|
Freakazoid_ posted:Star trek knew what holodecks were for and they didn't hide it, and neither will we when holodecks become reality. The best thing is the thing where Picard is typing on a typewriter before someone bursts in with plot which implies Picard is content sitting in a hologram pretending to do paper work
|
#
?
Dec 25, 2017 04:43
|
|
|
|
| # ? May 8, 2024 08:23 |
|
|
Speaking about porn holodecks... https://www.youtube.com/watch?v=Ihuv5b6nrbA&t=380s
|
|
#
?
Dec 25, 2017 11:30
|
|
|
Owlofcreamcheese posted:Hey, someone write a scifi story where in the future instead of netflix streaming a video file it just streams some 3D assets and a general script outline for a movie and then your personal filter bubble AI renders a realistic 3D movie out of it on the fly personalized to you the way like google news mostly only shows you news it already knows you'd like. So like, your own personal copy of Juno II does or doesn't go more into the anti-abortion angle depending on what it knows you'd want and you can only really intelligibly talk about movies with people that exist in your own personal sphere, the same way news now basically might as well be from an alternate universe if you talk to people from another political "tribe" and the news they consume. I expect this to appear in games first. Candy Crush could already create levels for me based on my personal preferences. They probably don't do it just because I wouldn't compete directly with other players so there'd be less incentive to purchase boosters. Doctor Malaver fucked around with this message at 17:59 on Dec 25, 2017 |
|
#
?
Dec 25, 2017 13:13
|
|
|
Capitalism.... bad?
|
|
#
?
Dec 25, 2017 17:19
|
|
|
Doctor Malaver posted:I expect this to appear in games first. Candy Crush could already create levels for me based on my personal preferences. They probably don't do it just because I wouldn't compete directly with other players so there'd be less incentive to purchase boosters. One of the resident evil games gets easier if you are bad but hides that that system exists from the player so you don't feel bad
|
|
#
?
Dec 25, 2017 18:09
|
|
|
Owlofcreamcheese posted:One of the resident evil games gets easier if you are bad but hides that that system exists from the player so you don't feel bad
|
|
#
?
Dec 25, 2017 19:50
|
|
|
Will Wright talked a lot about this concept before making Spore. But not just from a difficulty context but the idea that the game could guess the narrative you want and build the game arc to match.
|
|
#
?
Dec 25, 2017 21:21
|
|
|
Trabisnikof posted:Will Wright talked a lot about this concept before making Spore. But not just from a difficulty context but the idea that the game could guess the narrative you want and build the game arc to match. It’s a shame Spore ended up not really having any real gameplay generation, it ended up just being randomized character generation. I wonder how the same concept would work out if they were developing it today or in 5 years.
|
|
#
?
Dec 25, 2017 22:04
|
|
|
Coherent story is hard to do via AI because for understanding basic building blocks of storytelling like justice, redemption, wanderlust and wonder, you'd already have to be slowly nearing Artificial General Intelligence levels. This whole topic slowly turned from scary and feasible to really nice but currently not quite there. Obviously, you can always do something like Dishonored and quasi-script all the spectral facets of what flavor of sadistic murderer the player is, but the amount of content needed for that will quickly get unmanageable even on an AAA budget, and yeah, procedural generation and AI is helping and we're definitely gonna see more and more of it, but I think it will start from there; letting the designer order literally man-decades of content to their massive convoluted meta-script, and only slowly the AI will move upwards from making new textures and models to rooms to buildings to cities and only after that meaningful off-script quests and characters and whatnot. It's a question if you count Dwarf Fortress which is a weird case where you keep wondering if there's actual depth and thought to the world or if you're starting to hallucinate due to the brain damage you've suffered by the interface. There's poo poo like legends, entire story arcs, all of the world is generated procedurally, but then somebody uses gold worth more than their literal life to carve a picture of themselves eating cheese and you realize it's more of cleverly dressed noise than something coherent.
|
|
#
?
