back to article Who knows the secret of the black magic box? Boffins seek the secrets of AI learning by mapping digital neurons

OpenAI Microscope: Neural networks, often described as “black boxes”, are complicated; it’s difficult to understand how all the neurons in the different layers interact with one another. As a result, machine learning engineers have a hard time trying to interpret their models. OpenAI Microscope, a new project launched this …

  1. Mage Silver badge
    Facepalm

    which groups of neurons are activated in a model when it processes an image

    They are NOT neurons in any biological sense. Can we have a proper name for this and stop pretending it's like how a brain works and stop using misleading terms like Learning (it's storage of curated data) and AI (It's artificial, but not intelligent or stupid).

    For a brief moment I wondered was it some sort of a scanner examining actual brains.

    Though the tool sounds interesting.

    1. Anonymous Coward
      Anonymous Coward

      Re: which groups of neurons are activated in a model when it processes an image

      Agree totally, 30 years ago they were ANNs artificial neural networks and they were trained (standard terminology for models) as opposed to learning (which implies they achieve this themselves, like a T2).

      ANNs are trained to model an approximation of some function and as you say is artificial but not intelligent...

  2. Pascal Monett Silver badge

    "see which groups of neurons are activated"

    One thing I don't get here : these "neurons" are a programming object, right ? In that case, why did nobody think of including an activity log for each neuron ?

    What is keeping that from happening ?

    If they included that, there would be no need to guess, the log would be clear. Analyzing the log would be a tedious affair, to be sure, but it would be clear.

    1. Anonymous Coward
      Anonymous Coward

      Re: "see which groups of neurons are activated"

      Good point, the neurons are always active but depending on the weights on the links may or may not contribute much to the decision output, so you do need a tool to light up which nodes are active in the context of the overall network activity.

      This is interesting but I'm not sure how much benefit to the world this is, how the network ends up being structured often relates to the initial random weighting assigned to the links.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like

Biting the hand that feeds IT © 1998–2022