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

Explorer Quest

Explorer Quest: Train a family member
// Learning from labeled examples
⏱ about 15 min

Explorer Quest: Train a Family Member

Rocket wants to train Raven, not a model. He picks a secret label: "Zoomy."

He shows her a ball. "Zoomy." A rug. "Not zoomy." A toy car. "Zoomy."

"What do you notice?" Raven says. "The zoomy things can roll!"

Nova cheers from above. "Here is a hint for next time," she says. "Try fewer examples. Is it harder?"

  1. A label is the right answer attached to an example.
  2. Training data is a big set of labeled examples.
  3. Machine learning finds patterns in training data.
  4. It needs a huge number of examples, usually supplied by people.
  5. Wrong labels or too few kinds of examples can teach wrong patterns.
◇ TRAIN A FAMILY MEMBER
The trainer picks a secret label, like "things that roll" or "things made of paper."
The trainer shows one object at a time and says "yes" or "no."
Use 8 labeled examples. The learner can watch but not guess yet.
Then the trainer holds up 4 new objects. The learner predicts yes or no.
Count how many predictions were right.
Play again with only 2 examples. Compare the results.
RoundTraining examplesNew objectsRight predictions
184
224
For a grown-up
Use safe household objects only. Nothing sharp or breakable.
Ask: did 8 examples or 2 examples work better? Why?
Ask: what happened if the trainer gave a wrong yes or no by mistake?
QUEST CHECK
  • Read the question.
  • Tap your answer.
In the game, the trainer saying "yes" or "no" is like adding a...
Which round should be easier for the learner?
The learner got 3 of 4 right. How many were wrong?
WHY THIS EXERCISECounting right and wrong predictions shows how well the learner was trained.
← Friday