Glossary
What is synthetic data
In short
Synthetic data is examples created by a program or a neural network, not collected from real life: texts, tables and images used to train and test models.
Machine learning · synthetic data
MMOOZZGG.com editorsChecked: 1 min read
For a child
Imagine a teacher who wants the class to practice problems about a shop, but she has no real receipts. She makes up similar ones herself: so many apples, such a price. The problems are invented, but they teach for real. A computer also sometimes learns from invented examples. They just have to look like real life, or it will learn the wrong thing.
In plain words
There are often too few real examples: rare breakdowns happen once a year, and real customer details cannot be shown to outsiders. Then examples are created artificially. A program draws thousands of photos of cracked parts from different angles. A language model writes customer questions in different words. A table generator produces records that look like a customer database, but with no real people. The model trains on these examples, but it is still tested on real cases.
For business
Synthetic data helps in three cases: real examples are few, they are expensive, or they hold personal data that cannot go to a contractor or to the cloud. For example, to test a chatbot, a team generates hundreds of request variants instead of exporting real chats. The cost is the work of the generator model plus people checking the result. The risks: artificial examples repeat the generator’s errors and biases and have few rare cases, and a model trained only on them handles real life worse. A generator trained on real records sometimes reproduces parts of them, so synthetic data is also checked for leaks. The rule is simple: train on a mix, and do the final test on real examples. The legal side is covered together with the rules of 152-FZ.
For a specialist
Synthetic data is produced by procedures, by simulation or by a generative model: a GAN, a variational autoencoder, a diffusion model or a language model. For tables, the goal is to keep the distributions and the links between columns without rows of real people. In computer vision, 3D scenes are rendered, and the labels come automatically. In training language models, it is common for the model to generate instructions itself. In the Self-Instruct method (Wang et al., 2022), the model created instructions, inputs and answers, duplicates were filtered out, and then the model was fine-tuned on the result. Shumailov et al. (2023) described a limit: when trained on generated content, models slowly lose rare events from the tails of the distribution. This effect is called model collapse. Synthetic data does not make data anonymous by itself: a generator can memorize training records. So for sensitive data teams add differential privacy and a check of closeness to the source rows.
Example
An insurance company wants to teach a model to find suspicious claims, but it has few real fraud cases. Analysts generate artificial claims based on known schemes and add them to the training set, but they check quality only on real cases.
Related terms
Questions and answers
Can synthetic data replace personal data?
Sources
Not sure whether your company needs this?
AI solutions map