A New Way To Predict Ship-Killing Rogue Waves (economist.com) 46
AI models can find patterns and make predictions, but their reasoning is often inscrutable. This "black box" issue makes AI less reliable and less scientifically useful. However, a team led by Dion Hafner (a computer scientist at the University of Copenhagen) devised a clever neural network to predict rogue waves. By restricting inputs to meaningful wave measurements and tracing how they flowed through the network, the team extracted a simple five-part equation encapsulating the AI's logic. Economist adds: To generate a human-comprehensible equation, the researchers used a method inspired by natural selection in biology. They told a separate algorithm to come up with a slew of different equations using those five variables, with the aim of matching the neural network's output as closely as possible. The best equations were mixed and combined, and the process was repeated. The result, eventually, was an equation that was simple and almost as accurate as the neural network. Both predicted rogue waves better than existing models.
The first part of the equation rediscovered a bit of existing theory: it is an approximation of a well-known equation in wave dynamics. Other parts included some terms that the researchers suspected might be involved in rogue-wave formation but are not in standard models. There were some puzzlers, too: the final bit of the equation includes a term that is inversely proportional to how spread out the energy of the waves is. Current human theories include a second variable that the machine did not replicate. One explanation is that the network was not trained on a wide enough selection of examples. Another is that the machine is right, and the second variable is not actually necessary.
The first part of the equation rediscovered a bit of existing theory: it is an approximation of a well-known equation in wave dynamics. Other parts included some terms that the researchers suspected might be involved in rogue-wave formation but are not in standard models. There were some puzzlers, too: the final bit of the equation includes a term that is inversely proportional to how spread out the energy of the waves is. Current human theories include a second variable that the machine did not replicate. One explanation is that the network was not trained on a wide enough selection of examples. Another is that the machine is right, and the second variable is not actually necessary.
Re: Hahaha (Score:2)
How natural were the laws in Flatland?
Re: (Score:2)
Practically all Flatland laws were copied from some sort of natural phenomenon. Sharpness inflicting wounds, fog getting thicker with distance, etc. But your point was?
Re: Hahaha (Score:2)
Is a sphere natural? Did a Flatland natural law explain its movement through the Flatland world correctly?
Re: (Score:2)
Are you an idiot? Yes, it appears you are.
Re: Hahaha (Score:2)
Huh? How do you think we determined things like the gravitational force? Do you think we just spitballed random formulae and checked if it fit?
Re: Hahaha (Score:2)
How come galaxies don't all have the same gravitational force?
Re: (Score:2)
You could read a book and you would not need to ask this question. You could even read the original book ( https://www.gutenberg.org/eboo... [gutenberg.org] ), or its encyclopedia entry ( https://plato.stanford.edu/ent... [stanford.edu] ) instead of shitposting, but here we are.
Re: Hahaha (Score:2)
You don't have a response, then? Cause you just posted two irrelevant links...
Re: (Score:2)
I posted the full story of how the theory of gravity was developed. A very interesting read, really, but a long one, so I understand why you'd not be interested.
Re: Hahaha (Score:2)
One link is about symbolic logic published in the 20th century. The other is Newton's classic that is largely concerned with classical mechanics and calculus. There are entire chapters dedicated to philosophy. I can only assume you have NOT read it, since you can't provide any further citation. I have no interest in trying to figure out what you think you read that you misunderstood.
Re: Hahaha (Score:2)
Here's the proof you're wrong: https://www.wikipedia.org/ [wikipedia.org]
Re: (Score:2)
This is only proof that you're dumber than a bag of rocks.
Re: (Score:2)
Some of my best friends happen to be rocks that I keep in bags. and in boxes, and sometimes out in the garden.
Re: (Score:2)
Please ask them to accept my apologies for comparing them to the GP. It was indeed rude and uncalled for.
Wait wut? (Score:4, Insightful)
The first paragraph claims that they devised some new way to trace exactly what the neural network is doing inside the black box. Which would be very cool.
The second paragraph explains that they fed wave data into a neural network pattern recognition algorithm, and then fit a 5-variable function to the output. Which is ABSOLUTELY NOT the same. They then observed that the result closely matches a known wave equation (booorrinng) but with an unexplained extra term (hm. Could be interesting) which is likely an artifact of the training data they used (oh. Boring again).
Worth publishing, but meh. Maybe some ML/CS or physicist can school me as to why this is cool. To me, this is like asking a ML algorithm to analyze movies of objects falling, and the algorithm spits out a quadratic equation. Not exactly earth-shattering research?
Re: (Score:2)
Re: (Score:2)
Re:Wait wut? (Score:5, Informative)
Now, since I have the knowledge to understand this sort of thing, here's my take:
1) The researchers decided upfront what the relevant variables can be. As they say in the article:
So this takes care of the "science" part.
2) The researchers then tried to build a GLM, which is a well understood linear regression model where you create all the combinations of variables you can think of with weights for each combination. Then you estimate the weights, and throw out terms where the weights are too small. What's left is a sparse linear model which you pair with a link function.
3) Whereas in classical statistics the combinations of variables are simple analytic functions of the input variables, here they allowed a fully connected NN to find the best fitting function for each such combination. So they are doing functional data analysis with a NN as the approximating family. Alternatively, you could call this approach a nonparametric fit of the interaction term.
