Module 02 — How Machine Learning Works

📖

Lesson Content

Read & Understand

Machine learning is the engine inside most modern AI. The core idea: instead of a programmer writing rules, you show the system thousands of labeled examples — "this is a cat," "this is not a cat" — and it learns the rules itself.

This learning happens through a feedback loop. The model makes a prediction. That prediction is compared to the correct answer. The difference (called the error) is used to adjust the model slightly. Repeat millions of times, and the model gets very good at its task.

There are three main flavors. Supervised learning is like learning with a teacher — labeled examples, clear right answers. Unsupervised learning is exploring without a guide — the model finds its own patterns. Reinforcement learning is like training a puppy — reward good actions, penalize bad ones.

Once trained, a model becomes a black box: input goes in, output comes out. Even the engineers who built it often can't explain why it made a specific decision.

Key Takeaways

ML learns from labeled examples, not hand-coded rules

Training is a feedback loop: predict → compare → adjust

Three types: supervised, unsupervised, reinforcement

A trained model is a "black box" — input in, output out

More quality data generally means a better model

🔄

The Training Loop

How a model learns step by step

Feed Training Data

Thousands of labeled examples go in (e.g. photos tagged "cat" or "not cat")

↓

Model Makes a Prediction

The model guesses — "I think this is a cat"

↓

Compare to Correct Answer

How wrong was it? This gap is the "loss" or error

↓

Adjust the Model

Small tweaks to internal weights to reduce error next time

↓

Repeat Millions of Times

Each pass makes the model more accurate

↺ Loop repeats until accuracy is good enough

How a model learns step by step

🗂

Three Types of Machine Learning

Visual analogy comparison

📚

Supervised

Learning with a teacher. Labeled examples, clear right answers. Like studying with an answer key.

🔭

Unsupervised

Exploring without a guide. Finds hidden patterns on its own. Like sorting books with no genre labels.

🐕

Reinforcement

Training by reward & penalty. Agent learns by trying things and seeing outcomes. Like training a dog with treats.

Visual analogy comparison

⬛

The Black Box Model

Input → Model → Output

Inputs

📷 An image

✉️ An email

💬 A question

🎵 Audio clip

→

⬛

The Model

billions of parameters

→

Outputs

🏷 "This is a cat"

🚫 "Spam"

📝 An answer

📄 Transcript

Input → Model → Output

✅

Self-Check Quiz

Click an answer to check your understanding

Q1 of 4

In the training loop, what happens after the model makes a prediction?

The model is shipped to production

New training data is collected

The prediction is compared to the correct answer and the error adjusts the model

The model deletes incorrect memories

✓ The feedback loop: predict → compare → adjust. This is repeated millions of times.

✗ The next step is comparing the prediction to the right answer, calculating the error, and adjusting the model accordingly.

Q2 of 4

Which type of ML is most like training a dog with treats?

Supervised learning

Unsupervised learning

Reinforcement learning

Transfer learning

✓ Reinforcement learning uses rewards and penalties — just like treat-based dog training.

✗ This describes reinforcement learning — the agent takes actions and receives rewards or penalties to shape its behavior.

Q3 of 4

What is "supervised" about supervised learning?

A human watches the model train in real time

The training data has labeled examples with correct answers

The model only works when supervised by an engineer

The learning is restricted to safe topics

✓ "Supervised" means the training data comes pre-labeled with the right answers — like learning with an answer key.

✗ "Supervised" refers to labeled training data — each example comes with the correct answer already attached.

Q4 of 4

Why do engineers call trained models "black boxes"?

They're stored in black server racks

They only work in dark environments

You can see inputs and outputs, but the internal decision-making is hard to interpret

The source code is secret

✓ Models learn patterns too complex to trace — you can see what goes in and out, but the 'why' is often opaque.

✗ 'Black box' refers to interpretability — the internal workings are too complex to easily explain, even to engineers who built it.

🧩

Exercises & Worksheets

Apply what you learned

Draw the Training Loop

Without looking at the diagram, sketch the 5-step training loop from memory. Label each step. Then compare yours to the diagram — what did you remember? What did you miss?

🎨 Visual

Classify the Learning Type

For each scenario, identify which type of ML it is: (a) A model grouping customers by buying habits with no predefined categories. (b) A chess AI learning by playing millions of games. (c) An email classifier trained on 100,000 labeled spam/not-spam examples.

📝 Categorization

Design a Training Dataset

You want to build an AI that identifies ripe vs. unripe bananas from photos. What data would you need? How many examples? What labels? Write a brief training data plan in 5–8 sentences.

🧪 Design

Your Own Analogy

The training loop was compared to a feedback cycle. Create your own analogy using something from your personal life that covers: data, prediction, error correction, and improvement over time.

💬 Creative