机器学习的原理及其在气候预测中的潜在应用

经历了两次“人工智能寒冬”之后,机器学习于近十年再次进入大众视野,且有腾飞发展之势,已在图像识别和语音识别系统等实际应用方面取得了巨大成功。从已知数据集中总结关键信息和主要特征,从而对新数据做出准确的识别和预测,分别是机器学习的主要任务和主要目标之一。从这个角度看,将机器学习整合到气候预测的思路切实可行。本文,首先以线性拟合参数(即斜率和截距)调整为例,介绍了机器学习通过梯度下降算法优化参数并最终得到线性拟合函数的过程。其次,本文介绍了神经网络的构建思路以及如何应用神经网络拟合非线性函数的过程。最后,阐述了深度学习之卷积神经网络的框架原理,并将卷积神经网络应用到东亚冬季逐月气温的回报试验,并与气候动力模式的回报结果相比较。本文将有助于理解机器学习的基本原理,为机器学习应用于气候预测提供一定的参考思路。

Machine learning and its potential application to climate prediction

After two "Artificial Intelligence winters",machine learning has become a subject of intense of media hype and come up in countless articles,showing a promising future.Machine learning has gained a big success in image recognition and speech recognition systems.Refining key message and dominant features from the train datasets and making accurate prediction on the never-seen-before datasets are the major task and the ultimate goal of machine learning,respectively.From this perspective,it''s feasible to integrate machine learning into climate prediction.Beginning with a simple example on finding the weights of a linear fitting,this study shows how machine learning updates weights through gradient descent algorithm and eventually obtains the linear fitting line.Next,this study illustrates the architecture of neural network and uses neural network algorithm to learn the true curve fitting a non-linear function.In the end,this study elaborates the architecture of deep learning such as convolutional neural network,and uses convolutional neural network model to hindcast winter monthly surface air temperature anomalies in East Asia.The results by deep learning are further compared with the hindcast by dynamical model-CanCM4i.This study will help to understand the fundamental of machine learning and provides insights how to integrate machine learning into climate prediction.