Beyond Black-Boxes: Teaching Complex Machine Learning Ideas through Scaffolded Interactive Activities

Brian Broll, Shuchi GroverAAAI 2023 · DOI

Introduction

As Artificial Intelligence (AI) becomes an integral part of modern technology, there is an increasing push to introduce these concepts at the K-12 level. However, most introductory curricula treat machine learning (ML) as a "black box." Students often use high-level tools like Google’s Teachable Machine, where they provide data and receive a model without understanding the underlying mechanics. This approach can lead to a perception of AI as "magic" rather than a technical, iterative process.

The AI & Cybersecurity for Teens (ACT) curriculum aims to address this by teaching high schoolers the "how" of ML—including complex topics like gradient descent, optimization, and generative adversarial networks (GANs)—without requiring advanced mathematics like calculus.

Pedagogical Approach: Scaffolding and Abstraction

The researchers developed a sequence of activities within NetsBlox, an extension of the block-based language Snap!. Their approach relies on several key pedagogical strategies designed to make complex algorithms accessible:

  • Scaffolded Programming: Using Parson’s Problems, students are provided with necessary code blocks but must arrange them in the correct logical order, focusing on algorithmic logic over syntax.
  • Layers of Abstraction: Activities move from high-level "subgoal" blocks to lower-level implementation blocks. This allows students to "lift the hood" and see the underlying code at their own pace.
Scaffolded blocks in NetsBlox allow students to explore machine learning components like error computation and gradient updates.

Gamifying Gradient Descent

A core challenge in teaching ML is explaining optimization. To teach Gradient Descent without calculus, the curriculum utilizes the "Find the Minimum" game. In this activity, students act as the optimizer, clicking on a blank screen to find the lowest point of a hidden function. They are guided only by arrows representing the "slope" (gradient) at their current position, building a functional intuition for how neural networks update their weights.

The "Find the Minimum" game teaches gradient descent intuition by having students act as the optimizer.

Cybersecurity as a Learning Context

ML concepts are grounded in practical security tasks to maintain student engagement. The curriculum utilizes several datasets and real-world scenarios:

  • Twitter Bot Detection: Using synthetic datasets to teach decision trees that distinguish human accounts from bots.
  • Sentiment Analysis: Leveraging the ParallelDots API to detect abusive text in chat applications.
  • Biometrics: Students collect their own mouse-movement data to train models that distinguish between human users and automated registration bots.

By framing these lessons through cybersecurity, students engage in adversarial thinking—trying to "fool" a model—which naturally leads to discussions about model vulnerabilities, bias, and generalization.

Key Insights and Results

The authors conducted pilot workshops with high school teachers to evaluate the curriculum's effectiveness. Key findings included:

  • Innovation & Appropriateness: The curriculum received high marks for innovativeness (4.5/5) and the appropriateness of combining AI with cybersecurity (4.3/5).
  • Teacher Confidence: Teachers' comfort levels with AI topics rose significantly, from 2.7 to 3.7 on a 5-point scale.
  • Engagement: Qualitative feedback highlighted that the curriculum was "intellectually honest," successfully moving beyond black boxes to show how models are actually coded.

Limitations and Future Work

While teacher feedback was overwhelmingly positive, the current data is limited to workshop settings. Future work involves piloting the ACT curriculum in actual high school classrooms to assess direct student performance across varying levels of prior programming experience. The researchers also plan to develop more differentiated materials to support a broader range of student abilities.

Summary updated

BibTeX
@inproceedings{broll2023beyond,
  author = {Brian Broll and Shuchi Grover},
  title = {{Beyond Black-Boxes: Teaching Complex Machine Learning Ideas through Scaffolded Interactive Activities}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i13.26898},
}
Figure