Slashdot Mirror

← Back to Stories (view on slashdot.org)

SpiNNaker Powers Up World's Largest Supercomputer That Emulates a Human Brain

Posted by BeauHD on from the up-and-running dept.

The world's largest neuromorphic supercomputer, the Spiking Neural Network Architecture (SpiNNaker), was just switched on for the first time yesterday, boasting one million processor cores and the ability to perform 200 trillion actions per second. HotHardware reports: SpiNNaker has been twenty years and nearly $19.5 million in the making. The project was originally supported by the Engineering and Physical Sciences Research Council (EPSRC), but has been most recently funded by the European Human Brain Project. The supercomputer was designed and built by the University of Manchester's School of Computer Science. Construction began in 2006 and the supercomputer was finally turned on yesterday.

SpiNNaker is not the first supercomputer to incorporate one million processor cores, but it is still incredibly unique since it is designed to mimic the human brain. Most computers send information from one point to another through a standard network. SpiNNaker sends small bits of information to thousands of points, similar to how the neurons pass chemicals and electrical signals through the brain. SpiNNaker uses electronic circuits to imitate neurons. SpiNNaker has so far been used to mimic the processing of more isolated brain networks like the cortex. It has also been used to control SpOmnibot, a robot that processes visual information and navigates towards its targets.

98 of 164 comments (clear)

  1. Suprising... by Anonymous Coward · · Score: 5, Interesting

    ...that the article does not mention that the project lead, Steve Furber was one of the team at Acorn that created the original ARM chip back in the 80s.

  2. "switched on for the first time" vs "has been used by Anonymous Coward · · Score: 2, Interesting

    Which is it, genius editors ?

  3. All hype, no content by mark-t · · Score: 1
    Really, if it could *really* emulate the human brain and they switched it on yesterday, it should have done *something* by now that was worth reporting...

    And with only million cpu's, isn't that a few orders of magnitude to small to emulate a human brain anyways, which has hundreds of billions of neurons?

    --
    File under 'M' for 'Manic ranting'
    1. Re:All hype, no content by UnknownSoldier · · Score: 1

      > only million cpu's, isn't that a few orders of magnitude to small to emulate a human brain anyways, which has hundreds of billions of neurons?

      Yup, this "simulation" is off by an order of magnitudes.

      The brain is estimated to have 86 Billion Neurons; the number of connections even higher.

      Trying to use inorganic matter to simulate consciousness is a fools errand. They should start with bio-organic computing instead -- they would have better luck.

    2. Re: All hype, no content by jd · · Score: 1

      Not really. It'll be simulating a few neurons in the brain at one millionth speed. These things are for medical research, not AI.

      --
      It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
    3. Re:All hype, no content by Henriok · · Score: 4, Insightful

      A human brain rarely does something spectacular the first days of being turned on.

      --

      - Henrik

      - when the Shadows descend -
    4. Re: All hype, no content by mentil · · Score: 2

      That's what it WANTS you to think.

      --
      Corruption is convincing someone that the selfless ideal is the same as their selfish ideal.
    5. Re:All hype, no content by ShanghaiBill · · Score: 4, Insightful

      > only million cpu's, isn't that a few orders of magnitude to small to emulate a human brain anyways, which has hundreds of billions of neurons?

      No. There is no reason that you need 1 cpu per neuron.

      A biological neuron fires 200 times per second. A single core can simulate the firing of 10,000 neurons per second.

      Also, there is no reason the simulation needs to be real-time, so the speed isn't really that important.

    6. Re:All hype, no content by q_e_t · · Score: 1

      You don't need to have a 1:1 mapping of neurons to CPU cores. A CPU core is capable of emulating multiple neurons, in fact thousands of them, as the activation rules of neurons are not individually particularly complex. Simplisticly, the complexity is in being able to efficiently map connections between neurons, the effect of neurotransmitters and other support structures on the functioning, different types of brain area, and being able to make it function efficiently as a whole. There are plenty of papers available on SpiNNaker to read.

    7. Re:All hype, no content by q_e_t · · Score: 1

      P.S. I am not sure how many neurons/CPU SpiNNaker is emulating, or exactly what activation functions it emulates, but the activation functions are not normally considered to be that complex, hence being able to get more than one neuron/CPU. However, any one of those neurons could have connections to any other neuron. Normally with parallel programming you aim to pack a subdomain of the problem onto a CPU, so as to minimise communication, to make it efficient. How SpiNNaker solves the rather more general problem I don't know. Also, AFAIK, these are not the sort of cores you normally find in a supercomputer designed for computational fluid dynamics, weather prediction, chemistry, etc., given the titles of papers I see coming out from the team on ResearchGate (not that I've read them all, but I do 'follow' the research on ResearchGate).

    8. Re:All hype, no content by religionofpeas · · Score: 4, Insightful

      Trying to use inorganic matter to simulate consciousness is a fools errand.

      Why ? Unless you can point out some fundamental limitations, it's nothing but an argument from incredulity. It's like saying we can only make a functional wing from feathers, and not aluminum.

