Responsible AI practices
Responsible Preparation for Datasets
Responsible models need balanced datasets that represent every relevant group. Balance matters most in high-stakes areas like hiring, lending, and criminal justice. Tools: SageMaker Clarify to find imbalance, SageMaker Data Wrangler to fix it.
Two steps to balanced data
1. Inclusive and diverse collection
Collect from a diverse range of sources, viewpoints, and demographics. A model trained mostly on middle-aged people will be less accurate for younger and older people. Diversity matters for any topic, not only data about people.
2. Data curation (labeling, organizing, and preprocessing)
- Preprocessing: cleaning, normalization, and feature selection to remove bias and errors.
- Augmentation: generate new examples of underrepresented groups when there isn't enough real data.
- Regular auditing: keep checking that the data stays balanced and fair.
Balance for the intended use case
Balance depends on the purpose. A system about cancer in children should be built on data about children, not adults.