Data-centric AI
Data-centric AI is an approach to building AI systems that treats the systematic engineering of training and inference data, rather than model architectures and hyperparameters, as the main lever for improving performance, a discipline Andrew Ng defines as “systematically engineering the data needed to successfully build an AI system.”[1][6] It is organized into three missions (training data development, inference data development and data maintenance) and measured through benchmarks such as MLCommons’ DataPerf.[3][4][5]
Origin and terminology
The term was propagated by Andrew Ng and promoted through a series of workshops; the paradigm emphasizes improving data systematically and at scale.[2] In a 2022 interview, Ng described a “data-centric AI movement” that he said could yield “small data” solutions to problems of efficiency, accuracy and bias.[1] LandingAI, the company Ng leads, restates the definition on its website.[7]
The available sources use “propagated” and “promoted” rather than stating who first coined the phrase, so its exact origin is not settled here.[2]
Comparison with model-centric AI
In the model-centric view, practitioners improve results by modifying model architectures and tuning hyperparameters. In the data-centric view, they improve results by refining and engineering the training data.[6] See also concept/data-labeling.
Frameworks and taxonomy
A 2023 position paper by Zha, Bhat, Lai, Yang and Hu, accepted to the SDM 2023 Blue Sky Track, organizes data-centric work into three missions:[3]
- Training data development
- Inference data development
- Data maintenance
The same authors’ survey in ACM Computing Surveys uses these three goals, reviews benchmarks across application domains, and discusses the blurring line between data and models.[4] The position paper lists open challenges that include data-model co-design, data bias and benchmarks.[3]
Benchmarks and competitions
DataPerf is a community-led benchmark suite for datasets and data-centric algorithms, presented at NeurIPS 2023 in the Datasets and Benchmarks track.[5] Its first iteration contains five benchmarks:[5]
- Vision
- Speech
- Data acquisition
- Data debugging
- Diffusion prompting
MLCommons maintains DataPerf, and its code, platform and baselines are open source.[5]
An earlier Data-Centric AI (DCAI) competition asked participants to improve a dataset while keeping the model fixed. One account places its launch at NeurIPS 2021; this date comes from a secondary description rather than a primary record.[5]
Practical relevance
See ai-data-lab-selection-criteria for operational criteria on selecting external providers of labeled data.
Limitations of the available evidence
- No disagreements between the cited sources were identified.
- The status of DataPerf rounds after the first iteration was not established.
- Market size and adoption statistics were not found in the sources reviewed.