I don't know if you still have any impression of deepmind? If your memory is a little fuzzy, then here we can give you some tips. A few years ago, an artificial intelligence ai known as the alpha dog swept the world in the global Go circle, not only defeating Li Shishi, but also crying ke Jie, the first person in Go at that time. The alpha dog artificial intelligence is from deepmind.
It may be that it has achieved invincibility in Go, so deepmind quickly transferred this artificial intelligence system to other fields, such as e-sports games. As a result, in "StarCraft 2", alpha dogs once again swept the world, beheading major professional masters under the horse. Perhaps winning too much for the human brain, Deepmind feels that there is no challenge, after all, the speed of computing in the human brain is not as fast as AI, so this time Deepmind puts the field of AI into another more complex environment - nature.
What many people don't know is that after going through the lone battle of Go and esports, Deepmind has been working with the Met Office for the past few years, of course, this cooperation is carried out in secret. The collaboration between the two is the use of artificial intelligence to predict the weather, and deepmind recently published a paper in the journal Nature, showing the results of the collaboration between the two companies. From the introduction, deepmind has designed a new machine learning model that can predict whether it will rain in the next few hours with great accuracy than the weather station's forecast.
Current weather forecasting is driven by powerful numerical weather prediction systems that solve physical equations, which can get predictions of the size of the Earth several days in advance. However, they have a hard time producing high-resolution predictions within two hours. Real-time forecasting fills the performance gap for this critical time interval, and advances in meteorological sensing have allowed high-precision radars to provide measured ground precipitation data at high frequencies. However, real-time weather forecasting is actually improved from the numerical weather prediction system, which is also a deep learning model that can be randomly simulated to obtain predictable and deterministic forecasts, and is applicable on a variety of spatial scales.
However, existing deep learning models show uncertainty in the increasingly blurred precipitation, and do not include small-scale weather patterns. In addition, existing methods focus on site-specific predictions rather than probability predictions for the entire precipitation field, which makes it impossible to provide consistent predictions in multiple spatial and temporal sets at the same time, limiting usefulness. Deepmind has designed a new set of deep models that can both learn from observational data and represent uncertainties on multiple spatial and temporal scales.
Deepmind uses the method of generating models by making detailed and credible predictions of future radars based on past radars. With such an approach, it is possible to accurately capture large-scale events while also generating many alternative rainfall scenarios, reducing the uncertainty of rainfall forecasts. Deepmind's model is also very good at predicting moderate to heavy rain events, which is a significant improvement over other competing methods. Deepmind said that after cognitive mission assessments by more than 50 expert meteorologists at the National Weather Service, experts would rate deepmind's method as the first choice in 89% of cases compared to widely used forecasting methods.
In terms of actual effects, deepmind's scheme provides more accurate probability predictions than other methods, and retains the statistical characteristics of precipitation on spatial and temporal scales. However, many scientists have also questioned, some scientists said that they did not see any revolutionary advances in deepmind's weather forecasting, and its model is also based on existing technology, and there is no detailed content of the actual assessment.
Of course, no matter what, at least from the current conclusion, the deepmind artificial intelligence prediction system is superior to the traditional solution in terms of accuracy and short-term real-time weather forecast. But as for replacing the old scheme with this system, or even laying off weather forecasters, it is still too early.
On the one hand, Deepmind said there are currently no plans to put it into use immediately, but the team hopes to eventually be able to use the model to provide real information for future weather reports; on the other hand, Deepmind also acknowledges that artificial intelligence will not replace humans, and weather forecasting requires expert and human involvement to ensure that the understanding of prediction is reasonable and then communicated to the public.