Auto-encoder training sessions (Day 1 13:30-15:00 & Day 2 13:30-15:00)
Facilitated by: Tennessee Leeuwenburg (Bureau of Meteorology)
This split-session tutorial provides an on-ramp to understanding and training machine learning models, through the example of autoencoders. Autoencoders are neural networks that learn compact representations of complex data and are widely used for dimensionality reduction, feature extraction, anomaly detection, and generative modelling. In Earth system science, they can be applied to large climate and weather datasets to identify patterns, compress high-dimensional data, and support a range of downstream machine learning tasks.
The session is aimed at science students, early-career scientists, or later-career scientists who want a quick orientation into the world of machine learning. Data scientists may also find it interesting from the perspective of the application to Earth system science.
The first session will be a presentation covering neural network architectures and machine learning training processes, to explain the terms and concepts involved. The PyEarthTools framework will be introduced, and some attendees can make an early start on using the worked examples.
The second session will give attendees the chance to train their own models, either by working through the examples included in a series of Jupyter notebooks, or by connecting their own data sources into pre-defined network architectures in similar fashion. A basic knowledge of the NCI ARE environment is recommended.
Attendees can attend both sessions, or choose the one most of interest.