LATENT LIBRARY DATA Data for AI Collecting, Cleaning and LabellingData That Models Can Learn From LATENT LIBRARY EDITORIAL

Data for AI

Collecting, Cleaning and Labelling Data That Models Can Learn From

Latent Library Editorial

IntroductionIntroduction

Ask experienced machine-learning practitioners what made their most successful project work, and few will name an algorithm. They will talk about data: the moment they realised the labels were inconsistent, the leak they found just before launch, the extra month spent collecting examples of the rare cases that mattered most.

Models learn only what their data shows them. A brilliant architecture trained on biased, noisy or unrepresentative data produces a confident, fast, wrong system. A simple model trained on clean, well-defined, representative data often performs remarkably well. Yet data work is rarely taught with the same care as algorithms.

This book is a practical guide to the full life cycle of data for AI:

Each chapter ends with a checklist, and exercises with answers and a glossary close the book. No advanced mathematics is required.

Chapter 1Why Data Decides Everything

The model is a mirror

A trained model is, in effect, a compressed summary of the patterns in its training data. It cannot know about situations the data never showed, cannot correct systematic errors in the labels, and will happily learn accidental patterns that happen to predict the label.

Three classic failures illustrate this:

  1. Missing situations. A defect-detection camera trained on images from a factory's day shift performs poorly at night, when lighting differs. Nothing was wrong with the model; the data never showed night conditions.
  2. Inconsistent labels. Two teams labelled customer complaints using different ideas of "urgent". The model learned a blurry average of both, satisfying neither.
  3. Shortcut patterns. A model meant to recognise a medical condition learned partly from incidental markers in images that correlated with where patients were scanned. It scored well in testing and would have failed elsewhere.

Data-centric thinking

When a model underperforms, the instinct is to try a bigger model or tune settings. Often the better first question is: what is wrong with the data? Common answers:

The cost curve

Fixing data problems is cheapest early. A clear label definition written on day one costs an hour. Discovering inconsistent labels after 50,000 examples have been annotated costs weeks. Discovering them after deployment costs trust.

Checklist

You have reached the end of the free sample.
The full book continues with 14 more chapters, plus everything else in its 80 pages.

Get the full book