Slashdot Mirror
IBM Creates Custom-Made Brain-Like Chip
An anonymous reader writes In a paper published Thursday in Science, IBM describes its creation of a brain-like chip called TrueNorth. It has "4,096 processor cores, and it mimics one million human neurons and 256 million synapses, two of the fundamental biological building blocks that make up the human brain." What's the difference between TrueNorth and traditional processing units? Apparently, TrueNorth encodes data "as patterns of pulses". Already, TrueNorth has a proven 80% accuracy in image recognition with a power consumption efficiency rate beating traditional processing units. Don't look for brain-like chips in the open market any time soon, though. TrueNorth is part of a DARPA research effort that may or may not translate into significant changes in commercial chip architecture and function.
I think this chip could govern the US more effectively than our current Congress.
Throwing more monkeys into the mix in the vague hope of getting Shakespeare out.
What does a brain-like potato chip taste like?
It is getting hard to figure out where IBM is on chips. Arguably the 4 main chips experiencing investment are: x86, ARM, Z-Series processors and POWER series 2 of which are IBM. OTOH there is no roadmap for POWER beyond the current generation. I'd love to know is IBM getting more serious about CPUs or pulling back?
I am just curious to know whether this chip can lead to the development of artificial brains to be used by Humans in future? And, I couldn't understand why this chip will not available in the open market.
Thanks to your capitalization, I've misread it as "development of artificial brains to be used by Hamas". (Yes, I need new glasses.)
Ezekiel 23:20
"Chips" mattered in the 1970s and 1980s. "Networks" mattered in the 1990s and 2000s. But now we're in the 2010s. We have transcended "chips". We have transcended "networks". We are in the era of "neurons".
The number of neurons in the brain varies dramatically from species to species. One estimate (published in 1988) puts the human brain at about 100 billion (10^11) neurons and 100 trillion (10^14) synapses.
100 billion divided by 1 million = 100,000 of these chips to reach the human neuron count.
100 trillion divided by 256 million = 390,625 of these chips to reach human synapse count.
Assuming Moores Law for these chips with a doubling every 24 months to be conservative.
2 of these on a chip in 2016
4 of these on a chip in 2018
8 of these on a chip in 2020
16 of these on a chip in 2022
32 of these on a chip in 2024
64 of these on a chip in 2026
128 of these on a chip in 2028
256 of these on a chip in 2030
512 of these on a chip in 2032
1024 of these on a chip in 2034
2048 of these on a chip in 2036
4096 of these on a chip in 2038
8192 of these on a chip in 2040
16384 of these on a chip in 2042
32768 of these on a chip in 2044
65536 of these on a chip in 2046
131072 of these on a chip in 2048
262144 of these on a chip in 2050
So we could be seeing human brain capabilities on a chip by mid century. Quite possible we'd see similar capabilities built as a supercomputer 10-20 years before that. Don't flame for the wild assumptions I'm making here - i know there are a lot, this is just intended as some back of the envelope calculations.
nà posebnega...
Is it coming soon?
Assuming of course this chip can hold 2 hr conference calls with 40 other chips and pound out 240 page Powerpoints.
Tinman Abort!
Don't look for brain-like chips in the open market any time soon, though. TrueNorth is part of a DARPA research effort that may or may not translate into significant changes in Real Dolls.
When can we hope to see reports of the first brown-faces executed by drones incorporating this exciting new technology?
Where are all the HAL 9000 jokes? HAL was built by IBM in "2001: A Space Odyssey", perhaps this is an example of life imitating art?
I couldn't understand why this chip will not available in the open market.
The millitary boys want to keep their toys secret to prevent the enimies getting them.
"Human brain has around ten-to-the-tenth neurons. By third year Mike had better than one and a half times that number of neuristors. And woke up." -- The Moon is a Harsh Mistress
A.
...bringing you cynical quips since 1998
I agree i your initial statement, but that's pretty much as it has been for at least 15 years or so. POWER9 is on the roadmaps, and the next generation zArch too. And they are sitting there like proxy boxes with nothing much spced, like it has been for almost all previous generations of their predecessors. What I'm concerned with is the lack of public roadmap for what they are planning in the HPC and super computer space. We had the very public Blue Gene project that began in 2001 with four projects; C, L, P and Q, but since the Blue Gene/Q came to life a couple of years ago, I have no idea what they are planning. It'd be nice to have some clue here.. Why not something from the OpenPOWER Foundation; A P8 host processor with integrated GPU from nVidia, on chip networking from Mellanox and programmable accelerators from Altera. But I haven't seen anything in that direction.
- Henrik
- when the Shadows descend -
Is this pulse recognition closer to Morse code or bar code? Obviously pulse code recognition could go beyond binary bit codes but the hardware must be seriously different from what we have now. It could even be analog.
