← AI Terminology
Data-Centric AI
Data-centric AI emphasises systematically improving datasets (labels, cleaning, coverage) as the primary lever for performance, rather than only chasing new models.
Advocated by Andrew Ng and others as an engineering discipline.
Advocated by Andrew Ng and others as an engineering discipline.
Why It Matters in AI
Many production wins come from fixing labels and slices, not bigger nets. Data-centric practice pairs with MLOps: version data, measure slice metrics, close feedback loops. Complements model-centric research.
Key Points
| Aspect | Description |
|---|---|
| LLM | Curriculum mixtures, preference quality, eval sets |
| Focus | Label quality, consistency, edge cases, balance |
| Culture | Dataset owners as first-class roles |
| Related | Active learning, synthetic data, MLOps |
| Practices | Error analysis → dataset patches → retrain |
| Vs model-centric | Architecture search secondary if data dirty |
Simple Analogy
Tuning a race car less than fixing the fuel and track conditions — better inputs often beat fancier engines.
Common Usage Examples
- Slice-based error analysis dashboards
- Version datasets like code
- Prefer fixing 500 labels over new architecture
- Document data changes in release notes
Summary
In short: Data-centric AI improves systems by improving data quality and coverage first — often higher ROI than endless model tweaks.