Hacker News Reader

Comment by smoldesu | Hacker News Reader

Honestly the hardest part about integrating AI into our daily lives is the fact that we don't really have AI yet. We have made great advances in the fields of machine learning and neural networks, but actually getting a computer to make educated decisions is hard. The current issue is that all of these models are black boxes: an ML model can guess which data will come out, but it can't really fully understand what it knows. It can identify slightly harder to notice patterns, but it can't actually think.

We adopted the phrase "AI" far too early in the field, and I suspect there will be another few decades before we have the technical and scientific capability to make real artifical intelligence a thing.

View on HN

> The current issue is that all of these models are black boxes

Not quite. It'd be great if that was the only issue with neural network ML. A much bigger issue is that neural networks have extremely limited applicability. They're great for classification problems when you can have huge training datasets, but for many common problems they're useless.

One obvious class of problems is time series prediction - extremely important in life and for business, and something neural networks are no good for.

There are many ML time series prediction techniques.

And as a sibling points out, pretraining (not to mention various other low data methods) make big NN models useable even on small datasets.

> There are many ML time series prediction techniques.

Yes, and they all give worse results than classic math and statistics approaches.

This, in my opinion, is the worst part of ML of all. Instead of clearly stating the limits and boundary conditions of the algorithms you start hearing responses to the tune of "you're doing it wrong" and "wait until X magic pixie dust makes it usable", like you're being sold snake oil instead of algorithms research.

Doesn't transfer learning level the playing field a bit? We already have excellent, off-the-shelf models for NLP, computer vision, ODT, etc... that just need fine-tuning for a particular business or domain problem. I think the 'huge training datasets' requirement is lessening every day.

If you're doing something novel, then sure. I can see that.

Predicting time series isn't "novel". It's the oldest data science problem in the world. And ML can't solve it yet.

Then go and use the right tool for the right job. I'm not suggesting neural networks are a panacea. I've worked with time-series problems recently and not used NN. NN not being good at one class of problems does not invalidate their usefulness for other classes, like your GP comment is suggesting.

> ...like your GP comment is suggesting

Really? What I literally wrote is "neural networks have extremely limited applicability".

> I'm not suggesting neural networks are a panacea.

That's good. Because ML practitioners are, even if they phrase it differently. (The idea is that since neural networks can fit curves then any problem can be solved by a neural network given enough layers and feature engineering grease.)

"Approximate Imitation"?

Reply on news.ycombinator.com