Jump to content

  • Log In with Google      Sign In   
  • Create Account

Interested in a FREE copy of HTML5 game maker Construct 2?

We'll be giving away three Personal Edition licences in next Tuesday's GDNet Direct email newsletter!

Sign up from the right-hand sidebar on our homepage and read Tuesday's newsletter for details!


We're also offering banner ads on our site from just $5! 1. Details HERE. 2. GDNet+ Subscriptions HERE. 3. Ad upload HERE.


video and ann


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
4 replies to this topic

#1 rouncer   Members   -  Reputation: 291

Like
0Likes
Like

Posted 02 March 2013 - 11:41 AM

Ive seen individual images feedbacking out of a neural network, but ive never seen motion video.

 

Im actually attempting this with htm (heirarchical temporal memory) right now, Ive got single pictures feedbacking out of a single layer, but I havent got up to temporal yet, im having an issue with what the output message is going to be for intermediate nodes.

 

Motion video would be amazing coming out of a neural network, but there doesnt seem to be anything on the internet about it?

 

I was thinking maybe there is huge potential for photorealistic games produced from recorded footage by hacking some inputs of the ann with controls.    

 

Any info on this subject?

 

thanks.



Sponsor:

#2 Tournicoti   Prime Members   -  Reputation: 683

Like
1Likes
Like

Posted 02 March 2013 - 01:12 PM

I don't know enought about this HTM model sorry, but a Mc Culloch & Pitts based neural model is a hyperplane. You can represent, visualize it as a plane in 3D, a line in 2D, in a visual application.


Edited by Tournicoti, 02 March 2013 - 01:46 PM.


#3 rouncer   Members   -  Reputation: 291

Like
0Likes
Like

Posted 02 March 2013 - 01:25 PM

yes, I also know that neural networks model functions, like you could even get a 3d landscape out of a neural network.   but my question is how to deal with learning motion video, instead of just static characters, or static volumes.

 

heres the best pdf on htm at the moment, and yes - it is motion video, just i cant get it to work. (not yet anyway, i was just maybe seeing if i could get some pointers.)

http://bias.csr.unibo.it/maltoni/HTM_TR_v1.0.pdf

 

but im not just interested in htm, i love ann's too.

 

It is a bit of an advanced topic, like- ive seen NOWHERE with any info about it.


Edited by rouncer, 02 March 2013 - 02:32 PM.


#4 rouncer   Members   -  Reputation: 291

Like
0Likes
Like

Posted 03 March 2013 - 12:56 PM

I just got temporal working, in my htm, its now pumping out 64 frame sections of a little man running and jumping up and down.

 

htmfeedback.png

 

the matrices on the screen are the regions, the yellow numbers are how many space types its stored at that node, and the pink number is how many temporal types at that node.   A problem I have now is if I let it record for too long, the amount of space types shoots through the roof (past the capacity of the memory) so I need some way to remedy this problem.

 

But its working. smile.png


Edited by rouncer, 03 March 2013 - 03:51 PM.


#5 rouncer   Members   -  Reputation: 291

Like
0Likes
Like

Posted 03 March 2013 - 05:56 PM

heres another screenshot, its 8 bit space type memory its at its limit after ~640 frames of unique data pumped into it.

bustedcapacity.png

im gonna try raising the limit to 16 bit, and see if thats enough to fit the game inside the finite space. smile.png

 

 

heres a choppy movie of it going, it actually goes full framerate +.

 


Edited by rouncer, 03 March 2013 - 06:32 PM.





Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS