unreliable because primary school boys and girls sound much alike [Authors2020c].…”

“…Content may change prior to final publication. interface design, learning ML by creating ML models offers new opportunities for bodily interaction and natural language [Authors2019a,2020a] [16], [35], [36], [47], [99]. For example, educational technology for learning ML principles allowed children to train PoseNet-based ML models to recognize different cheerleader poses (Fig.…”

“…Another described it as requirements analysis, collecting the training data sets, training the model, evaluating the model, and deploying the model to work in apps [Authors2020a]. Also parts of the workflow can be black-boxed: One study focused on just two stages of the workflow, data labeling and evaluation, and black-boxed the other parts of the workflow [35], [36].…”