← Back to course
AI Explorers 6-8 / Week 07 / Wednesday
3/6
Week 07 · Three Kinds of Bias

Wednesday

Explorer Lab: The Lopsided Deck
// When a fair-looking score hides a gap
⏱ about 20 min

Wednesday: Explorer Lab: The Lopsided Deck

Rocket fans out twelve blank index cards. "Today we make bias on purpose, so we can see how it works."

Raven raises an eyebrow. "On purpose? Is that allowed?"

"In a lab it is," Rocket says. "We control everything and measure it."

Nova drops a small cup of counters onto the table. "Here is the lazy rule we will test," she says.

"Always guess whatever label shows up most in the training cards."

Raven grins. "That sounds easy. It also sounds like a trap for comets."

"Then measure it and find out," Nova says.

Your lab

You will build a lopsided training deck, test a lazy guessing rule, and measure accuracy for each label.

  • Twelve index cards or paper squares
  • A pencil and colored pencils
  • A cup and ten counters, such as buttons or paper squares
  • An envelope for the test cards
  • A partner, if you can find one
Safety first
Use only paper, pencils and counters for this lab.
Keep small counters away from younger children.
  1. Make a training deck of 8 cards: draw 5 stars, 2 planets and 1 comet. Label each back.
  2. Make a test deck of 4 cards: 1 star, 1 planet, 1 comet and 1 satellite streak. Seal them in the envelope.
  3. Count the training labels. The most common label is the lazy rule's only answer.
  4. Have your partner draw each test card. The lazy rule guesses the same label every time.
  5. Mark each guess right or wrong in the results table.
  6. Now rebalance: add or redraw cards so each label has 2 training cards.
  7. Notice the lazy rule now has no single winner. Write one better rule using the card features instead.
Test cardTrue labelLazy rule guessRight or wrong?
1star
2planet
3comet
4satellite streak
PREDICT FIRST
  • Read the question.
  • Tap your answer.
Your training deck has 5 stars, 2 planets and 1 comet. What will the lazy rule always guess?
How many of the 4 test cards will the lazy rule get right?
Accuracy for the star test card: how many right out of 1?
Accuracy for the comet test card: how many right out of 1?
Which label was never even in the training deck? Type one word: satellite, planet or star.
Lab ideaTrue or false?
The lazy rule was unfair even though nobody meant it to be.?
A label missing from training can still be learned from the data.?
Rebalancing the deck fixed the sample, but you still needed a better rule.?
WHY THIS EXERCISEA fair sample helps, and a good rule must still use real features.

Great lab work. Keep your cards. Tomorrow you meet the other two kinds of bias.

← Tuesday