Dec 26, 2017 00:38
|
|
|
Teal posted:, but then somebody uses gold worth more than their literal life to carve a picture of themselves eating cheese and you realize it's more of cleverly dressed noise than something coherent. Hello we are on the internet that sounds relatively normal and totally realistic.
|
|
#
?
Dec 26, 2017 01:01
|
|
|
ElCondemn posted:It’s a shame Spore ended up not really having any real gameplay generation, it ended up just being randomized character generation. I wonder how the same concept would work out if they were developing it today or in 5 years. Probably the same because Spore was dramatically simplified. They developed a much more complicated game but then cut it down dramatically because they decided it was too complicated for popular audiences.
|
|
#
?
Dec 26, 2017 03:30
|
|
|
Hieronymous Alloy posted:Probably the same because Spore was dramatically simplified. They developed a much more complicated game but then cut it down dramatically because they decided it was too complicated for popular audiences. I think people like the idea they sat on a near complete miracle game but it's way more likely the tech demo never worked or was fun like you would expect for a game based on procedural interactions more advanced than ever seen now that came out in 2008
|
|
#
?
Dec 26, 2017 03:49
|
|
|
Owlofcreamcheese posted:I think people like the idea they sat on a near complete miracle game but it's way more likely the tech demo never worked or was fun like you would expect for a game based on procedural interactions more advanced than ever seen now that came out in 2008 Dude, give me some credit: I don't just make poo poo up. There are lots of sources on Spore being dumbed down deliberately. https://forum.quartertothree.com/t/how-spore-ended-up-cute-and-dumbed-down/48324
|
|
#
?
Dec 26, 2017 04:13
|
|
|
And the sad thing is that our former god, Will Wright, was the one behind making the creatures look like reject Maguzi designs. That seems to be a common theme we're running into nowadays— creators deliberately making works more cartoony/kid-friendly/lackadaisical than needbe and then hilariously misunderstanding what makes something "dark" or "mature" when the inevitable fan backlash kicks in. We all remember the tragedy of George Lucas, and I know that Craig McCracken (the Powerpuff Girls/Foster's Home guy) actually grew to not like the Powerpuff Girls movie because it was "too dark". I'm convinced Akira Toriyama developed this mindset too. There's my needless introduction of children's media into this discussion. In news that also gives me back some of my child-like wonder: How generative AI is changing art and design quote:You've heard a lot about AI in art and design in 2017, with Adobe trumpeting its Sensei machine-learning technology for everything from image recognition in Lightroom to tech previews of turning photos into 'hand-drawn' sketches and animated graph creation at this year's Adobe Max conference. Yuli Ban fucked around with this message at 04:42 on Dec 26, 2017 |
|
#
?
Dec 26, 2017 04:37
|
|
|
Hieronymous Alloy posted:Dude, give me some credit: I don't just make poo poo up. There are lots of sources on Spore being dumbed down deliberately. I'm saying the smart man's version is very very unlikely to have been good. They showed the early video and it looked cool but it screamed "this doesn't work as a game"
|
|
#
?
Dec 26, 2017 04:53
|
|
|
Owlofcreamcheese posted:I think people like the idea they sat on a near complete miracle game but it's way more likely the tech demo never worked or was fun like you would expect for a game based on procedural interactions more advanced than ever seen now that came out in 2008 Spore's gameplay is really, ridiculously simple. I'm not even saying this as a value judgment on the game or anything, it's just that mechanically every part of the game is incredibly basic. There were and are games that are drastically more complex in every way and still extremely popular, so it's not exactly a wild claim to say that Spore's simplicity was a design choice and not a technological limitation.
|
|
#
?
