[nengo-user] Fwd: Eliasmith's TED talk

Alan Grimes ALONZOTG at verizon.net
Wed Feb 12 02:20:23 EST 2014


###############################
-------- Original Message --------
matthewdeutsch at hotmail.com

A dude on Facebook posted this. Figured I should tell you about it
since you were flipping a $#!+ over this guy.

http://www.youtube.com/watch?v=g2HHJfovb5E 		 	   		  

#################################################


Mathew, you need some industrial strength IA...



Lets strain ourselves to imagine that Eliasmith is even slightly wrong about some tiny part of his
design. Even in that case, the framework that his software establishes should be general enough to still 
be incredibly useful, ie you would just have to swap out a class here and there....

Now, the question becomes: What can we do about its performance? 

We could just wait for The Big Weenie to come out with his AI, but then that's lame... (I just read an
article today on Russia 2045 confirming that Google is in an AGI sprint). 

Eliasmith himself is interested in attempting to apply neuromorphic hardware, essentially using analog 
computing... (Maybe Steve should get in touch with him...) 

But doing that seems to give back many of the advantages of having an AI, ie I doubt that analog states 
will be easy to precisely copy between machines. 



Here's my current idea: 

Stick one of these in my desktop (currently has 32GB of ram! ) 
http://www.xilinx.com/products/boards-and-kits/EK-V7-VC707-G.htm
(Actually, there's a new vertex coming out with like ten times as many gates!) 

Then install a fiber optic 10gbe transceiver, run the cable through the wall, down the stairs, 
along the ceiling, through the bathroom, and finally to the rack in the basement...

Then load up the rack with one or even two of these:

http://www.ncorehpc.com/browndwarf/

That should have enough of the right kind of compute to make Nengo run noticeably Less Slow (tm). ;) 

Teslas are faster but they aren't as general in terms of the types of algorithms you can run on them. 

Anyway, when I finish running my current batch of number theory codes, I'll get serious about trying 
to get Spaun running on the hardware I have now. 





More information about the nengo-user mailing list