The pedagogy of AI

In the previous lesson, we learned that Machine Learning (ML) is about "teaching" rather than "telling." But just like human students, machines can be taught in different ways depending on what we want them to achieve.

On the AWS Certified AI Practitioner exam, you will be given business scenarios and asked to identify which learning method is being used. Master these three, and you master Domain 1.

---

1. Supervised Learning: The Student with a Teacher

Definition: Training a model on a Labeled Dataset. The computer is given the input (e.g., a photo) AND the correct answer (e.g., "This is a cat").

Think of it like a set of Flashcards.

• Front: Image of a square.
• Back: "Square." The computer looks at millions of these until it can recognize a square without looking at the back.

Key Tasks:

• Classification: Putting things into buckets (e.g., "Spam" vs. "Not Spam").
• Regression: Predicting a number (e.g., "What will this house sell for?").

AWS Example:

Amazon Rekognition landmark detection. The model has been "supervised" (told) what the Eiffel Tower looks like.

---

2. Unsupervised Learning: The Student as an Explorer

Definition: Training a model on Unlabeled Data. The computer is given no answers; it is simply told to "Find a pattern."

Imagine giving a child a bucket of mixed Legos and saying "Organize these." You don't tell them how. They might organize them by Color, by Size, or by Shape. They are discovering structure without a teacher.

Key Tasks:

• Clustering: Grouping similar items (e.g., "Group these 1 million customers into 5 'Segments' based on their spending").
• Association: Finding rules (e.g., "People who buy beer also tend to buy diapers").

AWS Example:

Amazon SageMaker clustering algorithms used to find "Anomalies" (outliers) in network traffic that might indicate a hack.

---

3. Reinforcement Learning: The Student with a Reward

Definition: Learning through Trial and Error based on a "Reward" system. There is no training data at start; the model (The Agent) explores an environment and gets "Points" for good actions and "Penalties" for bad ones.

Think of Training a Dog.

1. Dog sits (Action).
2. You give a treat (Reward).
3. Dog repeats the action to get more treats.

Key Tasks:

• Robotics: Learning to walk without falling.
• Game AI: Learning to win Chess by playing itself millions of times.
• Supply Chain: Finding the "Optimal" path through a complex warehouse.

AWS Example:

AWS DeepRacer. It is a 1/18th scale racing car that you train using Reinforcement Learning to stay on a track.

---

4. Comparison Table for the Exam

Feature	Supervised	Unsupervised	Reinforcement
Data Type	Labeled (Answers provided)	Unlabeled (No answers)	Interaction (Reward/Penalty)
Logic	Mapping Input to Output	Finding hidden patterns	Maximizing a score
Common Use	Prediction / Classification	Discovery / Segments	Decision Making / Games
Memory Hook	"Flashcards"	"Sorting Legos"	"Dog Training"

graph TD
    A[Machine Learning] --> B[Supervised: Labeled Data]
    A --> C[Unsupervised: Pattern Search]
    A --> D[Reinforcement: Reward/Penalty]
    
    B --> B1[Classification: Categories]
    B --> B2[Regression: Numbers]
    
    C --> C1[Clustering: Groups]
    C --> C2[Dimensionality Reduction]
    
    D --> D1[Robotics]
    D --> D2[Game Strategy]

---

5. Summary: Which One to Choose?

When answering exam questions:

• If the question mentions "labels," "answers," or "pre-defined categories," it is Supervised.
• If the question mentions "finding groups," "hidden trends," or "unlabeled data," it is Unsupervised.
• If the question mentions "trial and error," "reward," or "autonomous navigation," it is Reinforcement.

---

Exercise: Identify the Method

You are a researcher at a pharmaceutical company. You have 10,000 chemicals, and you want to see if they can be grouped into "Families" based on their molecular weight and solubility, even though you don't know the families yet. Which learning method should you use?

• A. Supervised Learning (Classification).
• B. Supervised Learning (Regression).
• C. Unsupervised Learning (Clustering).
• D. Reinforcement Learning.

The Answer is C! You are discovering "Families" (clusters) from data where you don't already have the "Answers" (labels).

---

Knowledge Check

What's Next?

We’ve covered the "Methods." Now, let's learn the "Lingo." To speak like a Practitioner, you need the right words. Find out in Lesson 5: Common AI Terminology Explained Simply.