• ayaya@lemdro.id
    link
    fedilink
    English
    arrow-up
    5
    ·
    6 months ago

    What you’re asking for is literally impossible.

    A neural network is basically nothing more than a set of weights. If one word makes a weight go up by 0.0001 and then another word makes it go down by 0.0001, and you do that billions of times for billions of weights, how do you determine what in the data created those weights? Every single thing that’s in the training data had some kind of effect on everything else.

    It’s like combining billions of buckets of water together in a pool and then taking out 1 cup from that and trying to figure out which buckets contributed to that cup. It doesn’t make any sense.

    • EnderMB@lemmy.world
      link
      fedilink
      arrow-up
      3
      ·
      6 months ago

      Respectfully, I worked for Alexa AI on compositional ML, and we were largely able to do exactly this with customer utterances, so to say it is impossible is simply not true. Many companies have to have some degree of ability to remove troublesome data, and while tracing data inside a model is rather difficult (historically it would be done during the building of datasets or measured at evaluation time) it’s definitely something that most big tech companies will do.

      • ayaya@lemdro.id
        link
        fedilink
        English
        arrow-up
        2
        ·
        6 months ago

        Sorry, I misinterpreted what you meant. You said “any AI models” so I thought you were talking about the model itself should somehow know where the data came from. Obviously the companies training the models can catalog their data sources.

        But besides that, if you work on AI you should know better than anyone that removing training data is counter to the goal of fixing overfitting. You need more data to make the model more generalized. All you’d be doing is making it more likely to reproduce existing material because it has less to work off of. That’s worse for everyone.