Lesson Overview
Students explore the fundamental concepts behind machine learning — supervised learning, training data, model accuracy, and overfitting — through a hands-on spam classifier activity, visual exploration of a neural network playground, and a structured discussion of algorithmic bias and ethics. They connect ML concepts to real-world applications like recommendation engines, facial recognition, and medical diagnosis — aligned to CSTA 3A-AP-13, 3A-DA-09, and 3A-IC-24.
Learning Objectives
Students will be able to:
- Distinguish between supervised, unsupervised, and reinforcement learning with real-world examples
- Train a simple classification model and evaluate it using accuracy, precision, and recall
- Explain what overfitting is and describe at least two strategies to prevent it
- Analyze a case study of algorithmic bias and articulate why training data quality and diversity matter ethically and technically
Lesson Phases (5E Model)
- 1Ask: 'Raise your hand if you used something powered by AI in the last 24 hours.' — Every hand goes up. 'How do you think Spotify knows what you'll like?'
- 2Quick-write: 'How do you think a spam filter decides what's spam? Write your best guess — no wrong answers.'
- 3Share responses — note the range from 'keyword matching' to 'it uses AI' — validate and build from both
- 4Frame the lesson: 'Today you're going to build your own spam classifier — and discover exactly why training data matters'
Assessment Strategies
Documented model accuracy + written explanation of which features were most/least useful and why
Student records which architectures worked vs. overfit — demonstrates understanding of the complexity trade-off
Scored on identification of training data problem, real-world harm, and proposed intervention — evaluated for specificity and evidence
Formative — look for correct definition of overfitting and a plausible, specific example of training-data-driven harm
Extension Activities
- →Build an image classifier using Teachable Machine — train on your own photos and test its limits
- →Research the EU AI Act or US AI Executive Order — how are governments trying to regulate algorithmic bias?
- →Explore how gradient descent works mathematically — connect to calculus concepts (loss function minimization)
- →Use the Code Academy Data Explorer module to work with real datasets and practice feature selection
At a Glance
Materials
- Spam classifier worksheet (50 email excerpts for manual sorting)1 per pair
- Teachable Machine or ML4Kids access (browser-based)1 per pair
- Neural Network Playground (play.tensorflow.org) access1 per pair
- Algorithmic bias case study card (one of: COMPAS, Amazon hiring tool, medical imaging, facial recognition)1 per pair
- ML Reflection Journal1 per student
Key Vocabulary
- Training data
- The labeled examples used to teach a machine learning model to recognize patterns
- Feature
- An individual measurable property used as input to a machine learning model (e.g., word frequency, pixel brightness)
- Model accuracy
- The percentage of predictions a trained model gets correct on test data it hasn't seen before
- Overfitting
- When a model learns training data too specifically and performs poorly on new data
- Bias (in ML)
- Systematic errors in model output caused by flawed or unrepresentative training data
- Neural network
- A machine learning model loosely inspired by the brain, composed of layers of interconnected nodes (neurons) that transform inputs into outputs
Standards Alignment
Digital Tools
- →Code Academy — AI & Machine Learning Module
- →Code Academy: Data Science Explorer
