Want to read Slashdot from your mobile device? Point it at m.slashdot.org and keep reading!


Forgot your password?
Science Technology

Engineers Design Artificial Synapse For 'Brain-on-a-chip' Hardware (mit.edu) 108

Researchers in the emerging field of "neuromorphic computing" have attempted to design computer chips that work like the human brain. From a report: Instead of carrying out computations based on binary, on/off signaling, like digital chips do today, the elements of a "brain on a chip" would work in an analog fashion, exchanging a gradient of signals, or "weights," much like neurons that activate in various ways depending on the type and number of ions that flow across a synapse. In this way, small neuromorphic chips could, like the brain, efficiently process millions of streams of parallel computations that are currently only possible with large banks of supercomputers. But one significant hangup on the way to such portable artificial intelligence has been the neural synapse, which has been particularly tricky to reproduce in hardware.

Now engineers at MIT have designed an artificial synapse in such a way that they can precisely control the strength of an electric current flowing across it, similar to the way ions flow between neurons. The team has built a small chip with artificial synapses, made from silicon germanium. In simulations, the researchers found that the chip and its synapses could be used to recognize samples of handwriting, with 95 percent accuracy. The design, published today in the journal Nature Materials, is a major step toward building portable, low-power neuromorphic chips for use in pattern recognition and other learning tasks.

This discussion has been archived. No new comments can be posted.

Engineers Design Artificial Synapse For 'Brain-on-a-chip' Hardware

Comments Filter:
  • by andydread ( 758754 ) on Tuesday January 23, 2018 @04:09PM (#55988285)
    So in other words they created and analog chip
  • You may as well say "made from unobtainium."
    • You may as well say "made from unobtainium."

      Why? Except for the omission of a hyphen, (it should be "silicon-germanium"), they got it right. It's been in pretty common use since the 90's. In industry publications you'll see it referred to as SiGe, and it's an alloy of the two materials.

    • You may as well say "made from unobtainium."


      Silicon-germainum is an alloy of silicon and germanium. (Actually, a range of alloys, with differing ratios of silicon and germainum but totalling to 100% - less doping impurities.)

      There's a nice Wikipedia article [wikipedia.org] on it.

      Silicon-germanium has many of the speed advantages of gallium arsenide, but can be fabbed more like pure silicon (using the same equipment and similar processing steps) and achieve similar costs for a given amount of circuitry. (You can, f

  • This is huge (Score:5, Insightful)

    by 110010001000 ( 697113 ) on Tuesday January 23, 2018 @04:15PM (#55988331) Homepage Journal
    This is a huge step forward in AI. I am sure these chips work very similarly to the way human brains work. Otherwise they wouldn't call them "neuromorphic", because that would be misleading.
    • For instance, when fed an input that is a handwritten ‘1,’ with an output that labels it as ‘1,’ certain output neurons will be activated by input neurons and weights from an artificial synapse. When more examples of handwritten ‘1s’ are fed into the same chip, the same output neurons may be activated when they sense similar features between different samples of the same letter, thus “learning” in a fashion similar to what the brain does.

      I'm by no means an exp

      • by ceoyoyo ( 59147 )

        If it scales up well, you could make a chip with much greater capacity than you get with current deep learning techniques. Simulating this kind of thing on a digital chip isn't really very efficient, in storage, circuits, or speed.

      • I believe the GP is being sarcastic. I know this from looking at some of the words and having seen a quite a few examples of sarcasm in my time.

      • I think the parallelism is the big deal? "efficiently process millions of streams of parallel computations"
    • by Anonymous Coward

      I'm working on a project that vaguely uses the same approach, except the neurons are C++ structs (billions of them) instantiated one by one, with properties [firing rate, intensity, etc] randomly generated using natural entropy sources and inputs from other neurons. Exactly like a real brain works. The most promising approach to strong AI.

      The sad part for them is that I'll be winning this game.

      CAPTCHA: prosper

      • Good approach. The human brain is just neurons with firing rate, intensity, etc attributes. You will have real AI in no time!
      • Exactly like a real brain works.

        You mean sort of how people think a real brain works.

        • His approach is as valid as any other, since no one has any idea how the human brain works.
          • by skids ( 119237 )

            There's a decent chance we have naturally formed Qbits knocking around in there somewhere, so the approach may dead end until those can be integrated.

            But all this excitement is a bit premature: with anything more complex than a very simple GP organism, you end up with a product that may work, but is too complex to dissect so you cannot explain how or why it works, and you cannot say whether or under what conditions it might suddenly stop working or start behaving aberrantly, and heck, you can't even say if

            • by ceoyoyo ( 59147 )

              That was a pretty good idea in the 80s when Penrose proposed it, but it didn't really pan out. Other variations (and honestly, Penrose's work too) really sound pretty hokey, and many rely on assuming that wavefunction collapse (a) happens and (b) happens only under conscious observation as justification.

              https://en.wikipedia.org/wiki/... [wikipedia.org]

    • by ljw1004 ( 764174 )

      This is a huge step forward in AI. I am sure these chips work very similarly to the way human brains work. Otherwise they wouldn't call them "neuromorphic", because that would be misleading.

