Lab 11 · Machine Learning Concept Lab: Train a Simple Classifier (No-Code)

🎯 Objective

Experience the full ML loop — data, training, testing, failure analysis — without code, so the concepts become concrete.

🧩 Skills You Will Practice

• Training data and labels
• Train/test thinking
• Confidence scores
• Dataset bias and limitations

🧰 What You Need

• A free no-code ML tool such as Teachable Machine in a browser
• Your webcam OR a set of your own simple images
• A results table in your notebook

📋 Step-by-Step Tasks

1. Choose a 3-class problem you can demonstrate live, for example Book / Pen / Phone, or Thumbs up / Thumbs down / Open hand. Define the classes before collecting anything — vague classes make vague models.
2. Collect training examples per class with variety: different angles, distances, lighting and backgrounds. Aim for a balanced set — roughly equal examples per class, and note the exact count per class.
3. Train the model and test it live with objects it has NOT seen in training. Record 10 test cases in your table: true class, predicted class, confidence percentage, correct or wrong.
4. Stress-test on purpose: change the lighting, move the background, hold the object partly out of frame. Record how confidence and accuracy change — this is model fragility, observed first-hand.
5. Create a bias case: retrain or imagine one class trained only against one background. Show or explain how the model learns the background instead of the object, and write the fix (varied, balanced data).
6. Write the evaluation report: accuracy out of your 10 tests, the two worst failure cases, one bias risk, and one real-world use where this error rate would or would not be acceptable.

✅ Expected Result / Deliverable

A trained, working classifier plus a written evaluation: 10 recorded tests, measured accuracy, 2 failure cases, 1 bias risk with its fix — the complete ML loop, documented.

⚠️ Common Mistakes

• Testing only with the same images used in training
• Collecting 90 examples of one class and 10 of another, then blaming the model
• Reporting accuracy without recording the actual test cases