3 ms·
It doesn't seem to be a very good idea to decide that since the discriminator cannot predict the output of the generator that prediction is impossible. I have b
by airza 4y ago
It doesn't seem to be a very good idea to decide that since the discriminator cannot predict the output of the generator that prediction is impossible. I have been chipping away at improving ML prediction of non-CSPRNGs for the last few years and prediction is very sensitive to hyperparameter choice (thus far anyway). It seems easy to imagine a PRNG produced in this way that seems usable.. until a more sophisticated network learns to predict it.
- sdenton4 4y agoIt would be interesting to try to turn the randomness tests themselves into additional loss functions, supplementing the GAN loss. Production-ready GANs in my experience are trained with a lot of auxillary losses. For example: HiFi GAN (for audio generation) uses four different discriminators to detect problems in the spectrogram and fine structure of the waveform, and the losses are more about matching the discriminator features than the actual classification result. Works great, but is hilariously 'impure.'
- shmageggy 4y agoIs your work published?