CONCEPT
Data-Centric AI
Andrew Ng’s campaign to invert the prevailing emphasis of the machine learning field: rather than holding the data fixed and improving the model—the dominant research paradigm—hold the model fixed and systematically improve the data, because in most real-world deployments the dataset is the decisive variable that the field has trained itself to disdain.
For most of the modern history of machine learning, prestige and attention flowed to the model. Researchers competed to design more sophisticated architectures, progress was measured by improvements on benchmarks everyone held fixed while iterating on the algorithm, and the overwhelming majority of papers were model-centric in exactly this sense.
Andrew Ng spent years arguing that this emphasis is, for most real-world applications, precisely backward. The data-centric AI campaign proposes the opposite: hold the model fixed and systematically improve the data—acquire more of it where it is thin, better of it where it is noisy, and remove what is mislabeled, ambiguous, or irrelevant. Engineering the data with the same rigor the field had reserved for the model typically delivers larger performance gains than any architectural tweak, especially in the settings where AI actually has to work: a hospital, a factory, a bank, with