What Is Machine Learning?

Step 1: Rules versus examples

How would you teach a computer to recognize a cat in a photo? You might try writing rules: “has pointy ears, has whiskers, has fur…” But a cat can be sleeping, hiding or half out of the frame, and the rules quickly fall apart.

There is another way. Show the computer thousands of photos and tell it which ones contain cats. Let it find the pattern itself. That is machine learning.

Some problems are better for rules:

def km_to_miles(km):
    return km * 0.621371

print(km_to_miles(10))

6.21371

A formula is exact and always right. Machine learning is for problems where the rules are too hard to write down: recognizing faces, understanding speech, recommending a video.

Step 2: The parts of a machine learning problem

Every ML project has the same four ingredients.

The process is: collect data, train a model on labeled examples, then use it on new data it has never seen.

Step 3: Training and testing

Imagine you study with flashcards. If your test uses the same flashcards, a perfect score doesn’t prove you learned anything, because you may have memorized them. A fair test uses new questions.

Machine learning works the same way. We split our data:

• Training set: examples the model learns from (usually about 70 to 80%).
• Testing set: examples kept hidden until the end, used to check how well it learned.

If a model does great on training data but poorly on testing data, it has memorized instead of learned. We will see this again in Session 5.

Step 4: Teachable Machine

Google Teachable Machine lets you train a model in your browser. Here is the workflow you will follow in the project:

1. Choose Image Project, then Standard image model.
2. Create classes. Each class is one label, like “pencil”.
3. Add samples to each class with the webcam or by uploading images.
4. Click Train Model. The computer looks for what is different between your classes.
5. Use Preview to test the model on something new.

Tips for a good model: Take many pictures. Change the angle, the distance and the background. Include different lighting. The more variety, the better the model learns the real difference.

Step 5: What the model really knows

Your model doesn’t understand pencils. It found numbers that tend to be different between your pencil photos and your eraser photos. If something new has similar numbers, it says “pencil”.

That is why it can be confidently wrong. If you show it a banana, it will still say “pencil” or “eraser”, because those are the only answers it knows. Good ML practitioners always ask: what examples did this model see, and what did it never see?

Think about it: If you trained a model only with photos of people with one hair color, how might it behave with someone different? We will come back to this in Session 5.