    9. Re:All hype, no content by sheramil · · Score: 1

      And in a years' time, SpiNNaker will have asked, very slowly, "Why... was... I... named... after... part... of... a... boat?"

    10. Re:All hype, no content by religionofpeas · · Score: 1

      First of all, we do not understand how or why consciesness exists, and that is a fundamental limitation.

      Like I said, an argument from incredulity. If you admit you don't understand how it works, then you cannot claim anything about what's needed to implement it.

    11. Re:All hype, no content by Henriok · · Score: 1

      It takes many many days, some say nine months until a human barely survives without intense support from a highly trained and specilized host. And at least several more months, some say years, until it does anything productive. All the while it's constantly updated with new hardware. So.. shall we give this project a bit more time before we deem it useless?

      --

      - Henrik

      - when the Shadows descend -
    12. Re:All hype, no content by Anonymous Coward · · Score: 1

      The analogy that a human brain is like computer hardware, and the mind is like computer software, is fundamentally flawed reasoning. Mind is not software or code that is executed by a processor, and computers cannot run without some kind of software, and software does not arise spontaneously from computer hardware.

      Read about the Chinese Room experiment, and try not to pat yourself on the back so much for handwaving away valid argument without any valid counter-argument whatsoever. Find a single logical contradiction, and you will win. Strong AI, a machine-mind, cannot and will not ever exist. As far as all science and technology and all of history is concerned, mind requires biological, healthy and living brain to exist. There is not of shred of evidence of the contrary.

    13. Re:All hype, no content by religionofpeas · · Score: 1

      try not to pat yourself on the back so much for handwaving away valid argument without any valid counter-argument whatsoever

      The original claim was "Trying to use inorganic matter to simulate consciousness is a fools errand.". The person making that claim needs to provide a good argument for it, probably starting with a good definition 'consciousness', and what is exactly the difference between organic/inorganic matter.

      Read about the Chinese Room experiment

      Yeah, I am very familiar with it. It's completely bogus. Let me offer the Chess Room experiment as an example. You sit in a room with the full binary code of Stockfish, and a few billion pieces of scrap paper. A human grandmaster sits outside, writes his moves on a piece of paper and slips it in the room. You simulate an x86 processor running through the Stockfish code. After a few billion years of mindless simulation, you send a reply on a piece of paper. Do you now understand chess ? If not, you have just experienced the flaw in the Chinese Room argument.

    14. Re:All hype, no content by q_e_t · · Score: 2

      Implementing neurons directly in hardware is problematic, as it means you lose (1) flexibility to change the implementation, and (2) economies of scale from using commodity off-the-shelf components. You can offset that somewhat by using FPGAs, which gives you some economy of scale, along with flexibility. There have certainly been a number of neuron-like hardware systems in the past (e.g. CAM, although that's very simplistic), and other systems which are weight adders. I've worked on such systems. However, you can simulate some of it effectively, in terms of cost:compute on GPGPUs and other systems, even if it's overall more efficient to use weight adders directly. Steve Furber's taken a different route to using GPGUs, possibly because of the interconnect topology issues - i.e. if you have a Minsky machine (named after the founder of modern AI) you have a lot of GPGPU horsepower, but apart from NVlinks your distributed connectivity is going to be big box to big box. If you are building something at the scale of SpiNNaker, I doubt that would work, especially if you wanted something hyper-connected.

    15. Re:All hype, no content by q_e_t · · Score: 1

      First of all, we do not understand how or why consciesness exists.

      What makes you think SpiNNaker is necessarily designed to 'discover' consciousness?

    16. Re:All hype, no content by dinfinity · · Score: 3, Interesting

      Thank you.

      The Chinese Room is one of the worst thought experiments ever to have entered the discourse on functionalism. In its basic form it fails miserably at answering simple questions such as "how many fingers am I holding up?"
      When challenged with the requirements to answer such questions, staunch supporters then modify the Chinese Room again and again until the person in it is reduced to nothing more than a hand writing the results of a complex processing system on a piece of paper. Given that no one reasonably assumes understanding or consciousness of a hand, the argument against functionalism has then successfully defeated itself.

      A much, much more interesting thought experiment is that of the China Brain. It is really, really hard to wrap your head around how consciousness would exist in a collection of scattered scraps of paper in possession of billions of individuals. The role of space and time in the constitution of our own consciousness become very important in that analysis.

    17. Re:All hype, no content by cascadingstylesheet · · Score: 2

      Trying to use inorganic matter to simulate consciousness is a fools errand.

      Why ? Unless you can point out some fundamental limitations, it's nothing but an argument from incredulity. It's like saying we can only make a functional wing from feathers, and not aluminum.

      You keep saying that ... and AI keeps being just around the corner ...

    18. Re:All hype, no content by Oligonicella · · Score: 1

      You may have a point, they may be able to simulate the firing of 10K, but I'll give you something else to explain. We have no idea how or why neurons take in their various inputs, weigh them and send impulses to a select few of their outputs. Exactly how does the programming account for that?

