,
9 tweets,
3 min read
Read on Twitter
Impressive! There has been significant progress in GANs over the past few years, but that's not really what we're seeing here... [1/8]
(Thread. Buckle up! 👇🔥)
(Thread. Buckle up! 👇🔥)
If you were to include images in a Tweet that promoted progress specifically in GAN research (relativistic discriminators, wasserstein constraint, unrolled steps, spectral norm, etc.), it'd look like this... Not quite so catchy! 🤓 [2/8] 
It's like a different field of research.
The progress in the latest NVIDIA paper was made 100% with domain-specific changes independent of GANs. The authors say so themselves in the introduction: [3/8]
The progress in the latest NVIDIA paper was made 100% with domain-specific changes independent of GANs. The authors say so themselves in the introduction: [3/8]
What makes the latest research shine are the expert insights in combining techniques often used in generation of images: coarse-to-fine approach, best-in-class neural style architecture, and an amazing new dataset! [4/8]
Attributing this progress to GANs doesn't help us as a community — let alone the general public. It's the new #DeepLearning buzzword, along with the shallow understanding that comes with it. 🤔
It's our job as scientists to communicate this better on social media. 📣 [5/8]
It's our job as scientists to communicate this better on social media. 📣 [5/8]
The "adversarial" part becomes much less important in cutting-edge applications. I'd bet you can achieve similar results with different algorithm, expert-crafted loss functions, or even statically learned losses. (I'd gladly advise that project!) [6/8]
I still believe GANs are the second best problem for every application. As soon as you understand what you're doing, you eventually & naturally switch to higher-quality expert crafted losses (which may be learned). 👩🔬 👨🎨 [7/8]
Research in GANs is of course important: it will make the second best solution closer to what an expert would come up with. Today we're seeing significantly more domain-specific progress as we integrate decades of knowledge & best practices in each field. Exciting times! ✨[8/8]
It would be fascinating to take NVIDIA's data+code (once they release it) and perform an ablation study on specific GAN features back to mid-2014, and see what difference it makes. How would you objectively measure the results? [9/8]
