If you are using bad feature engineering or you aren’t using cross validation then yeah, you’re probably going to overfit, but that’s not really a problem with neural networks. They’re not an all purpose tool that can solve all of humanity’s problems, but they definitely have their uses.
They’re not an all purpose tool that can solve all of humanity’s problems, but they definitely have their uses.
the problem is that humans are too stupid to understand this, and will try to use them for everything
I know there are ways to mitigate overfitting, but I figured explaining it wasn’t important for a comment made for the layman (I also have my doubts about the real world effectiveness of some of these techniques anyway, but that’s just based on my own experiences, so I didn’t mention it). Things have gotten better than when I first started playing with NNs, but overfitting is still an issue, especially since good, large datasets are hard to come by. The lack of generalization, whether through overfitting or as a natural consequence of NNs is a very large issue if there were to be used in a potentially life threatening scenario. But again, without the ability to know why a neural network has made given it’s answer they cannot be trusted with decision making.
