2026 ASEE Annual Conference & Exposition

This paper presents an educational approach designed to demystify machine learning (ML) for high school computer science students by moving beyond abstract, "black-box" tools. The project aims to enhance students' understanding of the complete ML pipeline, including data preparation, model training, and evaluation by having them program a Single Layer Perceptron for binary image classification. Our methodology employs a four-phase, scaffolded progression: 1) introducing foundational ML concepts using Google Teachable Machine; 2) utilizing a custom interactive simulation tool to visualize and understand the Perceptron algorithm's mechanics; 3) programming the algorithm in Java to classify numerical digits from the MNIST dataset; and 4) evaluating the trained model's performance on test data.

This approach was implemented in a high school computer science classroom with 25 students. Instructional effectiveness was measured using pre- and post-tests to assess conceptual knowledge and surveys to gauge student engagement and self-efficacy. The results indicated a statistical improvement in students' post-test scores, demonstrating a gain in their understanding of both general ML concepts and the mechanics of the Perceptron. Survey feedback confirmed that students found the interactive simulation and hands-on programming experience beneficial to their learning. The findings suggest that combining interactive visualization with direct programming provides an effective path for teaching ML algorithms in a high school setting, and bridging the gap between theory and practical implementation.

The average learning gain was about 25% improvement in post-test scores and 80% agreed or strongly agreed that the lesson improved their understanding of ML.