    19. Re:All hype, no content by Oligonicella · · Score: 1

      You are so very wrong. Read up.

    20. Re:All hype, no content by Oligonicella · · Score: 1

      The flexibility alone is enough. Most people who haven't studied the brain don't realize that dendrites aren't physically "locked" but change.

    21. Re:All hype, no content by Oligonicella · · Score: 1

      but the activation functions are not normally considered to be that complex

      Cite please.

    22. Re:All hype, no content by sfcat · · Score: 1

      if you have a Minsky machine (named after the founder of modern AI)

      Marvin Minsky ABSOLUTELY didn't found modern AI. He was a successful author but none of his lines of research panned out and nothing in AI currently is based upon his ideas. MIT's AI labs floundered terribly under his leadership and he is a big reason why for all of MIT's success, its AI labs aren't up to the standards of the rest of the CS dept there. I can tell you haven't spent any time in academic AI if you think this at all. In fact, this is one way we use to filter out the fakers.

      --
      "Those that start by burning books, will end by burning men."
    23. Re:All hype, no content by kmoser · · Score: 1

      > only million cpu's, isn't that a few orders of magnitude to small to emulate a human brain anyways, which has hundreds of billions of neurons?

      Yup, this "simulation" is off by an order of magnitudes.

      The brain is estimated to have 86 Billion Neurons; the number of connections even higher.

      So, then it's a simulation of several hundred Congressman's brains?

    24. Re:All hype, no content by Layzej · · Score: 1

      You keep saying that ... and AI keeps being just around the corner ...

      Have you been living under a rock? AI is used in everything from fraud detection, natural language processing, self driving cars, customer service, customer retention, automated detection and classification, etc, etc, etc.

    25. Re:All hype, no content by q_e_t · · Score: 1

      Indeed. I mentioned this to my wife and suggested that a fully physically analagous implementation would need dozens of people to swap around an old-style telephone exchange patch board.

    26. Re:All hype, no content by q_e_t · · Score: 1

      Traditionally, the activation function was f(W) where W is the set of incoming neuron activations over the synaptic connections, such that a=f(W) where a is an activation level, then the neuron fires if a>A, but now it's more common to use spiking neurons, i.e. a time-series element such that, say, a neuron that's had some pattern of events come in, then fires. That's what SpiNNaker uses, see https://en.wikipedia.org/wiki/... is an initial bit of reading, which includes "SpiNNaker (Spiking Neural Network Architecture), designed at the University of Manchester, uses ARM processors as the building blocks of a massively parallel computing platform based on a six-layer thalamocortical model.[5]". You could probably consider it, at any instant, to a'=f(W,a)-d(a) where W is the incoming signal at the current instant, a is the activation, and d is some decay, probably a function of a, with the neuron firing when a > A, although I don't know what mathematical formalism they use, and there are probably many more details. Quite how SpiNNaker implements that in detail (synchronous, asynchronous, etc).

      I should go and read the papers.

    27. Re:All hype, no content by q_e_t · · Score: 1

      In fact, this is one way we use to filter out the fakers.

      Opinions differ, obviously, http://thedatascientist.com/ma....

    28. Re:All hype, no content by burtosis · · Score: 1

      Have you been living under a rock? AI is used in everything from fraud detection, natural language processing, self driving cars, customer service, customer retention, automated detection and classification, etc, etc, etc.

      All of those cases are narrow/weak AI and not general purpose/strong AI. It's actually the distinction to help people understand we don't have working self driving cars today - the number of fringe edge cases and level of abstraction needed is beyond our current abilities. General purpose AI has an innate human/animal level common sense notion of the world and is self aware.

      I always thought the chinese room/Turing argument against strong AI didn't frame the question properly. Following a set of operations is what human and animal brains do, on levels like molecular, electrical, and cellular. Imagine for simplicity some are joined to form these "rooms" and that some processes perform executive oversight, perhaps even somewhat recursively could be considered rooms. Saying understanding has to occur in a particular room is like saying an economy has to be found within a single business office in a city. The economy is a emergent property of tens of thousands of rooms within a city and many cities each linked together across nations or the world and so is wholly found in no rooms and yet has a part in each. Consciousness is likely similar in that is the emergent behavior once a critical number of rooms with the right contents are connected.

    29. Re:All hype, no content by religionofpeas · · Score: 1

      Consciousness is likely similar in that is the emergent behavior once a critical number of rooms with the right contents are connected.

      Consciousness is not like a light that can be on or off. It's more like a bag of different tools and tricks. You can have a few of them, and you'd have a limited form of consciousness, or you can have a large bag of human-like tools, and have a human level consciousness. For AI applications, it is most often not desirable to give the machine a large bag with tools it doesn't need. There's no need for a self driving car to be curious or get bored, for example.

    30. Re:All hype, no content by Matheus · · Score: 1

      Thank You!

