This is a bit over-hyped and not exactly a breakthrough. It’s not doing true weather prediction but rather extrapolating the movement of radar images. This is nothing new. I remember as far back as the 1990s TV weathermen would draw a line across a line of moving rain and the computer would predict which town it would get to and when.
Slapping “AI” on this 25 years later is an good example of the whole present PR move of labeling things as “AI” that are just rather basic data analytics.
No. Drawing the line on the map was just used to decide where you wanted to show the prediction for. The forward looking prediction models themselves have been around for decades.
It is, but in many cases its still pretty easy to figure out. It depends on how complicated the problem is. If it's truly an incomprehensible nonlinear cesspool, then its pretty much impossible. If it's secretly a linear regression with a few forks and inflection points, its easy.
The problems neural nets are used for do tend to fall into the former, but the techniques to unravel it are getting better.
What exactly is "true weather prediction" then? Using a crystal ball?
And yes, extrapolation is not new, but deep learning is just much better at it.
> labeling things as “AI” that are just rather basic data analytics
They are using a GAN to do this, you can't get more AI than this. If this is not AI, then what is?
Getting accurate short term predictions is very valuable for e.g. prediction of solar panels energy production to avoid starting coal peaker plants, or to predict flooding.
I get that it's popular to dunk on anything called AI to show that you have not fallen prey to the PR bullshit machine, but in that case, it's 100% unwarranted in my opinion.
the point is that when academics and private entities start throwing around the word "AI" for anything digital they are cashing in on other people's hard work and diluting the positive effects of actual AI research.
I generally agree with your points, but nowcasting is definitely the less exciting problem for ML to solve. Also, applying a GAN to nowcasting is definitely not a novel idea.
You'd hope that given that modern ML methods can handle crazy high dimensional data that someone would be able to do something with weather forecasting beyond current models. However, there are many factors that render this a 'hard' problem.
I guess, starting off with nowcasting is a neat way for DeepMind to become acquainted with the topic matter, but I'd hope they have their ambitions set a little higher.
Sure, that might not be a Nobel prize worthy breakthrough, but it is still very much AI, and it is still very much "true" weather prediction. I just don't understand where the gatekeeping on both point is coming from.
Some folks around Hacker News are convinced that AI is just another hype-cycle.
Yes, this paper is an incremental improvement over the state of the art of hand crafted algorithms. However, it should be viewed as an incremental step towards greater things!
I'm sure DeepMind, the team that recently (basically) solved the protein folding problem, will deliver exciting advancements in the field of weather forecasting in the coming years.
The AI has been another hype-cycle for at least two times already. "AI winter" is a term from the seventies, and it re-emerged the second time in the late eighties. But maybe the third time the charm, who knows.
I think it's fair to say that we've moved forward leaps and bounds since the the 80s. There have been major advances in compute and AI methods since that era.
There's certainly much further to go and we may end up hitting another wall sooner than later, but AI is delivering tangible results today in ways that it never could in previous eras.
This same argument was just as valid in 1970 (AI has never been done before at all) and in 1985 (they did actually improve it, and the hardware, drastically). Still, in both cases walls were hit and hype died. So, learning from this dataset, I would predict something similar.
In any case, the weather on a scale larger than 2 months is effectively can't be modelled because butterfly effect is real: there is just not enough raw data being collected, and it's not precise enough; and IIRC the required amount of data and floating-point precision for more or less accurate modelling grows exponentially wrt the time period being modelled.
For instance, every modern smartphone today has AI inference accelerator in it. Not even the largest supercomputer in 1985 (or even 1995) had anything close to the AI capability of even a phone today. Some are even beaten out by watches.
Also, there were no practical applications of AI in 1985. There was a lot of speculation, but nothing came out of it because the technology just wasn't there yet.
Today computer vision, voice recognition, voice synthesis, driving assist mechanisms, face recognition, and many other fields have been advanced considerably through the use of AI technologies.
We have yet to even scratch the surface of what is possible with our current level of development.
Yes, we'll hit a wall soon enough, but we'll hit that wall having developed a range of useful products. The previous iterations were considered hype because nothing came out of them, but that can not be said for the current environment.
