Abstract

In recent years the exploitation of Machine Learning in many different domains has expanded considerably due to the increasing availability of large datasets and compute power. Supervised techniques (such as Deep Neural Networks) have made impressive progress in solving hard problems in image and speech recognition and are now gaining more popularity in the Weather and Climate domain. Unsupervised learning has received less attention but may be a source of data-mining tools to deal with future high volume, diversity and dimensionality of data produced from both observations and models. This talk will give an overview of applying an unsupervised technique (the Self-Organising Map or SOM ) to investigate relationships between cloud and cloud-controlling variables in high-dimensional observations and climate model data. SOMs are an effective dimension-reduction and clustering technique suitable for handling non-linear relationships in data. In this work we aim to explore whether SOMs are better able to identify non-linear relationships within the data than standard techniques, and whether this can provide a useful tool for assessing how well such relationships are represented in climate models. Our findings so far indicate that the SOM seems to emphasise relationships between variables that are shown as much weaker or non-existent with only standard correlation analysis and points to some interesting avenues to investigate further in model development.