Short Description

In machine learning, classification algorithms can be derived from a generative or a discriminative paradigm. Generative classification models (e.g. Naive Bayes) model the joint distribution over features and class labels, while discriminative models (e.g. logistic regression, Support Vector Machines) directly aim to maximize classification performance without considering the distribution over the input features. While discriminative classifiers have a better asymptotic classification performance, generative classifiers can make use of unlabeled data, i.e. for datasets where only a few of the features have class-labels, only generative classifiers can be used (“semi-supervised learning”).

To make use of the advantages of both paradigms, generative-discriminative hybrid models have been proposed. The aim of this master thesis is to study and implement some of the most prominent methods and apply them to classification tasks in speech signal processing.

Your Tasks

• Literature review of theory and survey of most prominent methods for generative-discriminative hybrids
• Mathematical derivation and implementation at least two hybrid classification algorithms (e.g. [1]) in MATLAB
• Apply the implemented algorithms to classification tasks in speech and audio, and experimentally evaluate the performance of the implemented algorithms under different conditions

Your Profile/Requirements

• High motivation
• Very good knowledge in machine learning and signal processing
• Good knowledge in MATLAB programming

Contact:

Sebastian Tschiatschek (tschiatschek@tugraz.at or 0316/873 4366)

References

[1] C. Bishop and J. Lasserre, “Generative or Discriminative? Getting the Best of Both Worlds,” Bayesian Statistics, 2007.