Synthetic data is data generated by models or by manipulating real data. It can be useful for training deep neural networks when there is insufficient real data or real data is hard to obtain. For example, 3D simulations can be used to emulate near accident situations, such as pedestraians crossing busy roads or in the dark; these will hopefully be rare in real gathered video data, but invaluable to help autonomous cars respond well to extreme events. Synthetic data can also help generalisation, for example by adding noise to images or addings cropped, rotated or resized copies. Synthetic data has also be sggested as a way to reduce bias in data sets.
Used in Chap. 8: page 105; Chap. 18: page 280; Chap. 24: page 375