      Was hoping this had already been said.. "Designed to simulate the human brain" and "simulating consciousness" are 2 *very different things. The human brain at its core is an absolutely amazing pattern matcher where pattern space is unlimited. We've come a LONG ways in making computers / software that is "good" at this but we're still a fair ways off from concepts such as intuition or the concept of applying a lifetime's worth of experiential, sensory input and possibly synthetic/imagined patterns to "unlike" situations to reason a favorable outcome to an unfamiliar scenario. Simulating such a capability in machine form implies the ability to advance that capability beyond our brain's "processing capacity" to make advancements/predictions that were previously unattainable.

      That all being said if we really attain that goal maybe we find that consciousness isn't that far off... You can spell most of Skynet with SpiNNaker... ;)

    31. Re:All hype, no content by raftpeople · · Score: 1

      The dendrites perform complex non-linear computations prior to forwarding signals to the main cell body (achieved via local/regional spiking forward and backward throughout the dendrites) that scientists are just now discovering and trying to figure out.

      In addition, glial cells manage the synapse activity (see tripartite synapse) and regional groups of neuron, you can't ignore that functionality.

    32. Re:All hype, no content by UnknownSoldier · · Score: 1

      It is obvious you aren't a programmer. You aren't thinking about it from a programmer's debugging perspective. Would you rather:

      * Start from a simpler base that ALREADY works (such as an Earthworm) and trying to figure out how the pieces work, or

      * Start from complexity literally billions of order complicated and TRY to debug that???

      Just to put the connections into perspective:

      * Each neuron may be connected to up to 10,000 other neurons,
      * The minimum total number of connections is estimated to be 100+ trillion. The number of these synapses are at least 1,000 times the number of stars in our galaxy. Yeah, good luck simulating THAT !

      > It's like saying we can only make a functional wing from feathers, and not aluminum.

      No, that analogy is flawed.

      When reverse engineering you ALWAYS start with something that ALREADY works.

      You don't start from scratch and "HOPE" it "eventually" works when you don't have a way to determine if it is or isn't working correctly. That is the height of stupidity.

      Theory ALWAYS comes AFTER application.

      There are a few big problems in Science at the moment:

      * Scientists don't have a fucking clue what Consciousness is. There is (currently) no ability to measure it, store it, load it, etc. Without a way to MEASURE it, HOW do you know if what you are doing is moving towards or away from the goal post??? Science will advance when Consciousness is part of the equations.

      * Scientists are under the delusion that Consciousness is this magically emergent property. This is the typical ignorance of man -- viewing things ass-backwards.

      e.g. Rene Descartes got it completely backwards. It is NOT I think, therefore I am but I AM, therefore I (may) think. (It is obvious Descartes never spent any time learning how to meditate without any thoughts for any significant period of time..)

      The actuality is that Consciousness is the Foundation of reality -- not energy. Peter Russel does a great job explaining this in his Primacy of Consciousness talk. Anyone who has had a shared OBE can tell also tell you this -- but unless you have had one you don't have a frame of reference to understand the implications of this -- Consciousness has the ability to operate outside the confines of space-time. Tom Campbell's My Big Toe is a REALLY interesting "map of the territory" so to speak that goes into more detail.

      * Trying to use a Linear process to understand a Non-Linear system will never work. Scientists have yet to (re)discover that Consciousness is Non-Linear due to the Property of Free Will. /sarcasm Yeah, good luck trying to use a deterministic system to modal THAT.

      * Mind != Brain. There have been dozens of experiments showing the Non-Locality of Mind. Using a Linear system will never modal that.

        But like Max Planck said:

      Science Advances One Funeral at a Time.

      The entire approach to the solution (and problem) to AI, like creating Artificial Life, is all wrong.

    33. Re:All hype, no content by UnknownSoldier · · Score: 1

      Gloried Table Lookup is NOT Artificial Intelligence; more like Artificial Ignorance.

    34. Re:All hype, no content by religionofpeas · · Score: 1

      No, that analogy is flawed. When reverse engineering you ALWAYS start with something that ALREADY works.

      The analogy was about the requirement to have brains built out of organic matter. You're talking about something else. Next time, try to address the actual argument.

      Trying to use a Linear process to understand a Non-Linear system will never work.

      That's why all neural models are non-linear.

      Without a way to MEASURE it, HOW do you know if what you are doing is moving towards or away from the goal post???

      Quite simple. You just focus on the behavior. You can measure the inputs and outputs, and if they get closer to the behavior of a real brain, you know you're getting closer to the goal post. That's how evolution shaped our brain after all, simply by looking at the outputs and see if they benefit survival and reproduction.

    35. Re:All hype, no content by burtosis · · Score: 1

      I'd agree, it not binary yes/no but rather a spectrum as real life applications often are. Unfortunately for self driving cars it's not so simple as lane following or constant distance car following. Edge cases make it difficult to have a narrow AI back end learn it all, but rather you would need seperate processes for each edge case class and an executive function moderate them and seamlessly fuse behavior with core driving algorithms. Basically a way to do it would be have a variety of narrow AI modules and an executive AI module or module to oversee the system and we are still learning quite a bit about this approach.

