Stage 2 · Intermediate · F6
Data Is the Model
The network that started the deep learning boom learned from 1.2 million photos, each one labeled by a person.
12 lessons · 118 minEasy
About this chapter
Change the data and the model changes more than any tweak to the network can manage. This chapter turns tables, photos and sentences into rows of numbers, then cleans them without lying about them. You will meet leakage, the bug that makes a model look brilliant until the day it ships, and the rare class, the reason 99% accuracy can be worthless. It ends with the human side: who labeled this, what it cost, who agreed to be in it, and what the card that ships with the data should say.
What you will be able to do
- 19 min
Rows, Features, Labels
Read any dataset as rows of features with one column of answers.
- 29 min
Pictures and Words as Rows
Say how many numbers a model reads for one image and for one sentence.
- 310 min
Where Labels Come From
Name where a label came from and what that origin costs you.
- 410 min
Cleaning Without Lying
Fix duplicates, wrong units and impossible rows, and write down every change.
- 510 min
The Holes in the Table
Choose between dropping, filling and flagging, based on why the value is missing.
- 610 min
Three Piles
Split rows into three piles and say what each one is allowed to do.
- 710 min
The Answer That Slipped In
Find target leakage before it makes a model look better than it is.
- 811 min
When Tomorrow Is Different
Recognize distribution shift and choose what to monitor after launch.
- 910 min
The One Case Accuracy Hides
Split, rebalance and test a table where the class you care about is rare.
- 109 min
More Data Without New Facts
Choose augmentations that keep the label true for the task.
- 1110 min
Who Is in the Table?
Check who is represented, who consented and who the dataset leaves out.
- 1210 min
The Card That Ships With the Data
Write a dataset card that states source, consent, gaps and intended use.