Assumed prior knowledge
It is assumed that the student has good programming skills (see for example the content of the courses 'Programmeermethoden' and 'Algoritmiek') and good knowledge about Data Science and Machine Learning techniques (see for example the content of the courses 'Data Mining', 'Data Science' or 'Kunstmatige Intelligentie').
The fields of Data Science and Machine Learning deal with large volumes of data. Complex algorithms such as Stochastic Gradient Descent, Gradient Boosting and Support Vector Machines are able to model this data and make predictions about future trends. Most of these algorithms have a high number of hyperparameters, that need to be tuned correctly in order for the resulting model to perform good. Properly tuned hyperparameters can determine the difference between mediocre performance and state-of-the-art performance. When presented with a new dataset, common problems that need to be addressed are: Which algorithm to use and how to tune the hyperparameters to obtain good predictive performance. The research field of Automated Machine Learning (AutoML) focuses on how to automate this process.
There will be several lectures presented by the lecturer (approximately 6) in which core techniques from the field of AutoML are presented. The other slots are filled by student presentations and discussion. Students will work in teams of two or three, analyzing and understanding a seminal research paper and presenting it to the other students. Furthermore, the students will conduct a small research project, implementing the method of the paper they elected to present. The students will report on their findings in a small scientific report. At the end of the course, the student should understand:
Basic principles of state-of-the-art hyperparameter optimization techniques, including (but not limited to) Bayesian Optimization and Hyperband.
How to tune the hyperparameters of complex algorithms, such as Convolutional Neural Networks and Gradient Boosting.
Important aspects of Algorithm Selection and Hyperparameter Optimization: how to determine a good search space and what are important hyperparameters.
Meta-learning: transferring knowledge obtained from prior experiences to new datasets (e.g., MAML).
The most recent timetable can be found on the students' website.
Mode of instruction
Total hours of study: 168 hrs. (= 6 EC)
The time will be divided between:
Preparing the lectures - weekly: read the paper that will be presented - once: prepare to present a paper
Attending the lectures
Research project - execute project - report writing
The weighting of the final grade will be:
20% presentation (video animation or oral presentation, depending on the number of students)
80% written paper, project work and programming assignment(s)
Recent and seminal papers from the AutoML literature. To be announced during class.
You have to sign up for courses and exams (including retakes) in uSis. Check this link for information about how to register for courses.
Please also register for the course in Brightspace, as we intent to have all relevant communication go through this platform.
Lecturer: dr. J.N. van Rijn
More information will be available on the dedicated website for this course, hosted on BrightSpace.