SpiNNaker, the Supercomputer Which Models the Human Brain

    0

    Stephen Furber is a Professor of Computer Engineering at the University of Manchester. He is also the creator of SpiNNaker (Spiking Neural Network Architecture):  a supercomputer developed for brain modeling.

    Andrea Macdonald Founder of ideaXme interviews Stephen Furber. Furber talks about his life and work at Acorn, where he helped introduce computing into schools and also developed the first ARM- Acorn RISC Machine- Processor. He describes his interest in understanding the human brain and why it is important. Moreover, he tells us about the workings of SpiNNaker 1 and the function of its individual parts as well as the plans underway to create SpiNNaker 2.

    Watch the full interview here:

    Understanding the complexity of the human brain for advancement in the field of computing

     “Well because it is one of the great frontiers of science and when people talk about scientific frontiers they are usually thinking about the unimaginably small, such as subatomic particles, or the unimaginably large, such as the square kilometer array which is looking into the origins of the universe and the far reaches of space. But we all carry this thing which is very human-sized around inside our heads – if we take it out, we can hold it in our hands – and we don’t know how it works. And this seems to me to be a big gap in our knowledge and a gap that’s probably very important to our future because we know that, for example, diseases of the brain cost the developed economies more than cancer, heart disease and diabetes put together. So, medically it’s extremely important. And we lack treatments because we don’t understand it. So, we don’t know how to design drugs to interrupt the disease pathways. Also, of course, it impinges on my professional area of computing, and computing is increasingly moving towards artificial intelligence.”

    SpiNNaker – a brain-like supercomputer with a million processors

    “So, the SpiNNaker computer is a supercomputer in one sense that’s been developed for brain modeling. We started the project about 20 years ago and we considered what we might be able to contribute to the scientific quest to understand the brain as computer engineers. And we wondered what we could do if we built a machine with about a million mobile phone processors in it and got them all working together so that we could support large-scale brain models. We realized very early on that even with a million processors, we’re not approaching even 1 percent of the scale of the human brain. But we are at a scale where we can possibly model entire mouse brains which will be a significant step forward. We built the machine around several principles, but we had to customize its architecture to the problem of brain modeling because conventional computers really struggle to support large scale brain modelling in anything like real-time because the brain is hugely connected.

     Each neuron inside your head – and you have about a hundred million of them – connects to, on average, about 10,000 others. And this requires the communications inside the machine to send messages not from one place to another — which is typically what’s required in computers — but from one place to many thousands of other places. So, we built a bespoke communication infrastructure inside the machine to tune it to this problem.”

    The architecture of a SpiNNaker chip

    “Each SpiNNaker chip is about a square centimetre of silicon. That’s not an atypical size for today’s chips. It’s made on a fairly old silicon process technology. And so we can manufacture about a hundred million transistors on a chip and those are divided into eighteen processing regions. Each processing region has a relatively small and energy efficient ARM core, that is the processor I helped develop back Acorn in the 1980s, though of course it has evolved through the hands of many thousands of people in the meantime. We use ARM cores with memory and various devices, and you can see 18 physical copies of that on the chip if you know what you’re looking at. At the centre of the chip, there is a thing that implements this connectivity.”

     “Yes. It is the router in the middle of the chip. Biologically, neurons communicate principally by sending spikes. They go “ping” every so often. So, one thing to ponder is that all your thoughts are spatiotemporal patterns of pings flowing between the neurons in your head. A ping is in effect a pure asynchronous event.

     So, there’s no information in its size or shape. The information is purely in its timing. So, in SpiNNaker each ping becomes a tiny packet of information. All that packet of information says is that neuron number 320 just went ping. We communicate that packet around the machine from chip to chip, across the machine which occupies 10 data-center sized rack cabinets. And we deliver it in a small fraction of a millisecond to every destination to which it has to go. And that’s the real-time requirement.

     So, the key to SpiNNaker is this router in the middle of each chip. Routing is not new. It is the basis of the Internet; all the information that flows when you watch a video on your computer that’s coming from an Internet source is flowing in packets across the Internet. But typically, the requirements there are to get very high data rates using very big packets and the requirement in SpiNNaker is to achieve relatively modest data rates but using tiny packets because each packet is basically carrying one ping.”

    The architecture of a single SpiNNaker chip. Each chip has 18 core processors and a router in the middle, along with various devices. Courtesy: Radio ideaXme https://radioideaxme.com/wpcontent/uploads/2019/04/SpiNN2_labeled-1024×963.jpg

    Building SpiNNaker 2 with better design and features for efficient brain modeling

    “I should say of course that we haven’t abandoned SpiNNaker 1 because we’re still doing a lot of work in using it and we have a whole raft of collaborators through the European Union Human Brain Project — neuroscientists, computational neuroscientists scientists, who are keen to map their models onto SpiNNaker 1. So a lot of what we’re doing is supporting users on SpiNNaker1 and working with users. But alongside that, SpiNNaker 1 is relatively old technology. We’ve had the silicon since 2011. So, we’re developing a second-generation chip on a much more up to date semiconductor technology. And we’re developing this in collaboration with chip designers at TU Dresden, in Germany. And the goal is to hit something around 10x performance per chip. That will allow us to put something perhaps of the scale of the small insect brain into a single SpiNNaker 2 chip which would then control a drone or a small robot, only with obviously the kind of capabilities you’d find in an insect. But still, that’s quite a lot of neurons to control some autonomous device.

    We’ve also learnt a lot in the development and use of SpiNNaker 1 about what’s important and that’s allowed us to tune the design of SpiNNaker 2 even more to the problem of modelling biological neurons in real time than we were able to tune SpiNNaker 1 because at that time we didn’t know what we now know after 10 years. There has been more learning since then that we are building into SpiNNaker 2.”

    Combining SpiNNaker with other technologies

    “There already is some progress – and it is going to get more dramatic – in brain prostheses, so that it does not displaces the human brain but adds to it. This is already happening of course. One clear example is a cochlear implant which is a treatment for certain forms of deafness. This is a piece of electronics which basically interfaces directly with nerves that go into the brain and is very effective. On a similar vein, retinal plants are emerging. They are not as fully developed as cochlear implants but they are certainly going in the right direction to enable us to restore sight to people with certain forms of blindness – and it is not a cure for all forms of blindness but for certain forms where the problem is in the biological retina. It’s becoming increasingly possible to think about replacing the biological retina in the physical eyeball with a silicon retina that connects into the optic nerve to restore at the moment a fairly low-resolution form of vision, but that will improve over time.”

    The original interview has been condensed by Iqra Naveed and Umama Nasir.

    Credits: Interview by Andrea Macdonald founder ideaXme Ltd.

    With permission of ideaXme, a global podcast, ambassador and mentor programme. ideaXme interviews the creators and innovators who shape our world.  They speak to all those who Move the human story forward! ideaXme Ltd.

    Connect and find out more about ideaXme here:

    Twitter @ideaxm 
    Instagram 
    @ideaxme 
    Facebook
     ideaXme

    Read, listen to, or watch ideaXme interviews here:

    Super ‘Human Like’ Computer ideaXme website link

    Audio ideaXmeiTunes
    Text and audio 
    Radio ideaXme

    Find ideaXme across the internet including YouTube, iTunes, SoundCloud, Libsyn, Medium, www.radioideaxme.com, Spotify, Radio Public, TuneIn Radio and more.

    LEAVE A REPLY

    Please enter your comment!
    Please enter your name here

    This site uses Akismet to reduce spam. Learn how your comment data is processed.