Jiachen-T-Wang Data-Centric Artificial Intelligence: A Survey
One-line decision
Use this skill when you need a comprehensive overview of data-centric AI techniques including data quality, augmentation, selection, and engineering. Avoid it when you are looking for a specific method rather than a broad overview.
Skill metadata
- Skill type: data-centric-ai-survey
- Paper kind: survey
- Actionability: medium
- Evidence quality: full_paper
Goal
Provide a comprehensive survey of data-centric AI techniques spanning data collection, labeling, preparation, reduction, augmentation, and pipeline engineering.
Problem signature
- Modality: general survey covering all data modalities.
- Data state: survey of methods applicable to data in any state.
- Scale regime: methods applicable at any scale.
- Model requirement: Survey; no specific model requirement.
Use when
- You need a broad overview of data-centric techniques.
- You want to understand the landscape of data engineering methods.
- You are planning a data-centric approach to ML.
Do not use when
- You need a specific method, not a survey.
- You already have extensive data-centric knowledge.
- You need implementation details for a specific technique.
Required inputs
- research_corpus: Literature on data-centric AI methods.
Optional inputs
- domain_focus: Specific domain to focus the survey on.
Outputs
- taxonomy: Taxonomy of data-centric AI techniques.
- best_practices: Guidelines for data-centric AI development.
Assumptions and prerequisites
- Data quality is as important as model architecture.
- A systematic taxonomy helps practitioners choose methods.
- Data-centric approaches complement model-centric ones.
Procedure
- Review data collection methods Action: Survey techniques for data collection and annotation. Why: Collection is the first step in the data pipeline. Note: See paper for details.
- Review data preparation Action: Survey cleaning, normalization, and transformation methods. Why: Preparation improves data quality. Note: See paper for details.
- Review data augmentation Action: Survey augmentation and synthesis techniques. Why: Augmentation increases effective data size. Note: See paper for details.
- Review data selection Action: Survey subset selection and active learning methods. Why: Selection focuses resources on valuable data. Note: See paper for details.
Parameters to set
- scope — Role: Breadth of the survey. How to set: Cover all major data-centric areas. Default/range: Comprehensive. Effect: Broader scope provides more guidance.
Validation checks
- The survey should cover all major data-centric techniques.
- Techniques should be organized in a usable taxonomy.
- Best practices should be actionable.
Failure modes
- The survey may not cover the latest methods.
- General guidelines may not apply to specific domains.
- The breadth may sacrifice depth.
Adaptation notes for VLM training
- Use the taxonomy to plan VLM data curation strategies.
- Apply data-centric principles to multimodal data engineering.
- The survey provides a framework for systematic data improvement.
Implementation notes
- Use the taxonomy as a checklist for data pipeline design.
- Refer to cited papers for implementation details.
- Update the survey with new methods as they appear.
Evidence from the paper
- The survey provides a comprehensive taxonomy of data-centric AI techniques.
- Data-centric approaches are increasingly recognized as essential for ML success.
- The survey covers collection, preparation, augmentation, selection, and engineering.
- Best practices are synthesized from extensive literature review.
Source paper
- Title: Data-Centric Artificial Intelligence: A Survey
- Year: 2023
- Venue: arXiv
- Paper ID: arxiv-2303.10158v3
- URL: http://arxiv.org/abs/2303.10158v3
- arXiv ID: 2303.10158v3