The Swiss engineering master's degree

Machine learning (ML) emerged out of artificial intelligence and computer science as the academic discipline concerned with “giving computers the ability to learn without being explicitly programmed” (A. Samuel, 1959). Today, it is the methodological driver behind the mega-trend of digitalization. ML experts are highly sought after in industry and academia alike.

This course builds upon basic knowledge in math, programming and analytics/statistics as is typically gained in respective undergraduate courses of diverse engineering disciplines. From there, it teaches the foundations of modern machine learning techniques in a way that focuses on practical applicability to real-world problems. The complete process of building a learning system is considered:

• formulating the task at hand as a learning problem;
• extracting useful features from the available data;
• choosing and parameterizing a suitable learning algorithm.

Covered topics include cross-cutting concerns like ML system design and debugging (how to get intuition into learned models and results) as well as feature engineering; covered algorithms include (amongst others) Support Vector Machines (SVM) and ensemble methods.

Compétences préalables

• Math: basic calculus / linear algebra / probability calculus (e.g., derivatives, matrix multiplication, normal distribution)
• Statistics: basic descriptive statistics (e.g., mean, variance, co-variance, histograms, box plots)
• Programming: good command of any structured programming language (e.g., Python, Matlab, R, Java, C, C++)
• Analytics: basic data analysis methods (data pre-processing, linear & logistic regression)

Objectifs d'apprentissage

• Students know the background and taxonomy of machine learning methods
• On this basis, they formulate given problems as learning tasks and select a proper learning method
• Students are able to convert a data set into a proper feature set fitting for a task at hand
• They evaluate the chosen approach in a structured way using proper design of experiment
• Students know how to select models, and „debug“ features and learning algorithms if results do not fit expectations
• Students are able to leverage on the evaluation framework to tune the parameters of a given system and optimize its performances
• Students have seen examples of different data sources / problem types and are able to acquire additional expert knowledge from the scientific literature

Contenu des modules

• Introduction (ca. 2 weeks): Convergence for participants with different backgrounds
• Supervised learning (ca. 7 weeks): Learn from labeled data
  Cross-cutting topics: such as feature engineering; ensemble learning; instance vs. model-based approaches, debugging ML systems
  Algorithms: e.g. kNN, decision tree, SVM, ensemble learning (bagging, boosting), graphical models (Bayesian networks), gradient based approaches
  bias-variance tradeoff: hyperparameter tuning, cross-validation, performance metrics
• Unsupervised learning (ca. 3 weeks): Learning without labels
  Algorithms: e.g., clustering, dimensionality reduction, anomaly detection, archetypal analysis
• Special chapters (ca. 2 weeks):
  Algorithms: e.g., reinforcement learning, recommender systems, hidden Markov / Gaussian mixture models

Méthodes d'enseignement et d'apprentissage

Classroom teaching; programming exercises (e.g., in Python 3,  Jupyter notebooks, Orange)

Bibliographie

• Aurélien Géron: "Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow", Concepts, Tools, and Techniques to Build Intelligent Systems, O’Reilley Media, 2022
• T. Mitchell, “Machine Learning”, 1997
• C. M. Bishop, “Pattern Recognition and Machine Learning”, 2006
• Simon Rogers, Mark Girolami: “A First Course in Machine Learning”, ISBN-13: 978-0367574642, Chapman and Hall/CRC; 2. Edition, 2016
• G. James et al., “An Introduction to Statistical Learning”, 2014