Master Thesis Projects

All information for students, EPFL Professors and companies/universities

DH Master Project Information Final 2021 (pdf)

Form Master Thesis Assessment 2021 (pdf)

Form Master Thesis Assessment 2021 (doc)

Info Sheet For Expert Master Thesis Project 2021 (pdf)

Presentation for students

DH Master Thesis Presentation Final (PDF)

Form for Master thesis projects outside of EPFL

Form Master Thesis Project Outside EPFL 2021(pdf)

Form Master Thesis Project Outside EPFL 2021 (doc)

Available Master Thesis Projects in DH:

Project Title: muse

Professor: Sarah Kenderdine

Project summary:

muse is a four-year research project for 24 Swiss museums that turns qualitative experience into quantitative data, in real time. Through a digital interface, inside galleries, and in real time, museums gain actionable impulses for strategic and exhibition planning.

The research project aims at developing frameworks for sentiment analysis that can enrich the analysis of visitor experience. The scope of the master thesis project includes natural language processing and textual analysis as well as automating tasks for data analysis and visualization. Join the team of engineers, designers, evaluation experts and analysts. Some travel within Switzerland to these museums may be required.

Contact: Prof. Sarah Kenderdine

Project Title: Harmonic Analysis using Voice-Leading Reduction

Professor: Martin Rohrmeier

Project summary: Harmonic analysis of musical pieces works very differently when done automatically than when done manually by experts. Part of the reason is that human experts are able to correctly distinguish chord notes from ornamentation notes. The goal of this thesis is to develop a model for surface ornamentation based on voice-leading operations and to use it for harmonic analysis.

Contact: Christoph Finkensiep

Project Title: Evaluating variability in harmonic annotations of Western classical music

Professor: Martin Rohrmeier

Project summary: Harmonic analysis is a different task for humans as well as computers. In order to improve algorithmic approaches, the Digital and Cognitive Musicology lab is creating large datasets of expert labels for Western classical music (Bach, Mozart, Beethoven, …). Naturally, different annotators might choose different labels in the same situation. This master project aims at investigating the variability in the harmonic annotations which has not yet been done extensively for Western classical music.

Contact: Fabian Moss

Project Title: Neural Architecture Search with PULSE

Professor: Martin Rohrmeier

Project summary: Despite the huge success of deep learning in various machine learning tasks, designing the network architecture remains a time consuming task to be performed by experts. The need to automatise this task has recently led to advances in neural architecture search (e.g. [2]). A similar endeavour is undertaken for other machine learning models, such as Gaussian processes (e.g. [1]). The PULSE framework [3] was developed to efficiently optimise the structure of machine learning models (e.g. neural networks or Gaussian process models) in large discrete spaces and is suited to unify and improve the approaches that are currently being developed.

The topic of this master project is to reformulate and unify the approaches in [1] and/or [2] in the PULSE framework. Prior knowledge of the PULSE framework is not required but experience with the design of neural networks and/or Gaussian processes is advantageous.

References:

[1] Duvenaud D, Lloyd JR, Grosse R, et al (2013) Structure discovery in nonparametric regression through compositional kernel search. In: Proceedings of the 30th International Conference on International Conference on Machine Learning-Volume 28. pp III–1166

[2] Liu H, Simonyan K, Yang Y (2018) DARTS: Differentiable Architecture Search. arXiv:180609055 [cs, stat]

[3] Lieck R (2018) Learning Structured Models for Active Planning: Beyond the Markov Paradigm Towards Adaptable Abstractions. Phdthesis, Universität Stuttgart

Contact: Robert Lieck

Project Title: Investigating a Perceptual Reality behind Music Theoretical Structure

Professor: Martin Rohrmeier

Supervisor: Dr. Steffen A. Herff

Project summary: A central aspect of music theory is identifying and analyzing structures in music. Some of these structures can be described with syntactical models akin to those used in language. Whether or not these theoretical models allow prediction on a perceptual level, or are predominantly useful on an analytical level is the subject of the present project.
In this project, you will design, implement, conduct, and analyze an empirical experiment that aims to test some of the perceptual predictions of music theoretical models in human listeners. In most cases the tested predictions will be related to harmony, rhythm, long-distance relationships between musical events, or perception of closure.

Prerequisites: Fundamental knowledge in data science / statistical data analysis, skill in an appropriate programming language (e.g., R, Python, or MatLab). An interest in experimental psychology and music.

Contact: Dr. Steffen A. Herff

Project Title: Modeling the Note Level of Jazz Harmony

Professor: Martin Rohrmeier

Supervisor: Johannes Hentschel

Project summary: Harmonic analyses in the form of chord annotations represent a coarse-grained segmentation of the score with a given chord label expressing an abstraction of a particular segment. Such abstract chord labels therefore enable researchers to investigate differences and common patterns in different tonal languages.
The DCML has access to a small corpus of Jazz transcriptions with included chord labels (changes). The aim of this thesis would be to relate the abstract labels to the concrete notes that the performers played in the respective segments and to look for significant stylistic differences in various performers’ realization of particular chords.
The project is aimed at a highly motivated student interested in Jazz harmony and its realization in particular Jazz recordings. It requires music theoretical knowledge and very good skills in programming and data analysis. Prior experience with machine learning is an asset but not a requirement. Methodological choices are to be discussed.

Contact: Johannes Hentschel

Project Title: Music Beyond Major and Minor

Professor: Martin Rohrmeier

Supervisor: Dr. Steffen A. Herff

Project summary: Western classical music as well as Jazz, Rock and Pop is commonly thought of as being either in major or minor key. While this is a good approximation in many cases, there are far more nuanced characteristics in the tonality of a piece – but these are challenging to uncover and accurately define.
The goal of this project is to perform a large-scale exploration of this space using unsupervised machine learning methods. In particular, the goal is 1) to compile a large data set of pitch-class distributions from different sections of musical pieces 2) to implement a Dirichlet mixture model in PyTorch and train it on this data set and 3) to adjust the number of clusters via cross-validation by performing a bilevel optimisation using the hyper-gradient.

Prerequisites:

• good Python coding skills
• experience with object oriented design and unit tests
• familiarity with PyTorch (or automatic differentiation in general)
• good understanding of the math behind probabilistic models

Contact: Dr. Steffen Herff