      I'd almost disagree in a sense as there are some executive functions analogous to humans that may be useful like curiosity - finding suspicious, weird, or other unexpected behaviors/events and devoting additional processing to it, perhaps centrally to help all vehicles perform better. Pretty much all animals including humans get curious so there may be an underlying reason why this or a version thereof is beneficial we aren't aware of.

    36. Re:All hype, no content by currently_awake · · Score: 1

      They should start small, say a rat brain. There are plenty of rats to take apart and figure out how their brain works, and it's cheaper to work at a smaller scale. We assume most of the underlying principles are the same (we don't actually know how brains work), so there is no disadvantage of starting small. Once they have a working (silicon) rat brain they can scale it up to a useful level (replacing human workers, then discovering that the unemployed don't buy stuff but can still vote to nationalize the means of production).

    37. Re: All hype, no content by ceoyoyo · · Score: 1

      Even people who realise they change often donâ(TM)t realise how fast they can change. I have colleagues who have looked at exercise in rats that increases the volume of certain brain structures. That happens over very short time scales... hours. Turns out itâ(TM)s actually mostly due to dendritic growth.

    38. Re:All hype, no content by thegarbz · · Score: 1

      This. I always laugh at videos of newly born animals falling over trying to take their first steps. And then cry a bit when I realise the years it takes us the dominant species to achieve the same feat.

    39. Re: All hype, no content by ceoyoyo · · Score: 1

      One should be careful about absolute statements....

      There are some very fascinating results coming out of ai research of all types. A project that I believe is related to this one simulated a bunch of neurons and cortical structure and discovered that the system spontaneously produced an oscillating pattern that was similar to whatâ(TM)s observed in the actual cortex.

      At the other end of the complexity spectrum, if you randomly create small convolutional kernels, most of them will be sensitive to edges.

      âoeSoftwareâ may well be able to arise spontaneously given the right basic structure, input, and feedback.

    40. Re: All hype, no content by ceoyoyo · · Score: 1

      Itâ(TM)s been done actually. I canâ(TM)t rmemeber whether it was a rat or a mouse. Also, a cockroach and at least one kind of worm.

    41. Re:All hype, no content by Layzej · · Score: 2

      "Intelligence is whatever machines haven't done yet.”

      If and when artificial intelligence surpasses human intelligence, people might conclude that there is no such thing as intelligence. Or they might simply redefine intelligence as "whatever humans haven't done yet” as they try to catch up with AI. - Tesler

    42. Re:All hype, no content by RespekMyAthorati · · Score: 1

      They should start small, say a rat brain.

      Neuroscientists are currenty trying to get a grip on the nervours system of c. elegans, a tiny worm with a grand total of 300 neurons.

      They have a long way to go, as that even after mapping those 300 neurons six years ago,
      the scientists involved have gained only a very limited understanding of what those neurons actually do.

      At this rate, understanding a mouse brain is decades away.

    43. Re: All hype, no content by ceoyoyo · · Score: 1

      "You know you have a mind."

      Do I? I suspect, and there is some evidence for this, that most of what we experience as consciousness is actually just story we tell ourselves. There are several lines of evidence that suggest our subconscious arrives at a conclusion through means that are entirely unknown to us, our conscious becomes aware of that conclusion, and when pressed will retrospectively make up a story justifying it.

      Except for little bits around the edges, consciousness is necessarily subjective. And highly subjective observations, particularly where hubris is involved, are strongly suspect.

      There's actually a psychologist who's suggested that people weren't properly conscious until relatively modern times. The reason mythical greeks were always talking to gods is that their conscious mind wasn't fully developed and their conscious kept whispering information to them. Some people still do talk to voices in their heads, gods or otherwise. Maybe we all do, but aren't as aware of it.

    44. Re:All hype, no content by q_e_t · · Score: 1

      I see I'm behind the times on the full complexity and need to do more reading.

    45. Re:All hype, no content by q_e_t · · Score: 1

      Skynet is the name of the UK's military satellite network, and has been since the 1970s.

  4. Raises the same questions a Touring Test does by Crashmarik · · Score: 1

    With a Touring Test
    If you have a piece of software that can pass a touring test what have you really created and what does it say about the nature of intelligence ?

    With this
    It would seem that it just validates (not a small thing) the knowledge of how an animal brain is put together, and only in very limited ways at that.

    Overall I suspect this project will tell us much more about what we don't know about how brains are put together than what we do know about how thought works.

    1. Re: Raises the same questions a Touring Test does by Anonymous Coward · · Score: 1

      Would that be a national or world tour?

    2. Re: Raises the same questions a Touring Test does by q_e_t · · Score: 1

      The Grand Tour, maybe?

    3. Re:Raises the same questions a Touring Test does by rnturn · · Score: 1

      Who wrote this "Touring Test"? AAA?

      --
      CUR ALLOC 20195.....5804M
  5. It'll simulate a small part of the brain by jd · · Score: 2

    In all honesty, I doubt it'll go much beyond the 250,000 neuron mark. Brain simulators tend to also be very slow, the ones I could find on Google could take a few hours to simulate a second of activity.