4) In the paper they do illustrate their method with allusions to a graphical belief network, but that is not really needed, it's just nice to read for people who aren't familiar with statistical modelling.
5) For their final output they don't want to keep the NN's for the variable cliques that remain, probably because NNs are extremely wasteful and brittle representations, so they define a family of common analytic expressions and look for the closest match to each NN, then replace the NN with that. This yields clean, continuous and expressive interaction terms which can be published and verified by future researchers (unlike NNs).
6) Does this approach "science"? My feeling is that it is a step in the right direction of using the NNs as plumbing in an existing tried and true approach. Does the GLM + final replacement of the NNs by analytical expressions bring real understanding? If it remains as is then no, but it may inspire future (human) theoreticians to find physical explanations for those particular combinations of variables.
Re: Wait wut? (Score:2)
By just how much were the inscrutable NNs better rogue wave predictors?
Re: (Score:2)
Re: (Score:3)
>Why wouldn't larger ships be less vulnerable to extreme waves than smaller ships ? :
They both are, and aren't. They can ride out large waves better than smaller vessels because of their weight: it gives them the ability to plow through where smaller vessels would have to ride the wave up. But larger vessels can't handle the stresses of larger waves, because of their own weight and length. Here's a quote about rogue waves from somebody who studies this stuff - Dr. Libe Washburn
"The waves are pretty dange
Re: (Score:3)
Re: (Score:1)
Because rogue waves are fucking big.
The highest one on records are far more than 100m/100yard high.
And usually when a ship gets hit: it vanishes without a trace. The wreck to be found decades later at an odd place.
Simply imagine: you have a 300m long container ship. Floating on the surface or the ocean. A second later it is under a 100m high avalanche of water. There is no gentle slop to climb the wave and ride it. It is not a tsunami, where you outside on the ocean most of the time notice nothing at all.
Re: (Score:2)
Occasionally, the end of the ship with the engines can limp back into por
Re: (Score:1)
That is not what "rogue waves" is about.
A "rogue wave", or "freak wave" is a monster wave that is created by overlapping waves going into different directions where several high amplitudes match at a point.
They can be absurd steep and 100m high or more.
But your explanation for "heavy weather wavers" is correct.
Re: (Score:2)
There certainly was a "reporting issue" until well into the 1960s, in that weather above a certain level generally killed all observers, which reduced the reported peak wave reports. As fixed installation well away from coast effects have increased, that reporting issue has decayed. Unfortunately, such installations also come with (increasingly) automated wave monitoring systems which have greatly improved the state of knowledge of actual sea states,
Re: (Score:1)
Not sure if I get your point.
I have seen lab videos where they experiment with artificial created rogue waves in an "aquarium".
What I mean with "rogue waves" is an event when waves going in different directions. Created by different storms in recent days. The amplitudes of such waves can "randomly looking" overlay and create very steep waves of 30m to 100m height.
If such a wave comes from the side, it is like tilting a desk from horizontal to 75degree or even 90 degree. The ship will simply fall down into t
Re: (Score:2)
Depending on the "fetch", up to weeks.
To get to those numbers from lab experiments, you need to do some very careful scaling (of viscosity, drag of wind on wave) from the lab observations to get towards field relevant predictions. But as you say, they are the consequence of previous storms interacting with the present weather. Which means your ship's master
Actual paper (Score:2)
instead of the article that was written by a moron - https://arxiv.org/abs/2311.125... [arxiv.org]
Re: Actual paper-Abstract (Score:1)
Abstract
Big data and large-scale machine learning have had a profound impact on science and engineering, particularly in fields focused on forecasting and prediction. Yet, it is still not clear how we can use the superior pattern-matching abilities of machine learning models for scientific discovery. This is because the goals of machine learning and science are generally not aligned. In addition to being accurate, scientific theories must also be causally consistent with the underlying physical process and
That's just an excuse (Score:2)
"AI models can find patterns and make predictions, but their reasoning is often inscrutable. This "black box" issue makes AI less reliable and less scientifically useful."
I posit that the #1 reason AI predictions are not useful is mostly because of the "convincing bullshitter" problem* that all LLM-driven AIs seem to exhibit. If that wasn't an issue, the lack of insight into their reasoning would not be nearly as problematic.
* Let's face it, "convincing bullshitter" is a more accurate description than the A
Re: That's just an excuse (Score:2)
Project, much?
Re: That's just an excuse (Score:4)
A "hallucination" in human terms is a temporary aberration, something mostly harmless that cannot be explained but will soon go away. That is the connotation which the AI industry wants you to believe, that those mistakes and alternative facts in their gigantic models are just temporary. Throw a few people at the problem and it will go away, nothing to worry about. And for those who ask for more details: ok just build other AIs whose job it will be to fix the flaws in the previous AIs. They'll figure it out. It's AIs all the way down.
Re: That's just an excuse (Score:2)
"the term "hallucinations" is a deliberate attempt to control the narrative and minimize the commercial impact of the blatant untruths being generated on a daily basis by such models and the people who promote them as solutions to everything."
Did you just describe mainstream economics?
Staying in the dark using a different technique (Score:2)
I predict that there will be no rogue waves... (Score:2)
Laugh all you want, but I got paid six figures by the government for a rogue wave study to produce that insight.
AI for weather patterns (Score:3)
Neural Nets are Equations, Too (Score:3, Interesting)
Too weak (Score:1)