Dec 26, 2017 04:55
|
|
|
Paradoxish posted:Spore's gameplay is really, ridiculously simple. I'm not even saying this as a value judgment on the game or anything, it's just that mechanically every part of the game is incredibly basic. There were and are games that are drastically more complex in every way and still extremely popular, so it's not exactly a wild claim to say that Spore's simplicity was a design choice and not a technological limitation. Yeah, I played Spore when it came out, and there was literally no part of it that wasn't done better and in more depth by a then-current game that I had already played. You could basically do a 1 for 1 correllation between each phase and another, better, more complex game (i.e., space phase was Masters of Orion but for dummies). It was very obviously dumbed down dramatically. Functionally speaking it was the Cliff's Notes version of five much better games. I mean argue about what it was dumbed down from sure Hieronymous Alloy fucked around with this message at 05:05 on Dec 26, 2017 |
|
#
?
Dec 26, 2017 04:58
|
|
|
Paradoxish posted:Spore's gameplay is really, ridiculously simple. I'm not even saying this as a value judgment on the game or anything, it's just that mechanically every part of the game is incredibly basic. There were and are games that are drastically more complex in every way and still extremely popular, so it's not exactly a wild claim to say that Spore's simplicity was a design choice and not a technological limitation. The original video was beyond anything modern games come close to. It's obvious the original failed and they decided to replace a second try with a minigame collection.
|
|
#
?
Dec 26, 2017 05:05
|
|
|
Owlofcreamcheese posted:The original video was beyond anything modern games come close to. It's obvious the original failed and they decided to replace a second try with a minigame collection. It's obvious they eventually put out a minigame collection. I don't think we can necessarily conclude from that that the original design "failed" except in the limited sense of "someone with veto power didn't like it." And people with veto power are sometimes really dumb.
|
|
#
?
Dec 26, 2017 05:09
|
|
|
Hieronymous Alloy posted:It's obvious they eventually put out a minigame collection. I don't think we can necessarily conclude from that that the original design "failed" except in the limited sense of "someone with veto power didn't like it." And people with veto power are sometimes really dumb. It seems very unlikely someone made secret procedural motion algorithms that worked far better than anything made to this day then just sat on it forever. None of that stuff was ever real
|
|
#
?
Dec 26, 2017 05:26
|
|
|
Is 'too complicated' even an issue with... ANY game? People will got nuts over even Dwarf Fortress, where the main barrier to entry is the interface rather than the complexity. I suppose if they're worried about the audience being primarily children or the Sims crowd, maybe? But it's not like kids haven't spent years getting the hang of difficult and complex games. This frothing contempt so many game developers seem to have for their audience ALWAYS backfires. Either that or creators get middle aged and start worrying about their (possibly hypothetical) kids not being able to play/enjoy their stuff and so dumb it down to what they think the kids will like.
|
|
#
?
Dec 26, 2017 05:26
|
|
|
Inescapable Duck posted:Is 'too complicated' even an issue with... ANY game? People will got nuts over even Dwarf Fortress, where the main barrier to entry is the interface rather than the complexity. I suppose if they're worried about the audience being primarily children or the Sims crowd, maybe? But it's not like kids haven't spent years getting the hang of difficult and complex games. Spore came out pre-minecraft so my guess is some extremely stupid executive decided "This is a science game right? That means it's for kids. The Sims sells well. We need to make this more like the Sims so children will like it."
|
|
#
?
