Synthetic data is artificially generated data designed to resemble observations from a target domain. It helps teams obtain training and evaluation data when real examples are difficult to collect, then organize the results into reproducible ML datasets. section 1.1.1
Use synthetic data when production data is scarce or expensive to label:
- Generate images and labels procedurally for computer-vision training, avoiding manual annotation of every example. section 2.1.1
- Work with domain experts to create synthetic fraud examples when identifying fraudulent transactions in existing data is laborious. section 1.1.1
- Treat generated data as part of the dataset lifecycle, including organization into training, validation, and test sets and versioning for reproducibility. section 1.1.1
Identity-document detection
An identity-document detector can use procedurally generated images whose labels are generated alongside the images. This supplies labeled visual data without annotating the entire dataset manually. section 2.1.1
