New Neural Hardware from the FACETS Project

Posted 26 Mar 2009 at 22:16 UTC by steve Share This

In 2007, we told you about Stanford's Neurocore chip which sported a 256x256 array of neurons, each designed to duplicate biological neurons down to the level of protein pores and ion channels. A new generation of neural network hardware is being developed by the Fast Analog Computing with Emergent Transient States (FACETS) Project. The FACETS hardware will have 200,000 neurons with 50 million synapses built on a single silicon wafer. Current prototypes are running 100,000 times faster than their biological couterparts, allowing a full day of neural activity to be simulated in less than a second. Physicist Karlheinz Meier, who coordinates the project, said,

We may then be able to make computing devices which are radically different and have amazing performance which, at some point, may approach the performance of the human brain – or even go beyond it!”

Aside from being much faster than computer simulations of neural networks, neural hardware is also much more power efficient, typically using only a tiny fraction of the power consumed by attempts to simulate large networks on super computers such as the Blue Brain Project. For more, see the Technology Review article, the Science Daily article, or check out the many technical papers available from FACET. (image courtesy of FACETS Project, Ruprecht-Karls Universität Heidelberg).

We just had a baby! This news is huge!, posted 27 Mar 2009 at 22:59 UTC by The Swirling Brain » (Master)

I think if this thing really works, it will be ultra huge news for robotics! I can't wait to play with this thing.

Only thing will be training this thing will be a bummer. Hopefully, they'll make it so that the intelligence can be sucked out and saved or reloaded, otherwise, every time you initialize a robot it will start off being a baby. For some tasks that would be good and others not so good depending on your project. Depending on how long it takes to learn (days, weeks, years), it may be impractical to reset. If you want to sell a robot with the smarts already in it, you'll need to be able to train the brain before-hand and then copy the smarts to a backup drive or another chip. One can hope it has the ability to backup its smarts.

Think of all of the crazy issues you can have with a brain that can modify itself? What if it truly changes its mind and decides to go postal terminator and kills everyone in the room? This thing could become good or evil!!! With hard coded programs the responsibility lies with the programmer, but with this thing, if really works like a real brain, then who knows what to expect! Wow! Exciting and Terrifying at the same time!

See more of the latest robot news!

Recent blogs

30 Mar 2017 shimniok (Journeyer)
25 Mar 2017 AI4U (Observer)
18 Mar 2017 mwaibel (Master)
25 Feb 2017 steve (Master)
16 Aug 2016 Flanneltron (Journeyer)
9 Jul 2016 evilrobots (Observer)
27 Jun 2016 Petar.Kormushev (Master)
2 May 2016 motters (Master)
6 Nov 2015 wedesoft (Master)
10 Sep 2015 svo (Master)
Share this page