SoylentNews is people
SoylentNews
A research team at Google has developed a deep neural network that can make fast, detailed rainfall forecasts.
The researchers say their results are a dramatic improvement over previous techniques in two key ways. One is speed. Google says that leading weather forecasting models today take one to three hours to run, making them useless if you want a weather forecast an hour in the future. By contrast, Google says its system can produce results in less than 10 minutes—including the time to collect data from sensors around the United States.
[...] A second advantage: higher spatial resolution. Google's system breaks the United States down into squares 1km on a side. Google notes that in conventional systems, by contrast, "computational demands limit the spatial resolution to about 5 kilometers."
[...] Google says that its forecasts are more accurate than conventional weather forecasts, at least for time periods under six hours.
[...] Interestingly, Google's model is "physics-free": it isn't based on any a priori knowledge of atmospheric physics. The software doesn't try to simulate atmospheric variables like pressure, temperature, or humidity. Instead, it treats precipitation maps as images and tries to predict the next few images in the series based on previous snapshots.
[...] Specifically, it uses a popular neural network architecture called a U-Net that was first developed for diagnosing medical images.
[...] To produce a weather forecast, the network takes an hour's worth of previous precipitation maps as inputs. Each map is a "channel" in the input image, just as a conventional image has red, blue, and green channels. The network then tries to output a series of precipitation maps reflecting the precipitation over the next hour.
Like any neural network, this one is trained with past real-world examples. Thousands of past real-world weather patterns are fed into the network, and the training software tweaks the network's many parameters to more closely approximate the correct results for each training example. After repeating this process millions of times, the network gets pretty good at approximating future precipitation patterns for data it hasn't seen before.
(Score: 1, Touché) by Anonymous Coward on Wednesday January 15 2020, @10:39AM (5 children)
Yeah, who needs that... just throw shit at something and see if it makes better predictions, we don't need science for that.
(Score: 4, Insightful) by PiMuNu on Wednesday January 15 2020, @10:54AM (1 child)
Jargon is "surrogate model" - basically perform a multivariate fit to existing data, then extrapolate off the fit to do the forecast. Neural network seems quite a good technique for fitting to data in higher dimensional spaces.
Nowadays, it is called "AI" or somesuch, but fitting to data and then extrapolating from the fit is a well-known technique dating back many decades (centuries?).
(Score: 0) by Anonymous Coward on Wednesday January 15 2020, @12:29PM
agreed.
it's actually exciting that they got a much more efficient short-term prediction strategy, but I don't expect it to ever work properly for medium-term predictions.
or if it does work properly, it will be equivalent in cost to physics-based models.
it's exciting for two reasons:
1. cheaper and more accurate for short-term prediction.
2. perhaps by analyzing the resulting neural network someone can come up with a physical interpretation and extract an analytic expression out of it.
(Score: 0) by Anonymous Coward on Wednesday January 15 2020, @02:39PM
This is essentially how they used to forecast weather, just in a more sophisticated way. Weather moves around from where it starts and tracking minor changes can give a sense of where it's going in the short term, starting systems are a little more tricky, but they'll grow until they stop and the very beginning is subtle enough that people won't care much if that's a little off.
This is intended for short term predictions and is likely to be quite accurate. Over multiple days, it's not like to work well, but we'll see.
(Score: 2) by DannyB on Wednesday January 15 2020, @03:23PM (1 child)
<no-sarcasm>
Idea:
Train an AI based on a large number of coin tosses to guess the next coin toss outcome based on previous coin toss history!
Now apply this to roulette. Based on past outcomes of roulette spins, predict the next spin outcome.
What's wrong with this?
Isn't each event independent of prior events?
With weather, ignoring physics is simply predicting what weather will do next based on what it did recently? Or am I missing something?
</no-sarcasm>
To transfer files: right-click on file, pick Copy. Unplug mouse, plug mouse into other computer. Right-click, paste.
(Score: 0) by Anonymous Coward on Wednesday January 15 2020, @07:42PM
That's idiotic, the weather isn't random and over short periods of time, this approach is fine. Whereas what you lost id's purely random.
Thus is great for periods of time to short for doing a proper model. or can also allow more complicated models to be run by allowing more time,