Dec 26, 2017 05:33
|
|
|
Anyway, to get back to the subject generally and not just Spore :ElCondemn posted:It’s a shame Spore ended up not really having any real gameplay generation, it ended up just being randomized character generation. I wonder how the same concept would work out if they were developing it today or in 5 years. Teal posted:
BrandorKP posted:Hello we are on the internet that sounds relatively normal and totally realistic. I'd definitely count Dwarf Fortress as a story-generating machine. Sure, in a sense you're just percieving patterns in the randomness, but. .. that's how narratives form, that's the root mechanic in the human brain anyway. I mean, read Boatmurdered: https://lparchive.org/Dwarf-Fortress-Boatmurdered/ That's a story, however it was generated. The trick is having enough different things being tracked and then connected to each other that you can get emergent narrative behaviors popping out of all that underlying structure and chaos. This is why other games that have tried to replicate "Dwarf Fortress but with a better interface" have generally failed -- see, e.g., Stonehearth, which is probably the closest anyone's come, but after four years in development they've spent the whole past year going back in and adding additional layers of background stuff (personality systems for individual town dwellers, likes and dislikes, fort history tracking, etc.) because it takes that level of granular detail and inter-connectivity to get the emergent behavior. I think a game along the lines of Spore could work in the same sense, it would just take a lot more layers of apparently unnecessary complexity before the cool emergent behaviors started to arise. Most rational game developers are going to look at all that apparently wasted effort and pointless detail and toss it out of the budget cycle, but it's a necessary part of the process, in the same way that they trained Watson to play Jeopardy by feeding it the internet. But Tarn Adams is a crazy genius man who spent . . like two decades now? .. . modelling everything.
|
|
#
?
Dec 26, 2017 05:48
|
|
|
This is it, creating systems capable of emergent behaviours is kind of at odds at creating curated experiences rich with content. This is why AAA games arguably have the simplest mechanics, it's not just appeal, but that every single thing in the world needs to be represented by something. If you're billing your game as a top tier $60 experience, you're saying that every piece of content needs to look great, and that costs. The level of modelling present within Dwarf Fortress is a byproduct of how content cheap all the representation is, they can focus on systems rather than art assets to bring the world to life with all this emergent possibility. AI might be able to do brushwork and a whole range of other things, but matching an art direction to form a cohesive part of a procedural experience, it ain't there yet. Any game that tries to build with complex systems tends to run up against the needs to produce assets to match. Typically they have to then limit themselves to some maximum level of fidelity to ever finish. The lack of complex games especially in the AAA space doesn't so much smack of contempt for player intelligence as it does run up against the realities of asset production at that level of execution.
|
|
#
?
Dec 26, 2017 06:10
|
|
|
Emergent behaviour in this age of Lets Plays and social media is marketing gold, though. Look at the popularity of Shadows of Mordor, Skyrim, any game where there's enough freedom and moving parts for weird poo poo to happen. Half the reason why every game is turning into a sandbox, it's one of the easier ways to let players feel like they're exploring something and not just following the rails on a carnival ride.
|
|
#
?
Dec 26, 2017 06:16
|
|
|
Maluco Marinero posted:This is it, creating systems capable of emergent behaviours is kind of at odds at creating curated experiences rich with content. This is why AAA games arguably have the simplest mechanics, it's not just appeal, but that every single thing in the world needs to be represented by something. If you're billing your game as a top tier $60 experience, you're saying that every piece of content needs to look great, and that costs. The level of modelling present within Dwarf Fortress is a byproduct of how content cheap all the representation is, they can focus on systems rather than art assets to bring the world to life with all this emergent possibility. Yeah, that's the exact problem Stonehearth's having: even with relatively simple voxel graphics, modelling a system complex enough for emergent behaviors turns out to be difficult enough that they haven't managed to catch up with Toady's one-man effort in four years, despite being a team of trained and talented professionals, funded by Riot Games, trying to match the efforts of one half-mad savant funded entirely by donations. Still, just because we're not there yet doesn't mean we won't be ever. Someday someone's going to figure out how to feed the Library of Congress to an AI and have it spit out original quality narrative content in response, and once that happens things are gonna be weird. Hieronymous Alloy fucked around with this message at 06:28 on Dec 26, 2017 |
|
#
?
Dec 26, 2017 06:22
|
|
|
Some fella wrote a whole bunch of words (and also I think a thesis?) about how you could make much more interesting procedurally generated worlds, and is making a game to show what he means. Cultures, complete with architectural styles, clothing styles, modes of address...and that's just one guy working with ASCII graphics. Imagine the man-hours of work that currently go into placing every house and tree in an open world game instead going into a bunch of shufflable styles and variables. Or getting dropped entirely. I have no idea which scenario will actually happen, if either.
|
|
#
?
