Hacker News Reader

Comment by mjsweet | Hacker News Reader

Just out of interest, could a GAN be used to clean up video, filling in all the blanks and creating a higher resolution, cleaner image and of course colouring too? From what I have seen of moving GAN images, they can jump around and be a bit jittery.

View on HN

It’s been actively worked on, you can see plenty of samples on YouTube. Like many ML applications, it sort of works, but the quality leaves plenty to be desired.

Yes, sort of: You can use a GAN to do almost everything described in the article. That's really interesting in a way, because they mention several distinct steps, and all of them seem readily amendable to machine learning, including the totally-unrelated-to-coloring lip reading they did to add sound, etc.

BUT: reading the article with ML in mind already, I noticed the frequent references to their work with historians and other experts to get it right.

In other words: ML could possibly get you results that are entirely plausible, yet wrong. They mention the color of buttons on the uniforms and, even less likely to be amendable to technology, commands and other speech shouted off-screen.

Probably the best solution would be a combination of the two: the ML algorithm would also learn from human examples and corrections.

I think this will be the solution in a lot of problems and industries: ML does the boring and repetitive work, and a human sanity-checks it and fixes the edge cases where ML is 'dumb'

Reply on news.ycombinator.com