Department of Computational Perception
Department of
Computational Perception
Johannes Kepler Universität Linz


Home  –  Mission  –  Teaching  –  People  –  Research  –  Media  –  Awards  –  Impressum


MACHINE LEARNING

<< back to overview


Neural Networks

In recent years, neural networks outperformed traditional algorithms on many image classification tasks such as MNIST, CIFAR or ImageNet. Thus, advances in training methods or structural improvements of neural networks are evaluated mainly on image data. The direct applicability of new methods to other types of data (e.g. time-series data such as audio and music) often remains unclear. To investigate this, we offer the following types of master theses or practical projects:

  • Deep Neural Networks for Audio Processing
    Your goal will be to acquaint yourself with cutting-edge NN techniques and apply them on audio/music related tasks. The procedure will be similar to the following:

    1. Select a recent method from this list of state-of-art systems for image classification, try to reproduce the results.
    2. Find a way to apply the method on an audio-related task such as music transcription, chord detection, beat tracking, onset detection, etc. We will provide you standard datasets for each of these tasks.
    3. Analyse the results.

  • Learning "Artistic Style" and Rewriting Music

    "A Neural Algorithm of Artistic Style" (see this paper) is a novel method of "repainting" one image in the style of another, using deep neural networks. Here you can find more examples. In this master thesis, you will investigate if and how this method is applicable to music (in form of audio or MIDI files). You will investigate questions such as:



Contact: Rainer Kelz, Filip Korzeniowski, Gerhard Widmer



Conditional Random Fields

Conditional Random Fields are powerful models for structured prediction, where labels (or, classes) of consecutive data points are strongly correlated. While CRFs are often used in audio processing, they still rely on manually designed features. Recent advances in this field aim at leveraging the power of automatically learned features (using deep learning) for CRFs. In this context we offer the following topics for theses and practical works:

  • Conditional Random Fields with (Deep) Neural Feature Functions for Audio Processing
    Explore and evaluate existing CRF models that integrate automatic feature learning, and if possible, improve them.


Contact: Filip Korzeniowski, Gerhard Widmer






last edited by kf on Sep 28, 2015