It's like saying there is a hype-cycle in cataract treatment because the development of method plateaued between babylonia and the age of reason.
“True weather prediction” as in the actual computer forecast modes that have been around for decades. Taking atmospheric measurements and creating 3D predictions of weather conditions in the future, including wind speed, precip type and amount, and so on. The example here is just basic image extrapolation analytics, which as pointed out by many, is not new.
I too can predict if it will rain in two hours' time (somewhat reliably) by looking out of the window.
On a more serious note, Yandex has been using ML in his Meteum platform since about 2016, I think? and this summer they've incorporated the user reports into it too: one can open the Weather or Maps widget and click "It's actually raining here right now" or "Actually, it's not raining here right now" button. Apparently that helps to improve the accuracy of predictions when combined with other data.
No idea how good their algorithm is though, I don't generally look at the weather widget on my phone anyway since looking out of the window is faster and works well enough.
I have also re-implemented this feature in All Clear on Android [1]. My main lesson from all this so far is that it is useful and fun to collect and show the data, and real insights can be had by users seeing their peers' data, but incorporating the data into a model usefully is a bit more difficult! Making numerical predictions that actively use the user's weather data (rather than just showing 'x users say it is raining nearby even though the forecast says it shouldn't be') is much tougher.
I too can predict if it will rain in two hours' time (somewhat reliably) by looking out of the window.
Years ago, I set up a simple website that screen-scraped the BBC's weather predictions, and compared them against the day's weather report to calculate a very crude and basic accuracy.
For the UK towns it monitors, a dumb prediction of "tomorrow's weather will be the same as today's" gives a 34% accuracy - which only falls to about 25% when predicting the weather for next week! Luckily, the proper weather forecasters do a bit better than this :)
Have you tried comparing the weather on a date with the weather on the same date but one year earlier? I heard an anecdote that it's actually one of the most reliable ways to predict weather but I kinda doubt it... and I am too lazy to actually check whether it is.
There isn't enough info in the article to know exactly what is happening, but the problem with most of these systems (my first attempt 10 years ago, or Dark Sky, others as well) is that they use existing rainfall and wind speed to make guesses about where the rain will move. They typically fail at making real weather predictions; new rain starting to fall is a simple example that is often missed by these kinds of predictions.
The article just says "maps" but doesn't say what the data is. Rainfall amount? Just rainfall or also wind maps? What other maps?
Curious to see how this works in reality. Most of these attempts can be impressive sometimes, but fail often enough to make it not reliable for real use. However I am extremely optimistic about the future of increasing weather prediction accuracy, including with AI models as well as traditional physics models.
Splendid. From now on for anything in my study involving correlation and causation to define future events, I will stamp AI on every sentence of the paper.
In the Netherlands there is buienradar.nl which also uses radar images to predict rain fall accurately. Already for more then 10 years. Before an algorithm was called ‘AI’ ;-)
Did you read the paper? It makes clear that, yes this has been done for years and is nothing new as such, but this new approach gives much more accurate and higher resolution results.
I only read the article, sorry! Its just that at the time I was often amazed by the accuracy of the service (for timing, duration and rain intensity). But I guess there is always room for improvement.
37 comments
[ 2.9 ms ] story [ 80.0 ms ] threadSlapping “AI” on this 25 years later is an good example of the whole present PR move of labeling things as “AI” that are just rather basic data analytics.
The chasm between drawing a line on a map and this system is enormous, and represents 30 years of research.
It doesn't have to be a breakthrough to be in the news, it's a routine advancement of the state-of-the art.
And it might be much less of a black box than you imagine.
The problems neural nets are used for do tend to fall into the former, but the techniques to unravel it are getting better.
What exactly is "true weather prediction" then? Using a crystal ball?
And yes, extrapolation is not new, but deep learning is just much better at it.
> labeling things as “AI” that are just rather basic data analytics
They are using a GAN to do this, you can't get more AI than this. If this is not AI, then what is?
Getting accurate short term predictions is very valuable for e.g. prediction of solar panels energy production to avoid starting coal peaker plants, or to predict flooding.
