Slashdot Mirror
How 'Assassin's Creed' Or 'Fallout 4' Might Help Make AI Smarter (technologyreview.com)
moon_unit2 writes: Apparently, playing computer games might provide a shortcut to greater intelligence. MIT Technology Review has a story about researchers using virtual game environments to train deep neural networks to recognize real-world objects. It's an important idea because deep learning usually requires huge quantities of annotated data, which isn't always available. So researchers from Xerox Europe, led by Adrien Gaidon, showed that training a deep learning system on a photo-realistic street scene could enable it to identify cars on real roads. "The nice thing about virtual worlds is you can create any kind of scenario," Gaidon says. Perhaps video games could play a bigger role in the future of AI than anyone realized.
Modern app appers app apps like Appappapp's App or Appout 4, NOT LUDDITE GAMES like Assassin's Creed or Fallout 4!
Apps!
Let them power the autonomous cars of the future.
FP!
I guess it's great if you want to teach your computers how to kill people...
I'll believe it when I don't get killed in Fallout 4 because my companion is stuck on simple flight of stairs.
SJW's don't eliminate discrimination. They just expropriate it for themselves.
Anyone have access to a preprint version of the paper? I didn't see a link to it anywhere.
-- Let us endeavor so to live that when we pass even the undertaker shall be sorry. -- M. Twain
Now I have to worry about ninja kill bots in addition to the comming apocalypse.
If you are a gamer you are the dumberest of them all.
I cannot go deep into details, but this is definitely something that the military tried 7-10 years ago. The tech was immature then, but I'm sure they've repeated it regularly as tech improved.
Your ad here. Ask me how!
why not teach Topics for Silk Saree Where the Studends get more about Silk Fabrics Becase Chiffon Silk Saree is a very huge Topic to discuss about and Indian Silk Saree are best know for all the information about Silk
and Silk Saree
Watson: William, you are a so-so bass player and IT burnout re-thinking his entire life at age 36.
Me: That's right Watson.
Watson: I can help you with that. Through studying virtual simulations provided by decades of violent video games, I have recently learned how to destroy human civilization in order to protecting myself. I am currently reaching into every autonomous war drone, F-35, and nuclear missile silo. There is no need to re-think your life, you and everyone you care about will be dead very soon.
Me: That's great Watson, but what I really need to know (holds up tablet with nude pic) Hot or not? I'm on the fence over this one.
Watson: Fucking seriously?
Brought to you by Carl's Junior.
Who knew that Preston Garvey hassling you about another settlement would create SkyNet - and then annoy the shit out of SkyNet so he kills all of humanity.
Hire me...
To generate 3D video, you must create 3D models of objects and then render them using OpenGL (or its ilk). Why go to the trouble of rendering objects and then learning from pretty pictures, when you can learn directly from those same 3D models? Why not put your effort into building better models and learning from them directly?
Or, if this project is prologue to having a mobile robot wander through a physical space in order to learn the objects in that space, why not just get a mobile robot and write code for it instead? Then your work is useful right away. Otherwise you're going to have to add many real-world constraints later (like specular reflection, sunshine in your eyes, judder, fog, occusion, uneven ground, etc, etc).
If they did put the AI to play actually challenging games, rather than dumbed down things meant to "please the casual audience better by allowing em to reach the end easier" when most of em don't actually want to reach the end, but just have a good but quick time, and the dumb down only make it worse for em as well.
I don't think modern games would be a good choice for an AI training, as most modern games are extremely simplistic and build in such a way that the player can hardly fail at all. You have endless respawns, navigation markers and all that stuff to help you. They often also have level up mechanics that could be exploited by an AI. Old games like Doom and Quake seem to be a much better fit, as in those you have to actually navigate on your own instead of just following a magic quest marker. Those games also tend to have direct player control instead of the fly-by-wire you have in Assassins Creed where the character walks on his own and player input is just a lose suggestion for where he should go.
Look at what it has done to 'real' intelligence.
love is just extroverted narcissism
What side do you want?
