|
|
Section for Cognitive Systems |
DTU Informatics |
|
Project 1: Song features for meta data classification
The demand for computational methods to organize
and search in digital music has grown with the increasing
availability of large music databases as well as the growing access
through the Internet.
The purpose of the project is to compare methods for predicting and
understanding higher level attributes of the song, so-called meta data such as genre or mood.
Detailed description:
- Identification of a number of methods for low-level feature extraction e.g. beat and Mel-Ceptrum (MFCC).
- Implementation of the feature extraction methods in MATLAB
- Test and evaluation, including estimation of accuracy.
- Design and implementation of a sparse classifier for predicting genre or mood.
- Test and evaluation: This involves estimating confusion matrices and comparison with
simple classification schemes such as Gaussian or linear classifiers.
- Building a Matlab demo.
Additional material
Teacher: Jan Larsen e-mail: jl@imm.dtu.dk.
Project 2: Music instrument detection
During the recent years we have witnessed an explosion in the amount of multimedia information available to the average user. This scenario could benefit from the appearance of intelligent tools that automatically extract knowledge from this information. In particular, a problem that has received some attention during the last years, is the decomposition of a music file into the waveforms from the different instruments.
It is reasonable to think that this separation process could be simplified by previously determining which instruments are present within a particular music clip.
The purpose of this project is to design a system for the automatic classification of music signals into a limited group of instruments. Although the system will be built to classify song segments where only one instrument is played at a time, an additional and interesting possibility consists in analyzing the ability of the system to detect the absence/presence of several instruments, when segments with multiple instruments are analyzed.
Detailed description:
- Extraction of features from a large database of musical instrument samples.
- Identification of relevant features for characterizing the most important and distinguishing properties of instruments.
- Performance evaluation of different methods for multiple-class classification. In particular we will use a sparse input (ARD) Gaussian process.
The evaluation includes instrument recognition accuracy rate, as well as the analysis of confusion matrices.
- Building a Matlab demo.
- Analyzing the performance of the system when it is presented music clips containing a mixture of different instruments
Additional material
Teacher: Jan Larsen, e-mail: jl@imm.dtu.dk
Project 3: Voice detection in music using sparse models
Music information retrieval (MIR) is a hot topic in machine learning. Detection of vocal parts of music
can assist MIR systems and be used for further refinements, e.g., recognition etc.
The purpose of this project is to design a system for the automatic detection of vocal parts of music signals using a sparse models.
Detailed description:
- Study of current state-of-the-art methods based on conventional music processing pipe-lines for voice detection
- Train high-dimensional linear models with sparsity constraints from log-spectra
- Train sparse Gaussian process models with from log-spectra
- Building a Matlab demo.
- Evaluating performance by comparing with state-of-the-art methods
Teachers: Jan Larsen e-mail: jl@imm.dtu.dk.
Project 4: Modelling fMRI data by kernel PCA and ICA
Brain imaging by functional magnetic resonance imaging (fMRI) was invented
in 1992 and has revolutionized neuroscience. By fMRI we have gained substantial
new insight in macroscopic information processing.
Principal and independent component analyses of fMRI have been used to
detect global patterns of activation in the human brain within a linear model.
Recently we sat an informal "world record" in mental state decoding precision
using non-linear features based on the so-called Laplace eigenmap.
The purpose of this project is to use non-linear feature
detectors based on kernel representations, e.g., kernel PCA or ICA.
A possible application would be the so-called Pittsburgh Mind Reading Competition.
Additional material
Teacher: Lars Kai Hansen, e-mail: lkh@imm.dtu.dk.
Project 5: Social networks - Community detection by spectral clustering
Human interaction can be described using graphs, so-called social networks.
One of the hot topics in current research is to identify highly interconnected
subgraphs: `Communities'. Principal component-like methods called spectral clustering
have shown effective for community detection in sparse networks.
The purpose of this project is to analyze community structure
in large graphs like the "movie actor network" with 125.000 movie actors
participating in 400.000 movies.
Additional material
Teacher: Lars Kai Hansen, e-mail: lkh@imm.dtu.dk.
