4 ms·
There is a mantra in ML that has been around for a while. It's that when training on synthetic data, your learned model is only as good as your generator model.
by Scene_Cast2 2y ago
There is a mantra in ML that has been around for a while. It's that when training on synthetic data, your learned model is only as good as your generator model.
- etiam 2y agoCatchy! And a really good point. Seems like there could be room for a couple of special situations with caveats though? With the GAN formulation your generator can be practically as good as your discriminator and your discriminator can probably be better than it would have been without adversarial regularization?
- davidguetta 2y agoisn't it obvious ? synthetic data was nice to generate an underlying distribution and avoid jumps in 'data holes', but with methods like diffusion now it's not as necessary