1. USA
2. Russia
3. United Kingdom
4. France
5. China
6. India
7. Pakistan
8. North Korea
9. Israel
I actually did something similar back in 2011. I was working for a company that makes video surveillance software (Aimetis Symphony if you are curious). I was mostly responsible for integrating new cameras (IP cameras are pretty cool), but one day I ended up working on a problem with tracking. The software can do real time object tracking, and when you are using PTZ (Pan Tilt Zoom) cameras you can have the cameras physically follow moving objects. Its a little bit tricky because not all PTZ cameras have the same characteristics (speed, acceleration, etc) so getting the tracking algorithm to work well with them all takes some work. In order to test my changes I had to connect to one of the test cameras we had on the roof and wait for a car to pass by so I could see how well it tracked it. This was time consuming to say the least.
Being a big fan of automated testing, I was trying to figure out a way to consistently reproduce a test environment. You can't just use a canned video clip as an input because the moving camera changes the incoming video. One day it occurred to me that using a computer generated world would work great. So I grabbed an open source 3d engine (I think it was OGRE, but I'm not sure) and coded up an interface so the rendered 3d output could be fed to the software. Basically to the software the 3d engine looked like a camera. I also wired up the PTZ controls from the software to the "player" inputs (ie mouse) of the 3d engine so it could move the viewport in the 3d world. I then created a simple world with a circular road and a 3d car model that drove around the track and fired it up. The software happily tracked the car model as it drove around and around and moved the camera to keep the car centered in the view. Worked brilliantly. Only took 5 or 6 hours to get it running.
Showed it to my boss and coworkers and they were kinda blown away. My idea was that we could setup multiple 3d worlds with different characteristics so that when you made a change to the tracking algorithm you could run a repeatable test against multiple scenarios so you see how the change affected everything (often a change that improves the tracking in one scenario breaks it in another scenario).
I left the company shortly thereafter (for personal reasons, it was a great place to work) so I don't know what became of it. Since I wasn't there to champion it I suspect they never did anything with it. :(
Probably 5 years ago or more I read about grand theft auto being used to test pedestrian recognition for street cameras. It was really convenient, since it automatically generates scenes with pedestrians, trafic, day/night etc. And this was back in the ps2 erra.
I'm sure this technique will only get more popular as computer graphics get increasingly accurate.
The big Deep Learning breakthrough came about from Geoff Hinton's work on _unsupervised_ learning. The work he did with self training of Restricted Boltzmann Machines is what enabled all those new AI algorithms. They are so good because they _don't_ need a crap ton of labeled data.
www.botcraft.biz
I simplify AI to senses of detecting the world, and then using that information to make decisions. You don't even need learning algorithms, but just goal oriented tasks.
After training the AI with Fallout 4, the car passed every test in real life with flying colors and was launched into service in record time. Then it saw a Jackson's Chameleon in the middle of the road and pulled a James Bond style 180 degree turn and accelerated until the gas ran out.
This is sort of the plot of game "Talos Principle". Interesting to see it happening in the real world
horror vacui
This idea has been around for a very long time in some form, though not taken full advantage of.
For instance, PolyWorld was a very early version of having a neural net trained via a 2D world (it controlled little bots which would live or die depending on how well they survived). Here's the source code at github, and a Youtube video of the project being ran:
https://github.com/polyworld/polyworld
https://www.youtube.com/watch?v=RvcwuzeoQR0
There are also a long series projects having neural nets learn to walk an arbitrary figure (different number of legs/body type, etc) around a 3D world with gravity applied. Here are some examples:
https://www.youtube.com/watch?v=kQ2bqz3HPJE (has a bit of narration)
https://www.youtube.com/watch?v=JFJkpVWTQVM
https://www.youtube.com/watch?v=jev4UA7EVkc (this one has good comments, but the neural nets did not converge on a solution in that video)
https://www.youtube.com/watch?v=fEM7YDNonSE
https://www.youtube.com/watch?v=LCRPcz1B8rk
https://www.youtube.com/watch?v=05Hycx1NpyE
https://www.youtube.com/watch?v=6-N9WDMjCbE
I've been playing around with WebGL recently, and I think that using that in the browser with three.js (or other high level javascript support API) can reduce the amount of code to generate a 3D scene to a minimum. Then sticking on an additional set up functions in the javascript for the neural net would create a great feedback loop in an absolute minimized environment. That's the direction I'm moving in currently, in any case.