Dec 26, 2017 06:32
|
|
|
Thinking on this line of procedural generation, and more broadly creative works in general, I think on how much of quality gameplay and narrative is born out through testing. The only way to determine the qualities of a piece is to be capable of experiencing it as a consumer, and then checking the design intent against what was actually experienced. I guess where this is of interest is how a million typewriters surely can generate many permutations, without intelligent evaluation how can they know they landed on a successful one. Intelligent evaluation is the lynchpin that makes it possible to procedurally generate with more variables, but as you say it’s weird. It delves into the environment of general artificial intelligence because evaluation is fundamental to decision making in general, whether it’s an automated level designer, artist, or an automated moderator. (and we all know how that’s working for the big guys right now - it ain’t)
|
|
#
?
Dec 26, 2017 08:42
|
|
|
I see no reason why you couldn’t just have another AI play through the generated content and provide a score as a feedback. That’s kind of in the spirit of the adversarial NN already.
|
|
#
?
Dec 26, 2017 09:12
|
|
|
mobby_6kl posted:I see no reason why you couldn’t just have another AI play through the generated content and provide a score as a feedback. That’s kind of in the spirit of the adversarial NN already. Yeah I think people are jumping to the general AI goal way too early, we have a long way to go and a lot of these problems can be tackled using current technology. It's just a matter of figuring out how to apply it in a way that works for the problem at hand. If spore were made today they could just run a NN cluster somewhere spitting out new species assets with all kinds of behaviors and designs that would be way more interesting than what we got. One of the problems I had with that game was that it pretended like there was actual evolution happening but every creature design "worked" and never changed, this is pretty easily solved today, even within the gameplay constraints they created for themselves.
|
|
#
?
Dec 26, 2017 16:06
|
|
|
Sounds like Spore did a Kickstarter hype pitch before there even was Kickstarter. With exactly the same results too (either turning into vaporware or releasing some lovely, barely playable crap to avoid fraud accusations) Never trust a gamedesigner/-developer, always demand playable results in exchange for money
|
|
#
?
Dec 26, 2017 16:23
|
|
|
Doc Hawkins posted:Some fella wrote a whole bunch of words (and also I think a thesis?) about how you could make much more interesting procedurally generated worlds, and is making a game to show what he means. Cultures, complete with architectural styles, clothing styles, modes of address...and that's just one guy working with ASCII graphics. Imagine the man-hours of work that currently go into placing every house and tree in an open world game instead going into a bunch of shufflable styles and variables. A contrast to this would be a Paradox series he doesn't talk about, namely Crusader Kings. There you have 10-20 thousand AI characters, with traits ranging from the common to the very rare, interacting with each other, creating emergent story lines in their interaction with the player and each other. I don't see any reason why such a system couldn't also be used in for example an RPG, moving the game away from custom made quests to ones that are procedurally generated according to the goals of the AI agents. Like, if an AI agent is basically a CK2 duke who wants to assassinate his rival, perhaps the player could then be tasked with actually carrying out the plot? Depending on how well the player performs, or the AI's plans, perhaps a war would be the next step, and suddenly the player is involved in a massively dramatic development which no developer has pre-defined. The question here basically becomes how far down the hierarchy you go in terms of having NPC's exist on the AI agent level - maybe some of them are high level versions (kings/politicians/bishops/rebel leaders and what have you) and others more low level with simpler scopes (small time gangsters and raiders) Actually, the CK2 approach almost seems mandatory if you auto-generate the world. At least, I have a hard time imagining pre-written story for a world that has not been generated. The cultural/political/religious layer the dude is talking about would also work really well with the emergent gameplay idea, as long as you add personality traits to the AI agents that can bounce off of those layers - like having a zealous xenophobe AI that weighs loving up culturally/religiously distinct AI's really highly, the greater the difference the more eager it is to assassinate, imprison, or start wars. Would also be hilarious if it resulted in as varied outcomes as you see in CK2. Like, one time you help a duke and rescue his wife, and you gain a friend for life and a safe haven from the people who aren't big fans of you. With different (hidden) personality traits, he eats his wife, sacrifices his first born to auto-generated Satan, starts a hell-war with all his neighbors and proceeds to proclaims you his most trusted and favored advisor.
|
|
#
?
