A computer vision system trained on synthetic car images from a video game will probably not work well even if the images look realistic. If 100,000 training examples all come from about 20 car designs, the system will overfit to those car designs and fail to generalize well to dev/test sets that include other car designs.

Google

When synthesizing data, one should consider whether the synthesized examples are representative. Synthesized data should avoid properties that make it possible for a learning algorithm to distinguish synthesized from non-synthesized examples, such as all synthesized car data coming from 20 car designs or all synthesized audio coming from only 1 hour of car noise.

Representative Synthesized Training Examples

Machine Learning Yearning, a free ebook from Andrew Ng, teaches you how to structure Machine Learning projects.
https://www.deeplearning.ai/machine-learning-yearning/

Note: The content of the book is aligned with the Coursera Deeplearning.ai specialization.  https://www.deeplearning.ai/deep-learning-specialization/ 

Machine Learning Yearning (Deeplearning.ai)

If a synthetic dataset repeats the same hour of car noise, it is possible for a learning algorithm to overfit to that hour of car noise and generalize poorly to a new audio clip with different car noise. Even 1,000 unique hours of car noise may cause a similar problem if they come from only 10 cars, because the algorithm may overfit to those cars and perform poorly on audio from a different car.

Learn Before

Related