17+ cv petite enfance
Find this animadversion offensive?
Choose your acumen beneath and bang on the Report button. This will active our moderators to booty action
Reason for reporting:
Your Acumen has been Reported to the admin.
I accept apparent this acquaint so abounding times, and every distinct time it’s explained ailing and it avalanche aback to the “OMG AI IS GETTING SMARTER THAN US AND WILL DESTROY US” meme. Thanks, Hackaday, for not falling into that trap, but still anecdotic it as “cheating” – and heck, alike an AI at all – is still a little clickbait.
It’s not cheating. This is what it’s advised to do (… and I don’t apperceive why you would administer it to artery map generation). It’s advised to do angel translation: to acquisition an algorithm based on commutual photos for how to construe amid the two domains. But CycleGAN is for unpaired angel translation, because you don’t consistently accept nice altogether commutual examples: you ability accept bunches of paintings by Monet, but you don’t accept the photographs area they came from.
The problem, of course, is that unpaired angel adaptation is a poorly-posed problem, because you don’t accept any way to adjudicator how able-bodied you did, added than allurement if the adapted angel looks like the others. As in, if you get a photo, and try to transform it into a Monet, all you can do is ask if it looks like added Monet paintings. Not “does it attending like Monet corrective this from this photo.” And because algorithms are stupid, what they’ll do is aloof accomplish a new algorithm “(input image)0 Monet painting”, and hey, this algorithm transforms aggregate into a absolute Monet painting! (They alarm this ‘mode collapse’ in the CycleGAN paper).
So what CycleGAN does is adds a new constraint: it creates its own pairs, acute that the algorithm be reversible. That way approach collapse doesn’t assignment – the atomic band-aid isn’t reversible.
But because it’s reversible, that agency back you use it for angel abridgement (where one dataset has abundant beneath advice than the other) it agency that the consistent algorithm is affected to bluffing (see amount 5 in the paper). Which is, of course, accessible back you anticipate about it. CycleGAN isn’t an AI, it’s aloof a filter, and it’s about applying a capricious “street map filter” assimilate the image. Which agency you’ve got to adumbrate the advice somewhere.
I’ve got a acrimonious activity that the authors advised the absolute agreement to authenticate this, because I don’t see why you would use CycleGAN for this use at all, because of advance you can alternation artery map bearing from paired examples (which you accept bags of), like pix2pix does, and again the accomplished affair goes away. So I would accept to assumption that this isn’t a case of “OMG AI afraid the authors” but added of them pointing out a limitation of the algorithm. But I could be wrong, in that they might’ve aloof blindly acclimated CycleGAN because they anticipation it would be fun, and again accomplished it’s got this limitation.