Dec 26, 2017 18:07
|
|
|
Raspberry Jam It In Me posted:Sounds like Spore did a Kickstarter hype pitch before there even was Kickstarter. With exactly the same results too (either turning into vaporware or releasing some lovely, barely playable crap to avoid fraud accusations) I think the cell section of the game is fundamentally the game they were trying to make. Where you could put 8 mouths on a thing and the engine would figure out on the fly how the physics would work for it biting something using the 3rd mouth on the left. Doing that in full 3D it seems really really obvious they never had any technology beyond making stuff walk procedurally and so the third mouth just ended up being "+1 to bite" or whatever in an RPG combat type system that wasn't really what anyone wanted because they clearly weren't going to make some rich RPG system and even good RPGs people tend to like for the story which this didn't have.
|
|
#
?
Dec 26, 2017 21:06
|
|
|
muon posted:GANs are indeed extremely cool, but I think you're (understandably) over-estimating how much they can do. The horse to zebra transform is a good example: these AIs do not "understand" a horse as a "a work animal with a mane and etc etc", they "understand" a horse as, essentially, an average over the thousands of examples presented to the GAN with a label of 00000100000 instead of 00001000000 (this one's a zebra). Now clearly that is still an incredibly powerful thing to "understand" and allows us to take the average of "zebra" and combine it with the average of "horse", but it does not imply that the GAN "knows" how to make a horse/zebra walk like humans can easily imagine. Now we could design a neural net to learn how to do that, but the only way we know how to do that today requires thousands and thousands of labelled data sets of horses walking before we could even begin to create that model. Neural networks need to have nuanced semantic understanding and ability to generalize to really be able to help create content in the way you're describing. That's an extremely difficult task that I don't know we'll solve by the 20s. It being pixel-based is a big problem. Watching some of the GAN examples in the YT video it has an extreme difficulty transitioning between orientations of the requested image. Example: It does pretty well on faces, because they're all mostly facing forwards. It has trouble with glasses/lack of glasses because how do you seamlessly transition a binary state? Sheep, though, that's a nightmare straight from Cyriak's inhuman video editing skills. Smooth animation may come from computer animation's source and output being fed in - give the network the skeleton and the resulting render, to teach it to do inverse kiniematics. See if you can feed it a horse skeleton and a picture of a real horse and have it figure out where the bones lay. Machine-learning IK would be pretty insane, because that's beyond pixels and into understanding the physical reality of what's happening in a picture. It'd be an interesting constraint on the outputs as well, since whatever it imagines would have to be skeletally possible.
|
|
#
?
Dec 26, 2017 21:23
|
|
|
And now witness this. Remember that time about a year and a half ago when a man fed Blade Runner into a neural network? Turns out that the neural net was able to remember Blade Runner and managed to transpose its aesthetic onto other films. The Neural Net That Recreated ‘Blade Runner’ Has the Movie Stuck in Its Memory The AI that made ‘Blade Runner: Auto-encoded’ transposed the aesthetic of the movie onto other sci-fi classics quote:But what happens when the Blade Runner auto-encoder watches other films? Broad tried it out. When shown another Philip K. Dick adaptation, A Scanner Darkly, and a Soviet Classic, Man with a Movie Camera, it could still recognize the composition of the frames, but it essentially transposed the aesthetic of Blade Runner: Auto-encoded. They were dimly-lit, plagued by visual noise, and dreamy. The auto-encoded versions clearly came from the same memory.
|
|
#
?
