Hacker News Reader

Comment by jjgreen | Hacker News Reader

Neural nets are a dead end. Fitting a new model to some problem advances our understanding of the problem not one jot, the best that you can do from there is stumble around hoping for a better fit.

View on HN

- they're not used for understanding a problem, they're used for getting accurate predictions (more accurate than alternative methods)

- they can be useful on their own without providing a path to something much more advanced

What if a problem is just an incomprehensibly complicated series of differential equations and a formal breakdown is simply impossible but a neural net adds value in parsing it?

Long terms we're almost certainly going to replace NNs with more robust first principles ideas for the things they solve, but short term you can't really argue that they're not solving things in computer vision that weren't possible to reach a few decades ago.

Hot take: No one knows what the future holds in AI. 15 years ago, people thought that neural nets were dead, but they've come back vindicated and with a vengeance.

> advances our understanding of the problem not one jot

Do they claim to do this?

I suspect the hangup is they are seen as part of artificial intelligence or machine learning, where intelligence and learning are words with existing baggage for many people.

You can’t get to the moon climbing progressively higher trees.

Ok but there's nothing on the moon

Classic HN taking analogies literally.

Reply on news.ycombinator.com