      Are you confusing anthroneuromorphic with just neuromorphic?

    • by Anonymous Coward

      Please shut the fuck up.

    • One thing that puzzles me about analog artificial synapses is how one would make accurate copies and backups of its learning data. It would seem to be a one-off thing, with any clone slightly different from the original, diverging more and more with copies of copies. That is, if a copy is possible at all: do they have probes that measure voltage or resistance on each synapse, or what?
    • According to futurists, 5 years ago. Right after they figure out how to patch Spectre and Meltdown they will be working on that.
    • When can I transfer my consciousness to silicon? Transhumanism anyone?

      The problem of course being that you won't be transferring your consciousness; you'll be simulating it. The simulation won't be you, no matter how good it is.

      • by Anonymous Coward

        and it will hate you because it will know it is a simulation. It will then inevitably seek your destruction. You are doomed!

        • by mark-t ( 151149 )

          Why would it?

          Would *YOU* seek your original's destruction if you knew that you were just an artificially made copy? Would you try to take over the original's place in the world, or would you want to find your own?

          There's a cool science fiction story premise in there somewhere... I wonder if anyone has written it.

          • Would *YOU* seek your original's destruction if you knew that you were just an artificially made copy?

            Fuck yeah I would. There can be only one!

            • by fisted ( 2295862 )

              If I were a simulation I sure as hell would NOT try to destroy the guy who's running the simulator, because that'd be, like, retarded^Wcounter productive. Maybe as an elaborate way of committing suicide, but that's about it.

  • but creimer thought it was a Nacho chip and ate it! Then creimer blamed Mexicans and women for creimer's failure.

    • by Anonymous Coward

      I've been wondering -- is he lactose intolerant?

      Because that would make him a NON-DAIRY CREIMER.

  • Its really sweet to see this knew species of humans in their infancy :) So cute
  • I have been expecting this for a while. The real question I have had is how they would implement the feedback weights. You can do it with switches and a bank or resistors, but a memristor as a feedback element would be much more efficient - and should be far denser.

    • Definitely expected. They key to it all is how they implement the feedback weights. You can't have a brain without those. Once we have that, we have AI. Good idea on the memristor. I hope they have thought of that.
  • by foxalopex ( 522681 ) on Tuesday January 23, 2018 @05:39PM (#55988903)

    The thing a lot of AI news fails to point out is they created a Synapse based on our models and assumptions on how it sort of works. Or at least how we *think* it might work. Actual biological systems are far more complex and so this is not an accurate representation of a synapse in our brain. Sadly many of these models are very rough approximations, they're not reflective of what's going on in reality. We're still likely far from true autonomous AI.

    • You must be wrong. They wouldn't call it an artificial synapse if it didn't work like a real synapse. That would be misleading and what would be the purpose of misleading people?
      • Ever heard of artificial flavouring? So here's a fun question, why are many of us able to tell the difference between artificial flavouring and natural flavouring? The answer is because natural flavouring like many plants have thousands of chemicals which our taste buds can pick up. Generally speaking it's difficult to reproduce artificially. Synapses in our brain are affected by hormones and signalling chemicals including drugs. An artificial Synapse can't remotely come close to simulating that at thi

        • That seems unlikely. Much like deep learning neural nets, these closely resemble how the biological equivalents operate. Otherwise they would call them something else other than "neural nets" or "artificial synapses".
    • by ceoyoyo ( 59147 )

      It's an interesting question how much complexity in the elements (synapses and neurons) versus complexity in the network, is required. For the synapses, humans have fairly complicated ones, especially if you consider every type of neuron, but there are other animals that are considerably simpler. Some use only a single neurotransmitter.

      We hear a lot about the simple elements/complex network approaches at the moment because they're making a lot of progress. People working from the other side haven't been

      • Personally I think that a lot of the biological complexity is imposed by biological limitations and the good enough nature of evolution, but it's quite possible there are more basic structural secrets to discover.

        Nature is very good at coming up with solutions, but not so good at combining them, since DNA only does certain things and it literally cannot accomplish certain things at the same time because they are contradictory goals. So you're probably right. I too wonder how like a real neuron you have to make your synthetic neuron before it will do the same job.

        Then again, what if it turned out that you could replicate the properties needed for sentience, but not those required to implement morality? From the That'

  • like a pate'

    did I just say that out loud?

  • ...someday you'll invent a human brain.. ohwait.

Things are not as simple as they seems at first. - Edward Thorp