合成数据Synthetic Data
Synthetic Data
用程序或模型造出来喂训练的数据：省标注钱，但有自己的毛病。Data generated by programs or models to train on: saves labeling money, but has its own problems.
合成数据指由仿真程序或模型生成、而非真人真事采集的训练数据。
它兴起于视觉领域：3D 场景渲染能造出像素级精确的标注，省下数十万美元人工。
核心难题是「合成到真实」的迁移：合成数据不完全像真实数据。
大模型时代，模型自己生成的文本也越来越多地用作训练数据。
Synthetic data is training data generated by simulation software or models rather than collected from real people and events.
It rose in computer vision: 3D-rendered scenes yield pixel-perfect labels at a fraction of manual cost.
The core problem is synthetic-to-real transfer: synthetic data does not fully resemble real data.
In the LLM era, model-generated text is increasingly used as training data.
模型用模型造的数据训练，效果与风险都在研究之中。听到「我们用了合成数据」，要问迁移验证做了没有。Training models on model-made data is an active area of risk and benefit. When a vendor says 'we used synthetic data', ask whether transfer validation was done.
回到名词地图看第1站的其他 34 个词 →Back to the map: the other 34 terms in station 1 →
- Nikolenko《Synthetic Data for Deep Learning》（综述，156 页）Nikolenko, Synthetic Data for Deep Learning (survey) ↗