This class is taught sporadically in either the summer or winter term. The class is taught in English.

Information for the current semester (if available):

General

In this course, students participate in the DCASE challenge (https://dcase.community/challenge2024/), opens an external URL in a new window. Participants can choose among the tasks outlined on the website. The tasks involve challenging real-world problems that researchers worldwide are tackling. Participating in this course and working on one of the tasks is a challenging venture and coming up with a deep-learning-based solution for a non-trivial real-world problem (as opposed to the assignments you typically tackle in exercise courses) is a meaningful experience.

Motivation

This course is for those who want to enhance their deep learning skills by working on a challenging audio-related deep learning problem. If you are looking for easy ECTS, this course is probably not a good choice. Achieving a good rank in one of the challenges, or even winning one of them, would be a remarkable achievement; however, competing against other (experienced) researchers is not easy, so your final grade will not depend on the rank of your submission.

Students of previous semesters competed successfully in a variety of tasks and came up with competitive and unique solutions. Here are some (of many) highlights:

- Rank 4 in DCASE'23, Task 1, opens an external URL in a new window
- Rank 7 in DCASE'23, Task 6, opens an external URL in a new window
- Rank 3 in DCASE'22, Task 1, opens an external URL in a new window

Prerequisites

First and foremost, you should be motivated to think deeply about solutions to challenging problems. On the practical side, this course requires prior experience with machine learning and familiarity with deep learning frameworks such as Tensorflow and PyTorch. Programming experience in Python is, therefore, also a non-negotiable requirement. Furthermore, code will be versioned and shared via GitLab; hence, familiarity with Git and GitLab is a big plus. We will introduce the basics of audio-signal processing and suggest a setup for a deep-learning pipeline in the first sessions.

Course Outline

We will form one or more teams depending on the number of participants. Each team will work on one of the ten tasks, for which a short description is already outlined on the DCASE website, opens an external URL in a new window. Each group will implement a complete data processing pipeline, from raw data and annotations to final predictions and evaluation. Ideally, this pipeline will include methods to facilitate the tuning of hyperparameters and model selection. We will, of course, assist you during this whole process with suggestions and helpful tips on methods and general approaches to such projects. We will provide datasets if we can or help find and distribute them, and we will also be able to offer a limited amount of (shared!) computing resources.

A bonus: the course ends with the challenge, which is on June 15th; you have more time to focus on your exams afterward.

Mode

We will have weekly meetings where you will give an update on your progress so far. We will set up a Zoom meeting for each session so that participants can also join remotely. Presenters will be selected from each group in a round-robin fashion. We may also review the code and ask individual group members clarification questions. Each group will need to produce a final written report detailing their approaches and lessons learned. The challenges will be done in teams, and the final grade is not dependent on the rank you achieve in the challenge (although it will be a bonus if you achieve a top rank). You will be graded individually. The grade is based on attendance, presentations, questions answered, and contributions to the code and final report.