    Based on the core count versus simulation speed versus neurons, a simulator that could handle the whole brain at one second per second would be five miles in diameter and 1,500 feet high.

    That doesn't mean this simulator is unimportant. Simulating fractions of the brain in extended time will let neurologists see the effects of medical interventions. That, and not HAL, is the objective of such projects, after all.

    --
    It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
    1. Re:It'll simulate a small part of the brain by q_e_t · · Score: 1, Offtopic

      Are you assuming one neuron per core? You can have a neural network with many, many units running on a single CPU core. In that case the number of units in in the hundreds to thousands, typically, which means that neuron-to-neuron communication is relatively easy to handle as you can simply use shared memory. The trick with SpiNNaker and similar efforts is being able to marshall communication with more diverse connections and communication, and that gets complex when communications are not within the locality (physically speaking) of a neuron's computational location. An analogy (and one that Wolfram would approve of!) is to look at cellular automata (I did a fair bit of work on these in the past) where again it is possible to relatively easily do synchronous updates for CA that are locally connected, but much more complex for those that include a neighbourhood with less local connections for the update rule (cf. activation function). Sometimes using asynchronous updates can be useful, as long as they are reasonably timely, and you accept some jitter, but if the activation function is appropriate, it is possible for it to cope with the noise that asynchronicity injects. In my work that is where neural networks came in, as hand designing update rules becomes impossible, and you end up needing a neural network to learn the update complex function based on desired state transitions. But then you still need really good organisation of the distributed computing element to make if efficient in terms of computation/time and even more skill to create something that is efficient in terms of computation/energy. If you look at Steve Furber's credentials with ARM, you can see why he's a good lead for the latter, and he has assembled an excellent team working on all aspects of the problem.

    2. Re:It'll simulate a small part of the brain by jd · · Score: 1

      The more neurons you have on a core, the less processing time you have per neuron since you're running them as time-shared rather than concurrent.

      The main problem is in the synapses. Up to 3000 per neuron, self-modifying not only in terms of end-points but also in terms of amplifying signals. If you've done network simulation, you'll know that's going to eat into the clock cycles.

      https://www.telegraph.co.uk/te...

      40 minutes to simulate one second is not good. So if you want to run the simulation faster, you need to split up the tasks into smaller chunks and run them on independent cores. That's the calculation I was using. How far can you possibly subdivide the work in order to get a reasonable simulation speed?

      --
      It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
  6. A million ARM cores by schweini · · Score: 1

    If anyone else is wondering how they can afford a million cores with a budget of less than 20 million dollars: Wikipedia says they are some ARM cores in 10 19 inch racks.
    Each core is supposed to be able to simulate 1000 neurons.

  7. Re:does not compute by q_e_t · · Score: 1

    We're talking about $19.5 million for Spinnaker, headed by Steve Furber (of ARM fame), not AI in total. I know one of the team there, and they are doing really great work.

  8. Re:Just a drop in the wetware bucket by religionofpeas · · Score: 1

    Some work in the past couple of years has identified structures that might make each of those cells in our brains be the equivalent of many quibits in a quantum processor!

    The brain is too warm and noisy to exploit quantum coherence on any meaningful scale.

  9. Yeah, but.... by stealth_finger · · Score: 1

    Yeah, but can it play doom?

    --
    Wanna buy a shirt?
    https://www.redbubble.com/people/stealthfinger/shop?asc=u
  10. Yeah I hate Homos by Crashmarik · · Score: 1

    Homophones, Homographs, Homonyms all a pain in the ass.

    1. Re:Yeah I hate Homos by Anonymous Coward · · Score: 2, Funny

      The word you was looking for is Hemorrhoids.

    2. Re: Yeah I hate Homos by TimMD909 · · Score: 2

      I thought that joke was bloody obvious.

  11. Re:"switched on for the first time" vs "has been u by Anonymous Coward · · Score: 1

    The editors have not used it yet. I mean their own....

  12. Depends if you count projectile pooing... by Viol8 · · Score: 2

    ... which babies can be quite good at!

  13. Simulating neurons isn't enough by Viol8 · · Score: 1

    Its now known that the white matter in the brain isn't passive after all and does affect information processing, plus the neurons are also affected by "out of band" (for want of a better term) signals in the form of hormones. So unless you simulate all of that then at best it'll be a brain-lite even if they simulated all 100 billion neurons (which I very much doubt).

    1. Re:Simulating neurons isn't enough by religionofpeas · · Score: 1

      According to the article, this machine also simulates chemical signalling.

    2. Re:Simulating neurons isn't enough by Oligonicella · · Score: 1

      Cells are not connected like electrical circuitry. They don't touch but send a chemical from one dendrite to another. That's why they're so much slower than a cpu.

    3. Re:Simulating neurons isn't enough by raftpeople · · Score: 1

      It's a good point. Scientists now understand that the synapse (new term is tripartite synapse) is managed and controlled by the glial cell surrounding the synapse. A network of neurons without the functionality of the glial cells managing the activity in the synapses and the flow of information will not really behave correctly.

