Fun Stuff by IBM


Very little actual information in the article. Would be interesting to see the development environment. At first glance it looks to me like the transputer concept updated with todays technology (which would be a good thing).
 
Last edited by a moderator:
Me too. Having a new CPU is good. But let the full machine come before going to the kind of fuzz IBM is trying to get.

If I cant try I cant trust
 
Me too. Having a new CPU is good. But let the full machine come before going to the kind of fuzz IBM is trying to get.


If I cant try I cant trust
Well it clearly is a commercial article indeed, however I like some of the concepts. But I come from a more biological viewpoint, Im interested for instance if this could operate without using the internal memory (or keeping it fixed) of the "neurons" since real neurons don't have memory, just a fireing pattern. I think we might learn one or two things via this architecture.
 
Is this some kind of dedicated, highly parallel hardware implementation of something resembling artificial neural nets? Could be interesting, but is probably going to be hard to program.
 
Title is deceptive, nothing fun happens at IBM, especially in my office space.
 
Last edited by a moderator:
manditory fun!
Grumpy-Cat-Yay.jpg
 
Last edited by a moderator:
Is this some kind of dedicated, highly parallel hardware implementation of something resembling artificial neural nets? Could be interesting, but is probably going to be hard to program.
I'm probably wrong, but it doesn't look like a neural net from a (cursory) glance. Rather like a lot of tiny programmable cores with (smart) hierarchical interconnects (which isn't bad, but old news).
 
It is a hardware implementation of a neural networking software called "Compass". The applications are huge - it can detect in real time people, cars, bicycles etc from a video stream - and can be thought of being somewhat massively parallel. It only runs at about 1KHz clock frequency, but the massive nature of the computations means that it knocks current multi-core operations into a cocked hat.

As it's a hardware implementation of a current software solution, it is programmed in the same way and there are large libraries of software already available for it.

D.
 
Me too. Having a new CPU is good. But let the full machine come before going to the kind of fuzz IBM is trying to get.


If I cant try I cant trust
Well it clearly is a commercial article indeed, however I like some of the concepts. But I come from a more biological viewpoint, Im interested for instance if this could operate without using the internal memory (or keeping it fixed) of the "neurons" since real neurons don't have memory, just a fireing pattern. I think we might learn one or two things via this architecture.
From a quick scan of the linked paper, I can tell you it claims it's a turing complete system overall, but personally I doubt each corelet is anything like a standard turing-complete CPU. Looking at the code examples, I've not seen any evidence of on-board memory, other than that needed to store connection settings (or perhaps that's in the network side in the implementation, dunno) and rules that define a pattern.


Where it differs from a standard neural net is in how it sets up the connections - or how it learns. The network is hierarchical to a certain extent and programs can direct their connections to any corelet in the local network, but have less control over connections to other networks. I haven't read up on how those connections are made, and whether they're established using machine learning rules automatically, or whether that has to be coded in at a configuration level. But each network therefore could also implement a learning algorithm if they were programmed to do so internally, but the intention I suspect is to at least have some networks dedicated to specifically engineered tasks, perhaps directing the learning of other systems, or perhaps doing specific things like decoding video streams for other networks to act on, to be the eyes of the system.
 
IBM has had many great achievements.

http://www.alfar.com/~amanda/pc110/specs.html

Pandora:  140 x 83.5 x 29 with 4.3" 800x480 screen made in 2008.

PC110:     158 x 113 x 33 with 4.7" 640x480 screen made in 1996.

PC110 had two type II PCMCIA and a CF port.  It booted PC DOS off of 4MB of Flash and ran on 8MB RAM on it's 486-33 CPU.

Considering the volume required for the PCMCIA and CF ports compared to the volument required for two SD slots - that IBM handheld was one densely packed machine.  I wonder how many layers are in that motherboard...  http://www.alfar.com/~amanda/pc110/pcb1.gif

Now they have a chip to simulate a neural cluster.  Amazing.
 
IBM has had many great achievements.

http://www.alfar.com/~amanda/pc110/specs.html

Pandora:  140 x 83.5 x 29 with 4.3" 800x480 screen made in 2008.

PC110:     158 x 113 x 33 with 4.7" 640x480 screen made in 1996.

PC110 had two type II PCMCIA and a CF port.  It booted PC DOS off of 4MB of Flash and ran on 8MB RAM on it's 486-33 CPU.

Considering the volume required for the PCMCIA and CF ports compared to the volument required for two SD slots - that IBM handheld was one densely packed machine.  I wonder how many layers are in that motherboard...  http://www.alfar.com/~amanda/pc110/pcb1.gif

Now they have a chip to simulate a neural cluster.  Amazing.
wow that's cool!
 
Back
Top