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

    •  
      CommentAuthorAngus
    • CommentTimeJan 7th 2022
     
    Hmmm...have only scanned it so far. But I can criticise it on the basis of sufficient ignorance.

    At first blush it seems like a very hard way to make a neural network. I don't see why they say it is fast. The thing will be dominated by the dynamics of the erbium gain sections which depends on pump, signal, etc. And how do you create many many synapses when you have to tease out individual cores to stick in the signals.

    Now I will have to see if they have addressed these cavils.
  1.  
    The world will know by mid-year. It's at least clear from a bunch of weasel words that they don't know how to train the bloody thing.
    •  
      CommentAuthorAngus
    • CommentTimeJan 7th 2022
     
    A question for the learned in these matters. A synapse has to do a weighted sum of the inputs, which is a linear process, and also produce a decision, such as passing a threshold, which is a nonlinear process Shirley?

    This thing only seems to have the first bit unless the nonlinear part is in the electronics after detection. Which would certainly cancel the advantages they claim for the optics.
  2.  
    The Perceptron was a linear discriminator and as such could chop the data with hyperplanes. Minsky & Pappert wrote a popular book showing that such purely linear systems were unable to solve certain problems, famously the XOR one.

    I couldn't tell on initial scan whether this thing was linear or not.
    •  
      CommentAuthorAngus
    • CommentTimeJan 7th 2022
     
    The optical part is.