Dec 26, 2017 21:40
|
|
|
Yuli Ban posted:And now witness this. Okay I don't like this article, and while I love what the guy is toying with, I hate the way he talks about it. Autoencoders are more like a glorified data compression algorithm that does it's earnest to figure out new patterns that will help it to fit in just a smidge more info on the stuff it's meant to learn into the horribly inadequate amount of memory it has. The side effect of that is that indeed, in ideal case, it does learn the "essence" of the entry data and finds the patterns that you'd struggle to uncover through classic engineery methods for you, and you can sift through the data in this new bastardized form and get a kind of "machine's notes" on it, and then for instance run clustering on that representation. It's not "mind of an AI", it's more like making a classroom full of of kinda special elemenetary graders write cliffnotes on a book and telling them "just describe whichever bits of each chapter found the most interesting" to all of them, and then averaging the results. It doesn't give insight on how AI works, or insight how eyesight works; it gives an insight on what's the lowest entropy description of a thing with a purposefully insufficient language for describing it.
|
|
#
?
Dec 26, 2017 22:57
|
|
|
Let me use this space to whine about human memory. Is like a lovely compression algorithm.
|
|
#
?
Dec 26, 2017 23:31
|
|
|
Yuli Ban posted:And now witness this. Isn't what he's describing just a Rorschach test for people looking at the product of neural nets? Scanner is a "dreamy" movie also and I've never seen a neural net creation that wasn't an extremely noisy fractal approximation of something. In fact the Scanner compression looks much more painterly fitting the rotoscoped style.
|
|
#
?
Dec 26, 2017 23:32
|
|
|
Tei posted:Let me use this space to whine about human memory. Is like a lovely compression algorithm. The critical difference between human memory and an autoencoder is that we build the representation on basis of abstract yet meaningful parallels and there's no proof this will ever work by feeding a "large enough" neural net based autoencoder of this type "enough data". The foundation of your memory of a piece of cake you ate yesterday is understanding of cocoa which fills in the flavour, the colour, the scent, the likely texture cocoa based cakes have. You don't need to remember the angles the slice had, because you've seen a whole cake before and have seen a circle sliced into wedges countless times before. You don't need to think about if it came served on porcelain or a patch of snow, because you know cakes come on platters, and those are usually made of porcelain. Feeding all those connections in a meaningful manner without "stubs" into a neural net might easily turn out to simply not work; even if you have the resources to make it huge and give it all the data and time it might possibly mean, NNs come without a guarantee of ever converging to the global minimum (e: of training error, which is in this case extremely hard to define); it might simply enough not figure out the right connections in stuff (and it's often the case, and the best you can do is shrug and keep trying new and new combinations of parameters and data representations and eventually something else to do). Teal fucked around with this message at 10:54 on Dec 27, 2017 |
|
#
?
Dec 26, 2017 23:42
|
|
|
|
| # ? May 8, 2024 08:23 |
|
|
Harik posted:It being pixel-based is a big problem. Watching some of the GAN examples in the YT video it has an extreme difficulty transitioning between orientations of the requested image. Example: It does pretty well on faces, because they're all mostly facing forwards. It has trouble with glasses/lack of glasses because how do you seamlessly transition a binary state? Sheep, though, that's a nightmare straight from Cyriak's inhuman video editing skills. That's not really an artifact of it working on the scales of pixels, but just an artifact of the fact that a neural network is basically a giant polynomial function. It's entirely continuous. This is true of neural networks regardless of the type of data being worked on. If there's any discontinuity in the dataset (like in your example, a person is either wearing glasses or not) then the best the generator network can do is try to limit the amount of input space that maps to the transition between those areas. For orientation, though, I imagine that that's a limitation of the dataset. People tend to take pictures from particular angles. Like, a person is either straight on or at 3/4 profile, and the dataset doesn't really represent enough intermediate space for the system to learn how to properly vary between those areas, so they may as well be discrete areas as far as the network is concerned.
|
|
#
?
Dec 27, 2017 16:47
|
|
