← Back to course
AI Explorers 3-5 / Week 06 / Thursday
4/6
Week 06 · Training Data

Thursday

More data, better patterns
// Learning from labeled examples
⏱ about 15 min

Thursday: More Data, Better Patterns

Rocket tries a shortcut. He trains a new model with only two tool photos: two hammers.

Then he shows it a ruler. The model has never seen a tool with no handle!

"What do you notice?" Raven asks. "It only knows hammers."

Nova blinks. "Here is a hint," she says. "How many kinds of tools are there?"

Machine learning needs lots of data

Machine learning needs a huge number of examples.

With only a few examples, a model can miss important patterns.

Rocket's two-photo model never saw a tool with no handle. So it could not learn that.

More examples, of many different kinds, help a model learn better patterns.

ModelTraining photosKinds of tool photos seen
Rocket's shortcut model2hammers only
The crew's model16tools with handles and tools without
WHICH MODEL LEARNS MORE?
  • Read the question.
  • Tap your answer.
Which model saw more examples?
Which model could learn about tools with no handle?
How many more training photos does the crew's model have?

Labels must be right

The labels in training data must be right.

Imagine someone labels a rock photo as plant by mistake.

Then the paper model might learn that lumpy things can be plants.

People check their labels carefully. Mistakes in the data teach wrong patterns.

StatementTrue or false?
Machine learning needs a huge number of examples.?
Two examples are plenty for any model.?
A wrong label can teach a wrong pattern.?
Examples of many different kinds help a model learn.?
WHY THIS EXERCISELots of correct, varied examples help a model find good patterns.
FIX THE TRAINING DATA
  • Read the question.
  • Tap your answer.
The model keeps missing tools with no handle. What is the best fix?
A rock photo is labeled plant. What should the crew do?
MAKE GOOD TRAINING DATA
  • ?Let the model find patterns.
  • ?Collect many examples.
  • ?Add a label to each one.
  • ?Check every label carefully.
  • ?Include many different kinds.
WHY THIS EXERCISEGood training data is big, varied and labeled correctly.
Rocket's shortcut model had how many training photos?
A wrong label can teach a wrong ____.
Does a learning model need a few examples or a huge number? Type few or huge.
WHY THIS EXERCISEMachine learning needs a huge number of examples.
Try it
On paper, draw two cats that look alike. Show them to someone and say "cat."
Now draw a very different cat. Would two examples be enough to know it?

Excellent work! Tomorrow we review training data and the people behind it.

← Wednesday