I get that it's popular to dunk on anything called AI to show that you have not fallen prey to the PR bullshit machine, but in that case, it's 100% unwarranted in my opinion.
You'd hope that given that modern ML methods can handle crazy high dimensional data that someone would be able to do something with weather forecasting beyond current models. However, there are many factors that render this a 'hard' problem.
I guess, starting off with nowcasting is a neat way for DeepMind to become acquainted with the topic matter, but I'd hope they have their ambitions set a little higher.
Here's an interesting paper on this topic:
https://royalsocietypublishing.org/doi/10.1098/rsta.2020.009...
Yes, this paper is an incremental improvement over the state of the art of hand crafted algorithms. However, it should be viewed as an incremental step towards greater things!
I'm sure DeepMind, the team that recently (basically) solved the protein folding problem, will deliver exciting advancements in the field of weather forecasting in the coming years.
There's certainly much further to go and we may end up hitting another wall sooner than later, but AI is delivering tangible results today in ways that it never could in previous eras.
In any case, the weather on a scale larger than 2 months is effectively can't be modelled because butterfly effect is real: there is just not enough raw data being collected, and it's not precise enough; and IIRC the required amount of data and floating-point precision for more or less accurate modelling grows exponentially wrt the time period being modelled.
For instance, every modern smartphone today has AI inference accelerator in it. Not even the largest supercomputer in 1985 (or even 1995) had anything close to the AI capability of even a phone today. Some are even beaten out by watches.
Also, there were no practical applications of AI in 1985. There was a lot of speculation, but nothing came out of it because the technology just wasn't there yet.
Today computer vision, voice recognition, voice synthesis, driving assist mechanisms, face recognition, and many other fields have been advanced considerably through the use of AI technologies.
We have yet to even scratch the surface of what is possible with our current level of development.
Yes, we'll hit a wall soon enough, but we'll hit that wall having developed a range of useful products. The previous iterations were considered hype because nothing came out of them, but that can not be said for the current environment.
It's like saying there is a hype-cycle in cataract treatment because the development of method plateaued between babylonia and the age of reason.
On a more serious note, Yandex has been using ML in his Meteum platform since about 2016, I think? and this summer they've incorporated the user reports into it too: one can open the Weather or Maps widget and click "It's actually raining here right now" or "Actually, it's not raining here right now" button. Apparently that helps to improve the accuracy of predictions when combined with other data.
No idea how good their algorithm is though, I don't generally look at the weather widget on my phone anyway since looking out of the window is faster and works well enough.
I have also re-implemented this feature in All Clear on Android [1]. My main lesson from all this so far is that it is useful and fun to collect and show the data, and real insights can be had by users seeing their peers' data, but incorporating the data into a model usefully is a bit more difficult! Making numerical predictions that actively use the user's weather data (rather than just showing 'x users say it is raining nearby even though the forecast says it shouldn't be') is much tougher.
[1] https://play.google.com/store/apps/details?id=com.allclearwe...
Years ago, I set up a simple website that screen-scraped the BBC's weather predictions, and compared them against the day's weather report to calculate a very crude and basic accuracy.
For the UK towns it monitors, a dumb prediction of "tomorrow's weather will be the same as today's" gives a 34% accuracy - which only falls to about 25% when predicting the weather for next week! Luckily, the proper weather forecasters do a bit better than this :)
https://weather.slimyhorror.com/
(Excuse the basic site, I set this up over 17 years ago, and with minimal tweaks it has been left to its own devices since then)
The article just says "maps" but doesn't say what the data is. Rainfall amount? Just rainfall or also wind maps? What other maps?
Curious to see how this works in reality. Most of these attempts can be impressive sometimes, but fail often enough to make it not reliable for real use. However I am extremely optimistic about the future of increasing weather prediction accuracy, including with AI models as well as traditional physics models.
We'll see I guess.
The paper is linked in the article and answers all your data questions, including links to where you can download the data yourself.
So when is some investor going to come and cover my grandma with benjamins?
https://darksky.net/app
https://news.ycombinator.com/item?id=28744467