Support for Machine Learning (ML) on PS5 and Series X?

[invictis]

They might be using ML for HDR Enchantments on X360 and OG BC titles.

I think you are right about that.
Well there's one use.

 

[Janne Kylliö]

I think the XSX is going to be able to out put all games at 4k, do I don't see a need for 1080p to 4k upscaling

I think deep learning super sampling is potentially a true performance multiplier, a game changer. Maybe XSX has it, maybe not. But in a year or two it'll be a must have feature in PC graphics cards.

For non ray traced games:

XSX: 4K -> 8K
XSS: 1080p -> 4K

For DXR games:

XSX: 1080p -> 4K
XSS: 540p -> 1080p

If Lockhart (XSS) is real thing and has DLSS 2.0 type super sampling technology, it could even rival or out-perform PS5 lacking this technology (three big if's here!). And have better IQ as well (no TAA).

Here you can see that even RTX 2060 will provide somewhat playable framerate in Control using all DXR effects and outputting 4K resolution.

 

[TS_JBG]

o with 97 TOPS on the XSX, and the use of DirectML, what can we expect for the XSX in all reality?

Not 97TOPS of machine learning or that would leave not a single flop to output something on the screen first

Secondarily, AI enhancement or image reconstruction technology seems the two only, and meaningful, type of operations it coould help with. And it could certainly help well.

 

[PSman1700]

I think deep learning super sampling is potentially a true performance multiplier, a game changer.

I think it's the biggest game changer so far, things like ssd, ray tracing, 3d audio enable improved graphics but not directly performance as this tech does. Offcourse a beast of a gpu is welcome, a GPU equalling 2080Ti/12+TF's of power will supplant that, but tech like VRS and DLSS (or whatever MS call it) will improve performance insanely. SSD's with their speeds and seek times will allow quick or no loading and maybe allow more assets.

 

[AlNom]

I’d like to see ML applied to screenshots and video clips. Perhaps even built into photomodes.

 

[iroboto]

I’d like to see ML applied to screenshots and video clips. Perhaps even built into photomodes.

ML applied correctly could have some interesting results. I have faith that there is going to be ML in games next gen, and daring enough individuals to include it for this generation.

 

[PSman1700]

ML applied correctly could have some interesting results. I have faith that there is going to be ML in games next gen, and daring enough individuals to include it for this generation.

It's one of the most intresting new features imo, that and ray tracing. All other aspects have more to do with speed/doing things much faster. Yes that opens up things for some intresting scenarios, but it's still the old doing things faster, like every generation has done basically.

 

[RobertR1]

I’d like to see ML applied to screenshots and video clips. Perhaps even built into photomodes.

No AL, NO! We don't need more photomode screenshot wars only now more BS'd by AI. It'll be enough of a shit show as it is, I promise. You're a sick man!

 

[Deleted member 11852]

ML applied correctly could have some interesting results. I have faith that there is going to be ML in games next gen, and daring enough individuals to include it for this generation.

I also like to remind folks of the downsides to machine learning and that it can be a be a liability causing unpredictable results. We all remember Microsoft's racist Hitlerbot, Tay and Apple's iOS predictive text issue. ML FTW! The instant you take an algorithm that adapts to new data and modify to an algorithm that can modify itself, unless who wholly control the input, you've just lost control.

Tread with caution, there by dragons here. And big dongs. Because you know if it's possible, it will be.

 

[iroboto]

I also like to remind folks of the downsides to machine learning and that it can be a be a liability causing unpredictable results. We all remember Microsoft's racist Hitlerbot, Tay and Apple's iOS predictive text issue. ML FTW! The instant you take an algorithm that adapts to new data and modify to an algorithm that can modify itself, unless who wholly control the input, you've just lost control.

Tread with caution, there by dragons here. And big dongs. Because you know if it's possible, it will be.

lol. Indeed. But those aren't the type to happen within games. Unless you are using a reinforcement algorithm, you can't train back a machine learning algorithm. Our devices will run the trained model, but won't be able to change it. You'd have to be able to change the model for those types of issues to happen.