Project 6: Sparse Factor Models
A factor model is a powerful tool for latent structure analysis in linear systems that is intended to explain a set of observed variables as a linear mixing of set of latent factors, for example as in principal component analysis (PCA) or independent component analysis (ICA). In many real life applications, the so called linear mixing is sparse, meaning that only a few of the possible latent factors contribute to explain a variable, so there is a need to systematically impose/promote sparsity in factor models. Bayesian models and sampling methods allows us to incorporate prior knowledge, i.e. sparsity, into factor models to be able to handle such applications properly while keeping the interpretability of the model.
The purpose of the project is to use Bayesian sparse factor models to find possible interactions between observed data and a set of latent variables, e.g. to relate gene expression levels (latent) transcription factors or song lyrics with (latent) emotional words/concepts.
Detailed description:
- Literature study (especially Shimizu et. al., JMLR 7, p. 2003, (2006)).
- Extending the existing InfoMax algorithm to causal modelling.
- Evaluation and demonstration on real and artificial data (for example from the above reference) using cross-validation.
Additional material
Teacher: Ole Winther, e-mail: owi@imm.dtu.dk.
Project 7: Markov Chain Monte Carlo for Dynamical Systems
We will consider Markov chain Monte Carlo (MCMC) methods for inference is dynamical state space models. We will focus on a particular class of algorithms called particle filters and will study both their theoretical properties and practical issues around better proposal distributions and joint state and parameter estimation.
The purpose of this project is to get a thorough introduction to both theoretical and practical aspects of MCMC exemplified by particle filtering.
Detailed description:
- Literature study on particle filtering.
- Theoretical analysis of algorithms.
- Development of proposal distributions that work for strongly multi-modal systems.
- Evaluation and demonstration on toy and real data examples.
Teacher: Ole Winther, e-mail: owi@imm.dtu.dk.
Project 8: The Netflix Prize - Low Rank Decompositions and Model Averaging
We will try to win the 1M US$ Netflix Prize in this project. The best systems developed so far have two important ingredients which we want to understand and improve upon using for example a Bayesian approach. These are models based low rank decompositions and model averaging, that is averaging in some clever way the predictions coming from different models.
The purpose of this project is to understand successful approaches to the Netflix Prize problem and improve on these.
Detailed description:
- Literature study on leader board methods
- Implementation of these methods and modifications in Matlab
- Evaluation and demonstration on the Netflix qualifying set.
Additional material
Teachers: Ole Winther e-mail: owi@imm.dtu.dk.
Project 9: Infinite mixture models
In mixture models, data is modeled as generated from one out of K possible mixture components. Often, it is not known how many components there is, so it is desirable to learn this from data. One way to do this sometimes denoted the "infinite" mixture model because the number of components is unbounded.
The purpose of the project is to understand and implement an infinite mixture model.
Detailed description:
- Literature study on infinite mixture models.
- Implementation of conjugate (and possibly non-conjugate) mixture models in Matlab.
- Evaluation and demonstration on toy and real data examples.
Additional material
Teacher: Mikkel N. Schmidt e-mail: mns@imm.dtu.dk.
Project 10: Slice sampling
In Bayesian data analysis inference usually entails evaluating a high-dimensional intractable integral, which is often approximated using Monte Carlo. Recently researchers have proposed different general schemes for Monte Carlo inference, of which one of the most promising is the slice sampler.
The purpose of the project is to understand and implement a general slice sampling procedure.
There are three major areas that could be investigated in the context of this project:
- Literature study on Markov chain Monte Carlo and slice sampling.
- Implementation of a general slice sampler in Matlab.
- Evaluation and demonstration on toy and real data examples.
Additional material
Teacher: Mikkel Schmidt e-mail: mns@imm.dtu.dk.
Project 11: Graph based semi supervised learning
In classification problems we are given a set of training examples with known class labels, and we wish to infer a classification rule. In semi-supervised learning, some of the class labels of are "missing" and need to be inferred. Interestingly, the unlabelled examples still contain useful information. A popular way of approaching this problem is to represent the data set as a graph and then to employ discrete regularisation techniques such as label diffusions.
The purpose of the project is to understand and implement a graph based learning algorithm.
There are three major areas that could be investigated in the context of this project:
- Literature review on graph based regularisation
- Implementation of a graph based semi supervised learning algortihm in Matlab.
- Evaluation and demonstration on toy and real data examples.
Additional material
Teacher: Christian Walder e-mail: chwa@imm.dtu.dk.