Train a classifier on bright photos of one thing and dark photos of another, and it will learn the wrong lesson perfectly. Today you watch it happen — and diagnose it before your teacher does.
On this page: AP-TESTABLE sections marked with a BI5 tag · CONTEXT — NOT ON THE AP EXAM anything with this label. Exam-tagged questions never test the context material.
No reading — this page is the reading. Retrieve 5.10: who was missing from the dataset? Today is the consequence of that answer.
Impact, rooted in Big Data. Unit 5 asked who’s missing from a dataset. A model trained on it inherits the gap.
Before anyone explains: what did it actually learn?
CONTEXT — NOT ON THE AP EXAM
A model can only learn patterns that are in its training set. It has no idea what was left out. If every apple in training was brightly lit and every orange was in shadow, the easiest pattern to learn is “bright = apple” — and the model will learn it, confidently. It will do great on photos like its training data and fail on anything else. Working on new, different data is called generalization; memorizing the quirks of the training set instead is (informally) overfitting.
AP-TESTABLE
Back to 5.10’s pothole app: reports came mostly from neighborhoods where people had the app. A model trained on those reports would learn that potholes happen in wealthy neighborhoods — and send repair crews accordingly.
CONTEXT — NOT ON THE AP EXAM
A classifier tells apples from oranges using two numbers per photo: color (1 = red, 3 = orange) and brightness (3 = shadow, 8 = bright). It predicts whichever fruit’s training average is closer. Pick a training set, then look at the test photos.
Five questions. Find what the data taught — not what the builders meant.
Pick a task (“is this a dog or a wolf?”, “is this handwriting a 1 or a 7?”). List what your training photos would show. Swap with a partner and find one kind of input their set is missing. The missing input is where their model would fail — and who it fails.