DAY 2 · HOW AI LEARNS

Blind Spot Hunter

Three apps. Three gaps in their training.
Can you spot where they will go wrong?

3 apps · About 9 to 12 minutes

How to play

Four cats sitting indoors on a wooden floor, four rabbits sitting on grass, and one cat sitting on grass with a question mark above it.

From Day 2: a model can struggle with examples unlike its training data—like the cat outside in the lesson. Find these gaps, called blind spots, in three apps.

1

Look at what every example in the training data has in common.

2

Predict which new examples the app will get right or wrong.

3

Run the test, then name the missing kind of example.

App 1 of 3Score 0
THE TRAINING DATA

Look for a pattern: what is the same in all the training examples?

App 1 of 3Score 0
THE TEST

Peek at the training data again
Choose a prediction for all 6 examples.
App 1 of 3Score 0
NAME THE GAP

Step 3. A blind spot comes from a missing kind of example. Which one was missing?

Hunt complete!

TestTrying a model on new examples it wasn't trained on.
AccuracyHow often a model gets the answer right.
GeneralizeTo handle something new correctly because it's close enough to the examples.
Blind spotA gap where a model does badly because nothing like that example was in the training data.

A model can generalize to what is close.

When it fails, name the gap in its examples, not the fault in the machine.