"Training datasets are incomplete operational representations of the underlying world, and their construction shapes what machine-learning systems can learn and evaluate."
Hardt and Recht explain that datasets do more than supply examples for algorithms - they help formulate the machine-learning problem itself and establish the measurements by which success is judged. Their analysis of dataset construction, representation and validity supports the book's map-versus-territory argument and the claim that choices made before training can determine which aspects of reality become visible to a model.
View Source