Vanilla 1.1.9 is a product of Lussumo. More Information: Documentation, Community Support.

  1.  
    First, create your scientist
    https://www.quantamagazine.org/machine-scientists-distill-the-laws-of-physics-from-raw-data-20220510/

    "My God" I exclaimed as I entered the laboratory, "Your tits are looking absolutely splendid today, Doctor Tannenbaum!"
  2.  
    It's not a neuron - it's a Heath Robins-on
    https://www.quantamagazine.org/how-to-make-the-universe-think-for-us-20220531/

    But now that I've read more, I have a sense of a great discovery.
  3.  
    Dillavou's "smart resistor" network is brilliantly simple.
    https://arxiv.org/abs/2108.00275
    I've been waiting since 1986 for a biologically feasible alternative to backpropagation to come along. I had thought Blake Richards had it with his pyramidal cell sims, but it appears not to have caught on for some reason. This new idea is just like using a wind tunnel as a computer. Let the natural physics do the talking. Decentralise everything. I'm already wondering about engineering optimisations - for example, using memristors so as to collapse the two networks into a single one.

    I should shut up for now and finish reading!

    ETA The refs show that they've already considered memristors. Late again!

    ETA2 I reckon the horse to back is the optical one anyway. Still, a hobbyist could still lash something together with smart resistors.
    • CommentAuthorthehard
    • CommentTimeJun 3rd 2022
     
    But can you make it with a billion neurons?
  4.  
    Why not?
    • CommentAuthorthehard
    • CommentTimeJun 4th 2022
     
    They say it's scalable, but it's also pretty large. My best guess is that miniaturization will result in the same heat problems than a normal SoC.
  5.  
    The smart resistor research has benefited in the interim from a couple of subsequent papers that seem to cement the view that this could be a useful technology. Not only can resistance updates be asynchronous but also update rates may be made a lot faster than the natural relaxation rate of the network.
    https://arxiv.org/abs/2201.04626
    https://arxiv.org/abs/2112.11399

    Here's a brief list of the advantages over backpropagation:
    1. It's purely local in both space and time. Each synapse (resistor) updates itself with recourse to only its local state, and thus there's no need for a global bus carrying information and control around the place. This drastically simplifies any chip layout.
    2. It removes the need for the massive amount of intermediate storage which backprop requires (GB of it). Memory is expensive in respect of both real estate and power.
    3. Its locality frees the design from any sort of architectural constraint, since the resistor mesh can be any topology one chooses.
    4. The design appears to be highly robust against both noise and damage. Try damaging a few GPU transistors and see what happens.
    5. Power dissipation can be very low. GPUs - not so much.
    6. A chip implementation promises to be extraordinarily dense. We are currently approaching one trillion transistors per chip and a resistor is comparable to a transistor in both size and complexity.

    One Achilles Heel remains - speed.

    Let's see what V2 has to offer.
  6.  
    Posted By: thehardThey say it's scalable, but it's also pretty large. My best guess is that miniaturization will result in the same heat problems than a normal SoC.
    Yabut you can't compare it to digital circuitry ackling at GHz. This is analog tech (mainly).
    • CommentAuthorkorkskrew
    • CommentTimeJun 4th 2022 edited
     
    Those who take comfort in the idea that at least "creatives" are safe from being replaced by AI's, behold and despair.
    https://www.youtube.com/watch?v=lbUluHiqwoA
    Reality is irrelevant - or at least it will be.
    ETA: Humans too, probably.
    • CommentAuthorAsterix
    • CommentTimeJun 4th 2022 edited
     
    That "two minute paper" lasted 9:50.
  7.  
    Yeah and he talks funny too!
    •  
      CommentAuthorAngus
    • CommentTimeJun 5th 2022
     
    It lacks insight.
    • CommentAuthorAsterix
    • CommentTimeJun 14th 2022
     
    •  
      CommentAuthorAngus
    • CommentTimeJun 14th 2022
     
    So - well before Homo sapiens.
    • CommentAuthorBigOilRep
    • CommentTimeJun 14th 2022
     
    Posted By: AngusSo - well before Homo sapiens.

    There's plenty of evidence of fire use by homo erectus - I think it's when you get past 1m years that it gets more controversial.
  8.  
    Posted By: AngusSo - well before Homo sapiens.
    I assume, perhaps wrongly, that Homo and human are not synonymous.
    •  
      CommentAuthorAngus
    • CommentTimeJun 14th 2022
     
    No, but until BoR corrected me I had thought the wisdom was that fire use started with us.
    • CommentAuthorBigOilRep
    • CommentTimeJun 14th 2022
     
    Posted By: AngusNo, but until BoR corrected me I had thought the wisdom was that fire use started with us.

    I think the arguments over early sites are always whether they utilised an already burning fire, or had the facility to actually start one.
    •  
      CommentAuthorAngus
    • CommentTimeJun 14th 2022
     
    Yes, the same thought had occurred about the AI discovery.
    • CommentAuthorAsterix
    • CommentTimeJun 18th 2022