Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

junyanz.github.io

Post reply on HN
Print view View original

71–80 of 146 posts

Re: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

#71

Post by oconnore »

fritzo wrote: post #46
thriftwy wrote: post #6

The next step is "turning pencil drawings into photos" and using it to fabricate evidence on grand scale. Why bother catching politicians doing something when you can just draw them in? Will wreak havoc on societies with weak politics/reporting culture.

This will hurt journalists reporting on facts more than it will hurt targets of smear campaigns. Journalist: "Here is a compromising photo of a politician." Politician: "Here are 1M photos of every politician doing every imaginable illegal act. Prove that your one photo is not similarly fabricated."

Good journalism is typically not dependent on "gotcha" images. It might hurt amateur twitter reporting, though.

Re: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

#72

Post by oevi »

rattray wrote: post #70

Why is it that Neural Net-based ML only seems to be claiming results with images and natural language? Maybe I'm out of the loop, but I haven't seen anything demonstrating results on "data" – the kinds of challenges that are actually valuable to businesses. Why is that? Are those just less sexy / more proprietary in nature, or is there something about those challenges that make NN's less useful to them?

Because this is B2B business and usually done by specialized software companies. These companies do not publish or open source their solutions because it is either against their own or their customer's interest.

Examples are the whole Predictive Maintenance sector, the medical sector (Computer Aided Diagnosis) or insurance companies which use NN to for all kind of analyses.

Re: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

#73

Post by habitue »

soVeryTired wrote: post #50

I've only recently started reading about deep neural networks, and the thing that strikes me the most about the literature is the lack of mathematics . Open a NIPS paper from 2010 or so, and you'll see extremely dense mathematics: nonparametrics, variational approximation, sampling theory, riemannian geometry. But from my (admittedly small) sampling of the convnet / RNN literature there really doesn't seem to be much…

So what we're seeing is these fields pre-deep learning were mathematical disciplines making steady progress on well-understood foundations. Then deep learning came in, was exceptionally effective at problems that had been difficult to crack, and people shifted focus because it seems weird to be diddling around with incremental gains on techniques that are significantly less effective.

What this created though, was a clean break from a field with a theoretical foundation like what we expect in CS usually, to an empirical foundation, like what we see in psychology, sociology, medicine etc. The theory will eventually catch up, but the fact that "it works" really trumps having a full understanding of why it works (at least for now).

Re: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

#74

Post by visarga »

rattray wrote: post #70

Why is it that Neural Net-based ML only seems to be claiming results with images and natural language? Maybe I'm out of the loop, but I haven't seen anything demonstrating results on "data" – the kinds of challenges that are actually valuable to businesses. Why is that? Are those just less sexy / more proprietary in nature, or is there something about those challenges that make NN's less useful to them?

It's not just images, audio and text. There's also behavior (reinforcement learning), such as in Atari games and AlphaGo.

Re: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

#75

Post by 317070 »

rattray wrote: post #70

Why is it that Neural Net-based ML only seems to be claiming results with images and natural language? Maybe I'm out of the loop, but I haven't seen anything demonstrating results on "data" – the kinds of challenges that are actually valuable to businesses. Why is that? Are those just less sexy / more proprietary in nature, or is there something about those challenges that make NN's less useful to them?

First of, there is. These NN are good at exploiting the 'spatiality prior' in some types of data, like text and images. It means that features in the data which are close together, should be combined when you climb in the hierarchy of features. Databases with columns and rows don't have that prior for instance.

Second, there is also the peer reviewing problem. You are still trying to explain a very abstract concept to your peers in a paper which is usually limited to 6 or 8 pages. Text and images make for very graspable examples in such a short paper. That's the reason why some other data with a spatial prior is not used as often, like time series or EEG-data.

So, there is a combination of those two elements at play.

Re: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

#76

Post by eriknstr »

Current title: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

Current link: https://junyanz.github.io/CycleGAN/

Previous title: Berkeley's software turns paintings into photos, horses into zebras, and more

Previous link: https://github.com/junyanz/CycleGAN/

The previous title, which was based on the repo description of the previous link, was much more informative to me.

Re: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

#77

Post by eriknstr »

kore wrote: post #24

This has also been making the rounds, Deep Photo Style Transfer: https://github.com/luanfujun/deep-photo-styletransfer

I inquired about the license for that one but unfortunately they can't allow commercial use.

https://github.com/luanfujun/deep-photo-styletransfer/issues...

Re: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

#78

Post by habitue »

27182818284 wrote: post #45
baq wrote: post #4

it feels like there's a new deep learning paper each week, ever so slightly bringing me closer to an existential nervous breakdown.

I started buying AI textbooks circa 2010 anticipating something was about to change dramatically. Sadly, I didn't work through the books like I had planned at all. Lacked the discipline to come home and work through them at the time. I really regret it now as things are blowing up and, as you said, it seems like every week there is something new and interesting

Luckily, deep learning is built on super complicated mathematical foundations, but the practice isn't. You could quickly get up to speed with a library like Keras and be productive, building things and trying them out without needing years of theoretical training. I'd highly recommend looking into it, most of the papers are relatively easy to read, and rely largely on empirical results and intuition, rather than deep theoretical proofs or the like.

Re: Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

#79

Post by reader5000 »

habitue wrote: post #73
soVeryTired wrote: post #50

I've only recently started reading about deep neural networks, and the thing that strikes me the most about the literature is the lack of mathematics . Open a NIPS paper from 2010 or so, and you'll see extremely dense mathematics: nonparametrics, variational approximation, sampling theory, riemannian geometry. But from my (admittedly small) sampling of the convnet / RNN literature there really doesn't seem to be much…

So what we're seeing is these fields pre-deep learning were mathematical disciplines making steady progress on well-understood foundations. Then deep learning came in, was exceptionally effective at problems that had been difficult to crack, and people shifted focus because it seems weird to be diddling around with incremental gains on techniques that are significantly less effective. What this created though, was a…

Also, deep net problems are different than traditional CS because we can mathematically define when say a list of numbers is sorted but not when "this image looks photorealistic".

Effectively, deep nets are good at tricking human visual cortex, so I'm not sure theres a deeper mathematical reason they work well, in the same way that the reason mp3 compression works so well is that they tested it against human hearing.

But yes, all deep learning is based on gradient descent which is a greedy heuristic algorithm and it's not a priori apparent it should do anything interesting. But it does.

Post reply on HN