A machine learning course oriented on climate science and energy sector applications and based on Python notebooks.
The amount of data produced in environmental sciences (energy and climate prediction) paves the way to new applications. While traditionnal statistical methods are still essential, advanced machine-learning methods are more and more needed to make sense of such big data, whether it is for analysis or to make predictions. One goal of machine learning is to extract identifiable patterns from these complex data sets. These patterns can then be used to take informed decisions. Another is to model relationships between different variables and then use these models to predict one variable from information of the other. Examples of such big data sets are:
- If we want to optimize the energy performance of a building, we can set sensors in different places of the building that will give us a good overview of the energy consumption and energy loss. Analyzing such data set with machine learning will help us predict or optimize our energy consumption.
- The IPCC provides forecast of the average temperature for the next 100 years. This forecast is based on about 30 different predictions made by complex models. Machine learning provides a way to track reliable patterns in this complex data set.
The objective of this course is to provide an introduction to statistical analysis and machine learning in order to help the students apply relevant methods to analyze specific data sets. Different families of unsupervised and supervised methods will be covered, while also providing a general approach to validate and test results. We encourage students to develop their critical thinking skills when facing a new dataset and applying a method in order to draw robust conclusions. We illustrate this course with examples from environmental sciences. These datasets correspond to concrete cases of analysis presented in the form of ipython notebook.
- Introduction,
- Supervised Learning Problem and Least Squares,
- Overfitting/Underfitting and Bias/Variance Decomposition,
- Regularization, Model Selection and Evaluation,
- Classification I: Generative models,
- Classification II: Discriminative models,
- Introduction to Unsupervised Learning with a Focus on Principal Component Analysis,
- Ensemble Methods for Regression,
- Introduction to Neural Networks,
- Neural Network Regularization and Introduction to Deep Learning.
After each course session of about 45 min, the students work on a tutorial for about 45 min during which they code the applications on real data, analyse the results and discuss them. The professors assist the students during the tutorials and make sure that all students are able to achieve the objectives of the tutorials.
Students by groups of two choose projects based on applications to environmental problems. The objective of the projects is to apply the statistical learning methodology from processing the intput data to evaluating the prediction skills of a model. The focus is on mastering the overall approach rather than on applying the most sophisticated models.
Technical support: All examples, tutorials and projects are based on the scikit-learn Python package. The students are thus asked to code in Python, but they do not need to be skilled in scikit-learn to join the course (see Prerequisites below). Students work either on a JupyterHub (priority), a school computer or on their personal laptop.
- Elementary data analysis in Python with numpy, pandas and matplotlib,
- Linear algebra (linear systems, inverse, eigenvalues and eigenvectors),
- Elements of probabilities and statistics (probability distribution and density, random variable, conditional expectation, variance, covariance, sample estimates).