[SIML] Learning Non-Stationary Independent Components from Time Series Data (Master or Bachelor Thesis)

Art der Arbeit:

Master Thesis

Betreuer:

Melodia, Luciano
Lehrstuhl für Informatik 6 (Datenmanagement)
E-Mail: luciano.melodia@fau.de

Lenz, Richard
Lehrstuhl für Informatik 6 (Datenmanagement)
Telefon +49.9131.85.27899, E-Mail: richard.lenz@fau.de

Beschreibung der Arbeit:

Finding independent components in functions is a challenging task in machine learning. Known as the cocktail party problem, it involves extracting the independent components from several signals from a mixed source. Only the number of components is known in an overcomplete scenario. In an overcomplete basis, the number of basis vectors is greater than the dimensionality of the input samples. The representation of an input is not a unique combination of basis vectors, however, overcomplete representations have greater robustness in the presence of noise, are more sparse, and have greater flexibility in matching structure in the data. Autoencoders, along with other forms of Independent Component Analysis (ICA), were used effectively for this purpose. Newer approaches for non-stationary data are to be tested.

Task: An implementation of time-contrastive learning and an evaluation of the independent signals. The signals are to be systematically mixed with each other and treated with noise. An analysis of the procedure as well as of the result should illuminate the different scenarios in natural data as isolated as possible. Mixes of EEG signals are to be separated as a data set.

Vorausgesetzte Vorlesungen bzw. Kenntnisse:

One or more lectures in the field of machine learning, deep learning, knowledge extraction or signal processing. Basic knowledge of linear algebra and analysis (not all of it should be forgotten). Good programming skills in Python or a similar language.

Bearbeitungszustand:

Die Arbeit ist bereits abgeschlossen.