1
$\begingroup$

I have been reading about Generative Adversarial Networks (GANs) and was wondering if it would make sense to train a generator function only to use it for creating more training data.

In a scenario where I don't have enough training data to build a robust classifier, can I use this limited data to train a generator that'll produce samples good enough to improve the accuracy of my discriminator (classifier)?

$\endgroup$
1
$\begingroup$

Yes and no depending on how you define "good enough samples". You will likely end up with a chicken and egg problem: you want to use the GAN to generate training data, but the GAN doesn't have enough training data itself to generate convincing enough samples.

Other techniques exist for data synthesis of training images. For example: adding noise, flipping axis, change luminosity, change color, random cropping, random distorsion.

| improve this answer | |
$\endgroup$
  • $\begingroup$ you're assuming image data. do we have good techniques for text for example? $\endgroup$ – user239558 Sep 14 '17 at 19:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.