. . . I wacky-parsed the title as: "IBM Creates Custom-Made Brain-Like Chimp.
. . . so just imagine where that thought train derailed me . . .
Schroedinger's Brexit: The UK is both in and out of the EU at the same time!
New glasses? What you really need is to turn off the television.
One estimate (published in 1988) puts the human brain at about 100 billion (10^11) neurons and 100 trillion (10^14) synapses.
http://en.wikipedia.org/wiki/Neuron
I'm sorry Dave. I'm afraid I can't do that.
If telephones are outlawed, then only outlaws will have telephones.
Imagine a Beowulf Cluster of these!
As good as any other theory, I guess.
Best Slashdot Co
>"Already, TrueNorth has a proven 80% accuracy in image recognition"
This is where the whole thing falls down. Software can emulate any hardware you build, even quantum. So the accuracy of a particular image recognition algorithm is irrelevant, given that you didn't need this processor for it, and the implication seems to be that this processor has to use a neural net of some kind. Neural nets are known to be less efficient and accurate than tuned algorithms on the same hardware.
Be wary of people citing soft output accuracy ratings of hardware chips. Now if someone were comparing the software implementation to the hardware implementation of a particular implementation, it might demonstrate performance improvements...
A commodore 64 running NATO COMMANDER on a datasette could run the George W Bush thought process.
Every time one of these damn 'neural computers' come out people tend to equate the number of neurons and synapses and think 'hey, if we can get to the number of human neurons... Presto!!!!1'
Brains are waay more complicated than just neurons and synapses. Just taking the neurotransmitters into account makes the whole charade crash down. Then there is the glial network that, surprise surprise, does an enormous amount of complex work. There's even recent research suggesting that the branching patterns of the neurons perform complex computations. There are chemical gradients in the brain that act as a sort of addressing system.
tl;dr Brain on a chip? Yeah fucking right.
Can't Purina use them in Zombie Chow? If so, I would rather feed that to the neighborhood zombies instead of my own gray matter?
Hardware advancements will be useful. But the old problem remains.
The problem is getting the pattern of this cellular logic needed ( typical nerve type operations are summations of positive and negative weights from broad interconnects (synapses) and layers of neurons) for whatever specific problem space is to be handled.
You grow/train into the pattern using training data sets over and over that feed corrections backwards through the layers (you supply the inputs and tell it what the right outputs are and the logic slowly forms to process that way). A key problem is that the more complex the net is the more likely that formation fails and wont produce the processing results desired.
Fixed problem spaces work (Kanji Alphabet stroke pattern hand input reading was an early success - with its limited standard input forms that people can adust to use). One logic pattern built up (months to get it right) then cloned into a multitude of devices that all work the same way on the same expected inputs.
As the article said they DIDNT include the learning system in the hardware. THAT is a mechanism far more complicated to implement (much of your brain's cells and operation goes to implement the building and modifying (and maintaining) of the patterns stored/formed in your neurons/synapses.
Unfortunately THAT is the part you need to get to work the most to have a useful learning brain-like function.
Beyond that, the real problem still is : WHO is going to tell the 'brain' what the result should be (and if it is right or wrong)? If you cannot feed fixed training data into a system like this (mass input set with results/conclusions/outputs precalculated ) and you try to do it dynamically, users (whoever its being customized for) will become VERY VERY tired of correcting the system over and over. New input-output patterns potentially can disturb what is already formed and remedial 'learning' is required to restore the previous functionality The disruption is ongoing and increases the more logic patterns get added.
So this really only counts as an early partial step in what such systems will have to be comprised of - the summations hardware is actually trivial compared to the learning processing which will be required.
or skynet for that matter...
> TrueNorth has a proven 80% accuracy in image recognition
A related project is working on detecting pornography in images. It's called True Peter North.
I'd also like IBM to tell us how it compares to Kwabena Boahen's chips, which have been eating IBM's lunch in this space for a decade.
I am just curious to know whether this chip can lead to the development of artificial brains to be used by Humans in future?
That's easy: No.
Kurzweil is the modern equivalent of a televangelist.
Required reading for internet skeptics
Let me know when a computer can "see" with a pair of cameras. Identify an object heading toward the cpu(not just the cameras) and adjust its motors to dodge the incoming.
That actually do already exist.
It's a car's collision avoidance system.
It's already standard option from some manufacturer (e.g.: Volvo) (and should become mandatory in EU somewhere soonish).
Some like Mobileye rely entirely on camera, while other are integrating other sensors in the mix, like radar, infra red lasers, etc.
But yeah I see your point: complex task require complex network, way much more than this chip.
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]