Artificial Datasets; useful for Ai purposes, what about Ai applications in other fields of science?

Question

In Artificial Intelligence, it's common to create sample 'fake' datasets and use them for the purpose of making more efficient algorithms from classification to regression. Datasets with data points distributed in different patterns, which can make the problem harder for Ai systems.

In some other areas of science, there are very few numbers of data points, measured from nature and gathered in a dataset. The lack of enough data points will make the job of producing more accurate Ai systems more challenging.

So.

Q: Does making new 'Fake' data - based on the real ones - useful in training better Ai systems?

I think we must take a look at this question from the perspective of the learning theory. Is it possible that 'fake' data points contain the real truth(or a part of the reality) about the real natural data?

p.s: I'm thinking of a Generator ANN(in fact a GAN) to match the probability distribution function of 'fake' data on the PDF of the 'natural' data.

Answer 1

Does making new 'Fake' data - based on the real ones - useful in training better Ai systems?

Yes

This is called data-augmentation. A common application is image recognition and computer-vision. The task is to recognize when photos contain objects (e.g. cats).

There's obviously no reason that a photo of a cat must always be taken from the same angle, so it's common to train the model on random rotations, flips, translations and crops/rescaling of the original image.

There are other, more esoteric forms of data augmentation such as mixup, which trains the model on convex combinations of pairs of samples. "mixup: Beyond Empirical Risk Minimization" by Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, David Lopez-Paz