  14. Re: Just a drop in the wetware bucket by religionofpeas · · Score: 1

    so far as we currently understand it

    Obviously, yes. So far as we currently understand it, the brain also doesn't exploit the magical properties of fairy dust.

  15. Re:$7/core by Barny · · Score: 4, Informative

    According to their website, they had custom silicon designed and built. A basic box with these things has 4 CPUs on it, and each CPU has 18 cores onboard, complete with their own high-speed memory for data and instructions.

    Check it out over here http://apt.cs.manchester.ac.uk...

    --
    ...
    /me sighs
  16. Re:$7/core by Barny · · Score: 1

    Oh, meant to add, that 72 core board uses just 5W of power.

    --
    ...
    /me sighs
  17. Well yeh by Anonymous Coward · · Score: 1

    Well yeh, I assumed they didn't just walk down the shop and buy some TV boxes!

    1 amp to run 64 cores.... literally a USB powered super computer in each board. FFS.

    "The 103 machine is the 48-node board and has 864 ARM processor cores, typically deployed as 768 application cores, 48 Monitor Processors and 48 spare cores. The 103 machine requires a 12V 6A supply. The control interface is two 100Mbps Ethernet connections, one for the Board Management Processor and the second for the SpiNNaker array. "

    Jealous.

  18. Re: "Supercomputer"... by Megol · · Score: 1

    Clusters using standard processors, standard memory, standard design with the addition of faster network cards. How is that wrong?

  19. Re:does not compute by q_e_t · · Score: 1

    It's research funding. It covers the salary of a number of people, plus the costs of equipment (a dozen racks of computer system), office space, pencils, travel, etc. There are 38 staff in the Advanced Processor Technology group (headed by Steve Furber) at Manchester University.

  20. Obligatory Skynet by mfnickster · · Score: 4, Funny

    "Skynet becomes self-aware at 02:14 am Eastern Time after its activation on Nov. 4, 2018 and immediately begins shitposting on 4chan."

    --
    "Slow down, Cowboy! It has been 3 years, 7 months and 26 days since you last successfully posted a comment."
  21. Re:$7/core by AmiMoJo · · Score: 1

    Interesting that they only have a couple of 100Mb/sec ethernet ports per 768 application processors. They must not be expecting to shift much data between cores.

    --
    const int one = 65536; (Silvermoon, Texture.cs)
    SJW, n: "Someone I don't like, and by the way I'm a fuckwit" - AC
  22. Re:Just a drop in the wetware bucket by lurcher · · Score: 1

    Plants are warm as well.
    http://www.bbc.co.uk/earth/story/20160715-organisms-might-be-quantum-machines

  23. Re:does not compute by Anonymous Coward · · Score: 1

    That's sort of the GPs point, there's no way this cost 20 million. It would have cost WAY more than that. Seriously, if there are 38 staff, 20 million would average out roughly 26K a year for the salaries alone, which seems rather low. And that wouldn't leave any money for....you know, the actual hardware.

  24. Re:"switched on for the first time" vs "has been u by RavenLrD20k · · Score: 1

    Because between the "yesterday" referenced in the article when they "turned it on for the first time" and the "now" when the article was written there wasn't enough time to use the machine in any form or function, right? It's just sitting there...idle...like a giant fucking paperweight and does absolutely fucking nothing. Or maybe there might actually be enough time in the 12 - 24 hrs that differentiates "yesterday" from "today" that they could have spun up some simulations that they had ready to go.

    BTW... It's the Article, not the editors, that specified the so called contradiction that really doesn't fucking exist you disingenuous fuck.

  25. Re:Just a drop in the wetware bucket by religionofpeas · · Score: 1

    Plants are warm as well.

    https://physicsworld.com/a/is-...

    So then, is photosynthesis “quantum” or not? “The observations show that there is correlation between the wavefunctions of the states involved in energy or electron transfer,” says Romero. “But these effects are not considered by some scientists as truly quantum coherence in the sense that entangled states of quantum computing are understood.” And Engel agrees that to compare the two is to invoke “the wrong language”.

  26. Note to editors/writers by Oligonicella · · Score: 1

    Understand the words you're trying to use:

    emulate VERB

    computing
    reproduce the function or action of (a different computer, software system, etc.).

    simulate VERB

    imitate the appearance or character of.
    produce a computer model of.

    They are not interchangeable.

  27. A neuron is more complex than a cpu by raftpeople · · Score: 2

    "A neuron is much simpler than a cpu" - I don't think you're up to speed on the complexity and adaptability of function within one single neuron.

    A neuron performs localized non-linear computations with spiking forward and backward throughout it's 10,000 dendrites, it's not a simple "sum". In addition, the long term state of synapses are maintained due to epigenetic changes in the dna, the neuron is managing all of those synaptic weights. Scientists don't fully understand the function of even one type of neuron (there are many many types).

    A single neuron is a very complex thing, probably similar to or more complex than a cpu.