Which is funny, lol. I mean they should have expected that. But yea, in games the results aren't perfect. In real time ML, you want faster and smaller (footprint), but that's going to come with less and less accuracy. So it really depends on what developers are willing to walk away with and if 80-87% accuracy is sufficient for their needs. I don't know.

 

[manux]

One interesting place to experiment with AI models would be pedestrian/human AI in something like grand theft auto. In essence try to make the extras more interesting in all kinds of ways(behavior, speech, voice...).

 

[Deleted member 11852]

lol. Indeed. But those aren't the type to happen within games. Unless you are using a reinforcement algorithm, you can't train back a machine learning algorithm. Our devices will run the trained model, but won't be able to change it. You'd have to be able to change the model for those types of issues to happen.

Machine learning is self-improving/adapting algorithms. Algorithms that don't change but have been developed through machine learning are not themselves machine learning. I assume we are talking about the former, rather than the later.

 

[Deleted member 11852]

One interesting place to experiment with AI models would be pedestrian/human AI in something like grand theft auto. In essence try to make the extras more interesting in all kinds of ways(behavior, speech, voice...).

Maybe, it depends what the AI is adapting too. You don't want the AI adapting to the actions of other stupid AI, but perhaps to player's behaviour? Even that could go wobbly quickly, i.e. you go on a killing spree for 30 minute and now all the AI pedestrians run away from you on sight you all the time. It'd be like a hideous bug you could never fix unless you start over.

 

[iroboto]

Machine learning is self-improving/adapting algorithms. Algorithms that don't change but have been developed through machine learning are not themselves machine learning. I assume we are talking about the former, rather than the later.

Yea we are referring to the latter; machine learned models I guess. I don’t expect reinforced learning, which is constant adaption and a change in trained behaviour (the former) to show up anywhere for games.

 

[Deleted member 11852]

Yea we are referring to the latter; machine learned models I guess. I don’t expect reinforced learning, which is constant adaption and a change in trained behaviour (the former) to show up anywhere for games.

I'm expecting the former because that is what the hardware in Series X is optimised for - I assume something similar in PS5. You don't need ML-optimised hardware to run code developed from applying ML techniques, that's ordinary CPU/compute code.

 

[iroboto]

Though that type of AI might work for games specifically designed to be broken or funny. I think there is a reinforced learning NLP dungeon master making rounds

 

[manux]

Maybe, it depends what the AI is adapting too. You don't want the AI adapting to the actions of other stupid AI, but perhaps to player's behaviour? Even that could go wobbly quickly, i.e. you go on a killing spree for 30 minute and now all the AI pedestrians run away from you on sight you all the time. It'd be like a hideous bug you could never fix unless you start over.

I would expect it to be a model that was learnt offline. The model wouldn't keep learning once it's released. In essence replace existing hardcoded rules with a model that was learnt offline.

Having something that keeps learning is whole another ball game.

 

[iroboto]

I'm expecting the former because that is what the hardware in Series X is optimised for - I assume something similar in PS5. You don't need ML-optimised hardware to run code developed from applying ML techniques, that's ordinary CPU/compute code.

It’ll run faster. We can design Deep Learning models from int 4 to float 64. Int 4 being the fastest and least precise. Generally speaking even if there were RL based AI in next gen consoles; it’s only got barely enough juice to retrain itself offline. It’s unlikely to retrain itself during gameplay especially not in real-time.

 

[Deleted member 11852]

Though that type of AI might work for games specifically designed to be broken or funny. I think there is a reinforced learning NLP dungeon master making rounds

This is why I always throw in the cautionary tale of the unpredictability of machine learning. Once you open that box belonging to Pandora, its difficult to close. There are things that ML could be really suited for, like multiplayer bots when there aren't enough people but this would be massively risky having bots learn from player behaviour because it can be manipulated/influenced.

 

[Jay]

I would expect it to be a model that was learnt offline. The model wouldn't keep learning once it's released to in a game. in essence replace existing hardcoded rules with a model that was learnt offline.

Having something that keeps learning is whole another ball game.

Could keep learning in multilayer games and improve bot ai etc