快猫短视频

We could soon better forecast catastrophic events like disease outbreaks and floods thanks to work that allows artificial intelligence models to make predictions from scant data.

鈥淔orecasting extreme events with AI will no longer need 鈥榖ig data鈥�, unlocking countless opportunities to accurately predict disasters where data is limited, from hurricanes to earthquakes to pandemic spikes,鈥� says , who leads a team at Bayer Crop Science researching machine learning.

Standard computer prediction models require large data sets to produce accurate results, but there isn鈥檛 always sufficient information available for unusual events like rogue waves or disease outbreaks. So, Pickering and his colleagues turned to a machine-learning technique involving sequential sampling, in which a data point is analysed, then another is picked if needed.

This 鈥渁ctive learning鈥� enables AI algorithms to learn from available information to label new data points that are particularly relevant to a specific outcome.

The method is intended to identify the components of a complex system that have the biggest influence on whether a rare event occurs.

鈥淲e start with a small amount of initial data, the AI trains on this and, using its current understanding of the system, leverages our decision-making functions to ask the system for the next best piece of information to learn the system best,鈥� says Pickering.

He says the method is useful, for example, if you have no data points and have resources to perform only 10 experiments to understand your system.

鈥淭he question is, what are the 10 data points that will allow the AI to be the best predictor possible,鈥� he says. 鈥淥ur approach is able to find the 10 points, sequentially one by one, that are going to be best to describe the system and the intrinsic extremes in it.鈥�

Using the technique across a range of fields, the team found it could accurately pinpoint the most dangerous pandemic spikes and discover and predict rogue waves, reducing the number of samples needed to make a prediction from 10 million to 100聽compared with standard techniques.

AI forecasting, from predicting crime to predicting health outcomes, is usually based on collecting data on past events and applying AI algorithms, says , CEO of Pinecone, a machine-learning search infrastructure firm.

This new technique could help in areas where a lack of data currently limits prediction abilities, he says. 鈥淜nowing what economic or weather conditions preceded some extreme event might give us a good warning if we see those same conditions again in the present, yet the exact same conditions never quite repeat,鈥� says Liberty.

More-efficient AI predictions may also help forecast other climate-linked events. For instance, says Pickering, an El Ni帽o heating event, possibly exacerbated by climate change, is expected to cause more hurricanes this year.

However, accurate computer simulations of the phenomenon are expensive, says Pickering, and perhaps only a few hundred could be simulated in time when the hurricane season hits.

鈥淥ur approach, coupled with these simulations, would be able to use those hundreds of simulations to efficiently uncover the likely and catastrophic scenarios that society must prepare for in 2023,鈥� he says.

Even with improved AI techniques, forecasting will remain an inexact science. 鈥淭he further into the future we wish to understand, the more difficult learning the system becomes and the more data is required,鈥� says Pickering. 鈥淓vents far into the future would be intractable to understand with this method, but events near in the future are much better candidates.鈥