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

GAN Theft Auto [video]

youtube.com

Post reply on HN
Print view View original

51–60 of 99 posts

Re: GAN Theft Auto [video]

#51

Post by roystonvassey »

andrepd wrote: post #22
ShamelessC wrote: post #10

Great work! Hacker News still seems to have a deeply skeptical culture with regard to machine learning - not sure why. There's always someone saying it's "not novel" and it's "just doing x". Overfitting is a known issues in machine learning, people. If you still think all neural networks are doing is memorizing the dataset completely in the year 2021 - you might want to revisit the topic. It is one of the first conce…

> Hacker News still seems to have a deeply skeptical culture with regard to machine learning Is... that a bad thing? Skepticism is good. When it's about something as hyped as "deep learning", even more so.

I think it’s a problem when it turns to - being skeptical for the sake of it.

Not been too long on HN but the top comments on most threads are a contrarian one (and one which I truly appreciate because it provides a different POv) but sadly because it is encouraged through the high upvotes, the crowd tendency is to regress towards this approach, even if sometimes the rigour of the critique is lacking

Re: GAN Theft Auto [video]

#53

Post by danuker »

godelski wrote: post #31
sentdex wrote: post #17

Earlier quoted context omitted.

Heh, yeah, tough crowd I guess. The full code, models, and videos are all released and people are still skeptical. I feel like 95%+ of papers don't do anything besides tell you what happened and you're just supposed to believe them. Drives me nuts. Not sure why all the hate when you could just see for yourself. I'd welcome someone who can actually prove the model just "memorized" every combo possible and didn't do an…

> I feel like 95%+ of papers don't do anything besides tell you what happened and you're just supposed to believe them. Honestly I think there's a big problem with page limits. My team recently had a pre-print that was well over 10 pages and we still didn't get everything and then when we submitted to NeurlIPS we had to reduce it to 9! This seems to be a common problem and why you should often check different version…

There's the Hutter Prize [1] - memorizing is useful (and arguably intelligent) if it's compressed.

http://prize.hutter1.net/

Re: GAN Theft Auto [video]

#54

Post by bobsmooth »

sentdex wrote: post #16
rasz wrote: post #13

Earlier quoted context omitted.

The first link exposes the trick employed by your model. >animation info and simulation data but did your model learn any of that? >explicit goal of the video you posted is to combat runtime constraints The trick to motion mapping is feeding a lot of data with accompanying inputs to build an atlas you can reference during playback. >first large-scale GAN successfully trained on GTA V Its really cool. The problem I ha…

> I immediately felt insincerity bordering on scamming the audience MFW I read this. Jeez man. Model size is 173MB. It didn't just memorize every possible combo. How the hell you went from our excitement about a fun project we shared on YT to accusing us of "scamming" the audience I really don't know. What a terribly rude and hateful attitude you have =/

The people on this website are terrible sometimes.

Re: GAN Theft Auto [video]

#55

Post by slver »

whalesalad wrote: post #27

GPUs: am I a joke to you? Instead of using them to render polygons, let’s use them to train neural networks that produce models that make them unnecessary. I’m oversimplifying - but pretty wild nonetheless.

Well, neural networks run (fastest) on GPUs.

Re: GAN Theft Auto [video]

#56

Post by slver »

dividuum wrote: post #2

Impressive. Makes you wonder if at some point in the future there isn't a game engine any more but tons of training material and you play in a generated dream.

Maybe not entirely, because just like a dream, the rules of a neural network tend to drift and be somewhat fuzzy.

Unless it's a high-concept game whose very goal is offering you a dream environment.

But I do believe neural networks will get into everything. They're the last missing piece of our compute model.

Re: GAN Theft Auto [video]

#57

Post by slver »

sentdex wrote: post #16
rasz wrote: post #13

Earlier quoted context omitted.

The first link exposes the trick employed by your model. >animation info and simulation data but did your model learn any of that? >explicit goal of the video you posted is to combat runtime constraints The trick to motion mapping is feeding a lot of data with accompanying inputs to build an atlas you can reference during playback. >first large-scale GAN successfully trained on GTA V Its really cool. The problem I ha…

> I immediately felt insincerity bordering on scamming the audience MFW I read this. Jeez man. Model size is 173MB. It didn't just memorize every possible combo. How the hell you went from our excitement about a fun project we shared on YT to accusing us of "scamming" the audience I really don't know. What a terribly rude and hateful attitude you have =/

I wouldn't call it scamming, but 173MB is not small at all. At the resolution of this model, you can easily fit the entire Titanic movie in 173MB. Maybe even have enough space for audio.

Furthermore no one is saying the model "memorized every possible combo". However imagine you have a set of keyframes (maybe even multiple fragments per frame) and you need to interpolate between them? Not that hard of a task, isn't it.

Models don't care about simulating our "intention" properly. They care about fitting the input in the simplest way possible. Think about a model like a lazy worker merely trying to look like it's working.

None of this makes NN less exciting, but it should inform us you can't go 0 to 60 in one step and hope the NN would have great insight about what it's doing.

We need models that make smaller conceptual jumps, i.e. models that understand 3D space, then models which understand transformations in 3D space, then models which understand citicscape, etc. etc.

Re: GAN Theft Auto [video]

#58

Post by eru »

sentdex wrote: post #15
rasz wrote: post #9

Earlier quoted context omitted.

What he meant is you overfitted the network with video footage. There is no game, just seemingly clever stitching and playback of learned footage similar concept applied to animations and implemented in a state machine https://www.youtube.com/watch?v=KSTn3ePDt50 and optimized with nn https://www.youtube.com/watch?v=16CHDQK4W5k

We had ~100GB of data (and that was gzip compressed data). The final model is 173MB. It's simply not large enough to have memorized every combo.

Don't think gzipping helped much with video data?

Re: GAN Theft Auto [video]

#59

Post by okamiueru »

Can someone explain a bit more on the long term applicability, or maybe other use cases that might be easier to appreciate?

The reason why I ask is that it seems very challenging to generate the training data for such systems. Could someone explain how this can go further than to just replicating X? So, if assuming some creative freedom, could you give an idea of what the long term application of this would be?

NB: please take my questions at face value without thinking I'm implying this isn't cool for what it is. I'm all for people having fun. I'm all for projects not needing to tackle some grander issue.

Re: GAN Theft Auto [video]

#60

Post by tiborsaas »

okamiueru wrote: post #59

Can someone explain a bit more on the long term applicability, or maybe other use cases that might be easier to appreciate? The reason why I ask is that it seems very challenging to generate the training data for such systems. Could someone explain how this can go further than to just replicating X? So, if assuming some creative freedom, could you give an idea of what the long term application of this would be? NB: p…

In the future we might have a fourth common media format besides pictures, videos and audio: GAN records.
Post reply on HN