The best way to make a machine learning model generalize better is to train it on more data… But only have a limited amount of data!
The solution is create fake data to add to our dataset.
This strategy has been effective in object recognition, for example you can flip and rotate objects, change color, rescale, etc…