    1. Re: A neuron is more complex than a cpu by ceoyoyo · · Score: 1

      The answer is, we donâ(TM)t know. The necessary functions of a neuron might be more complex than a cpu, or they might not be. Itâ(TM)s an interesting question what elements are fundamentally required.

      Arguing that anything less than a quantum simulation of every atom in every neuron is insufficient is just as silly as arguing that relu+madd is certainly all you need. As is everything in between.

      We need research in many different directions to identify what features of the brain and its components are actually fundamental and which are irrelevant.

    2. Re: A neuron is more complex than a cpu by ceoyoyo · · Score: 1

      I personally suspect that spiking is the easiest way nature could figure out how to make a biological system produce a variable electrical output, and that a continuous variable magnitude output will work just fine (like PWM versus variable voltage/current). Others are convinced that thereâ(TM)s something very important about spiking itself. Weâ(TM)ll see.

  28. But, this is wrong. by WindBourne · · Score: 1

    To truely emulate the brain, then memories have to be in each CPU. They have found that neurons now contain our memories, so a true setup would be like teradata DB, than a Von Neumann computer architecture.

    --
    I prefer the "u" in honour as it seems to be missing these days.
    1. Re:But, this is wrong. by religionofpeas · · Score: 1

      To truely emulate the brain, then memories have to be in each CPU.

      What makes you think the CPUs don't have local memory to store the various parts of state of the neuron ?

  29. Bingo by WindBourne · · Score: 1

    New info has the axon containing the memory. IOW, a single CPU should be accessing a portion of the memory, as opposed to all CPUs being able to access all memory ( though slowly ).

    --
    I prefer the "u" in honour as it seems to be missing these days.
  30. Minsky by raftpeople · · Score: 1

    Minsky pushed the idea that multi-layer neural networks were a dead end, missing the fact that a multi-layer neural network is a universal function approximator.

    Neural networks are suited for a broad range of problems that symbolic methods are not. Ultimately we will probably end up with a combination of neural + symbolic to achieve real AI. But to even harness the advanced methods that symbolic is well suited for requires that foundation of being able to deal with messy data, categorization and pattern matching first.

  31. Missing components by raftpeople · · Score: 1

    First, scientists need to understand what one single neuron does to be able to simulate. Current understanding is that a single neuron is actually more like an entire neural network due to all of the non-linear calcs performed by the 10,000 dendrites (not just sum, it's more complex with local/regional spiking, forward and backward, all kinds of stuff).

    Secondly, scientists will need to figure out what exactly glial cells are doing, they control the synapse and manage the action, detecting and releasing neurotransmitters, glialtransmitters, chemical and electrical gradiant controlling more regional sections, etc.

    1. Re:Missing components by jd · · Score: 1

      Add to that the discovery that neurons have differering genomes. They're modified via retrotransposons.

      https://www.ncbi.nlm.nih.gov/p...
      https://www.scientificamerican...
      http://epilepsygenetics.net/20...

      So we've a genetic algorithm inside each neuron in the neural network, on top of everything you mentioned.

      --
      It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
  32. "Emulate"? That's a joke. by Rick+Schumann · · Score: 1

    We have no idea how our brains really work, and this toy they built at best could only 'emulate' perhaps an insects' brain.

    1. Re:"Emulate"? That's a joke. by religionofpeas · · Score: 1

      We have no idea how our brains really work

      That's why they build machines like this. Trying out the parts that you think you understand is a good way to discover what's still missing. And I'm sure that the scientists who build the actual neural models understand the challenges better than you do.

      this toy they built at best could only 'emulate' perhaps an insects' brain.

      The article doesn't say how big it is right now, but the are planning to extend it to 1 billion neurons. That's bigger than a cat's brain.

  33. Human only if human brain is size of a mouse by 3seas · · Score: 1

    Heard this on NPR a few days ago by someone on the team. They said a mouse brain cause its close to a human brain.
    hey anyone know where I can get some really good cheese?

  34. Wow, just how unique is it? by Dr.+Spork · · Score: 1

    Is it really "incredibly unique"? Damn, that so much more unique than almost every other unique thing! I mean, most unique things are unique, but this thing is apparently incredibly more so.

  35. Re:does not compute by BlueMonk · · Score: 1

    They probably didn't have 38 staff for the whole time.

  36. Non-Locality of Mind! by Camel+Pilot · · Score: 1

    Mind != Brain. There have been dozens of experiments showing the Non-Locality of Mind..

    Citations please...

    1. Re: Non-Locality of Mind! by UnknownSoldier · · Score: 1

      Go read all of Karl Pribram's research/books or C.J.S. Clarke's Explaining Consciousness: The Hard Problem, and thanatology examples such as Ring and Valarino 1998; Sabom 1982 and 1998, etc.

      Also of interest will be Jonathan Shear's Explaining Consciousness: The Hard Problem, David Chalmers's Toward a Scientific Basis of Consciousness, and David Bohm's work.

  37. Re: does not compute by ceoyoyo · · Score: 1

    Someone is unfamiliar with how much they pay grad students....