AI is promising to transform weather forecasting

I always thought that the idea of AI was that it fed itself with the information it needs to give a correct result. Rather like the human brain but perhaps quicker at sorting through information from all over the place.

All machine learning does is find patterns (during learning) and predict the outcome (during inference). For AI (or “deep learning”), we need a lot more data because there is much less preconception of the patterns to find, the neural network finds the patterns by itself.

An example of “classical” machine learning is the Haar cascade face detection algorithm. We have a set of 160,000 theoretical black / white patterns on a 24×24 image, and the ML algorithm ranks them based on their ability to determine if a 24×24 grayscale image is a human face or not. After the learning phase, it allows to quickly rule out an image just based on the most important patterns.

Deep learning would not define so many preconceptions about the patterns, the image’s size and framing, etc., and thus require a lot more input images to re-create the intermediate patterns (it’s called “unsupervised” machine learning). That’s all the difference there is. The goal and the data are still provided to the model (it’s not Sci-Fi AI in that regard, meaning it can’t self-determine goals or be aware of the limit of its knowledge).

It should indeed be a good fit for predicting weather, but there’s also the limit based on the chaotic system. We cannot hope to have much more reliable predictions a lot more in advance.

As for chaotic systems (which is different from qualifying complexity, although related), I don’t think studying them will allow for predictions with certainty, I think it is theoretically impossible. The goal is more to predict how likely it is that some events will happen, or to obtain statistics of how the system behaves on average. For example, you could determine how likely it is that a planet gets ejected from a 3 body system after a certain time (and that’s certainly something we’ve not solved, apart in simulations). You will not be able to predict with certainty when the planet will be ejected (except just before it happens). AI could also help here, but we still won’t be able to call it a “solved problem” because we’ll only have simulation results about the mathematical soundness or accuracy of the model. An example of how this is useful (despite no certainty in the results) is climate models: from the basic chaotic biosphere system, determine averages in the distant future (which is what climate studies are all about).

And I would think the meteorologists at ECMWF understand complexity better that most of us.

Not at all IMO. It’s all about gut feeling as of today. It’s much like the stock market. Choosing randomly is usually better than following so called expert analysts. (There are strategies you can follow, buts that’s another concept entirely).

Weather is more predictable than the stock market though, but as mentioned above, garbage in is garbage out. There is no reason to believe in increased accuracy and detail without adding measurement points to feed the calculations.

It’s rather weird how strong the faith is in black boxes. The less you know about how it works, the stronger the faith

Protein folding:

I know about that. It can just as well be seen as a complete failure of neural networks as a win. It is better than any human, but it doesn’t even scratch the core of the problem.

The thing is, there are about 10 to the 300 different ways to fold a typical protein. Only one is correct, only one particular way will make the protein work inside an organism, and nature does it in ms. How is a complete mystery. If anything in nature is magic, than this is it. This is the true magic of life. It’s so unimaginable that believing in God seems like the better choice. An AI can do 90%. We are still left with 10 to the 299 possibilities.

What you say about the stock market has been confirmed by studies several times.

For sure. By analysts tons and tons. Interesting you left out my comment about strategies

much like the stock market. Choosing randomly is usually better than following so called expert analysts.

Like many one-liners, the above is wrong. What is true is that a tracker beats some 90% of experts, but the reason is obvious (the FTSE100 – or actually any index – is self-selecting for well run companies). And similarly buying up say ten FTSE100 stocks, stuffing the certificates in your drawer, leaving them there for 10 years, also beats most analysts. There are other factors which confuse the debate e.g. even a monkey can make money in a bull market…

Getting back to wx, the biggest problem for us pilots has been that, in Europe, the data has since for ever been a revenue source for the government (taxpayer funded!) wx agencies, and only recently has this typically European exploitation been cracked with ECMWF data (usually in windy.com).

If ECMWF use pattern recognition, neural learning, whatever, that is all good, but is it “AI”?

Peter wrote:

but is it “AI”?

I would say by popular consensus, yes. Of course anyone is free to come up with their own definition of anything. But the term AI is bandied about right now specifically to cover neural networks. I can tell you that when you are trying to raise money from Silicon-valley style investors, you had BETTER have a good story on why your product is applicable to AI – even if it is breath freshener or sock suspenders. (Next year’s fashion will of course be something else).