02455 Project

Project 1: Music genre classification using neural networks

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 build a system for music genre classification and using a Bayesian neural network classifier.

Detailed description:

Implementation of multi-time scale features extraction techniques
Setting up a MacKay Bayesian Neural network simulator for classification
Test and evaluation of genre classification. 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, Phone: +45 45253923.

Project 2: Adaptive Regularization in Neural Network Filters

The project concerns neural networks for prediction of time signals e.g., Santa Fe time series competition data. The quality of the predictions made by the network is measured in terms of the so-called generalization error. A simple criterion for optimization of the neural network architecture is to choose the network with minimal generalization error. Optimal Brain Damage (OBD) prunes the network structure is order to minimize generalization error. However, also regularization of the networks weights is very important for facilitating training and will further help in improving generalization.

The purpose of the project is to implement a neural network simulator which features adaptive regularation and network structure optimization and test it on numerous time series predction problems.

Detailed description:

Implemention of a neural network simulator in MATLAB which enables the training of a two-layer network with a single linear output neuron.
Signal preprocessing, trending and representation.
Extension of the simulator to include OBD and various forms of adaptive regularization, from cross-validation to Bayesian optimization of regularization parameters.
Test the methods on various time series prediction problems.

Additional material

Teacher: Jan Larsen, e-mail: jl@imm.dtu.dk, Phone: +45 45253923.

Project 3: On-line Training of Neural Network Filters

The project concerns neural networks for prediction of time signals. In situation where the signals are non-stationary the model is continuously required to be adapted in order to track non-stationarities. This calls for on-line training algorithms of network weights. Potentially also the network structure and amount of regularization should change gradually.

The purpose of the project is to implement a neural network simulator which features on-line training and possibly adaptive regularation and network structure optimization.

Detailed description:

Development of on-line training methods which inlude regularization.
Implemention of the on-line neural network simulator in MATLAB which enables
Extension of the simulator to include methods for network structure optimization.
Test and validation of the methods on various time series prediction problems, including:
- robutness and quality of on-line training methods.
- determination of memorylength via the algorithm's forgetting factor.
- validation of training methods by estimating generalization error.
- possibility of gradually changing network structure and amount of reguarization.

Additional material

Teacher: Jan Larsen, e-mail: jl@imm.dtu.dk, Phone: +45 45253923.

Project 4: Comparison of Perceptron Neural Networks and Relevance Vector Machines

The project concerns feed-forward neural networks and relevance vector machines for prediction of time signals.

The purpose of the project is to provide a comparative study of a classical feed-forward neural network and the Relevance Vector Machinefor the time series prediction.

Detailed description:

Reading and understanding the classical feed-foward neural network and the Relevance Vector Machine.
Implemention of algorithms i MATLAB
Test and validation of the methods on various time series prediction problems, including:
- robutness and quality.
- validation of training methods by estimating generalization error.

Additional material

Teacher: Jan Larsen, e-mail: jl@imm.dtu.dk, Phone: +45 45253923.

Project 5: Bayesian Signal Detection: Linear systems

This project concerns detection of signals that can be expanded in terms of a low-dimensional basis set of functions. Examples: Detection of periodic signals in noise, detection of signals of a given shape with unknown delay and strength, detection of narrow band signals using ``slepian basis functions''

Possible applications in Neuroimaging

detection of activation in brain scanner signals
Cleaning signals for narrow band noise

Additional material

Teacher: Lars Kai Hansen, e-mail: lkhansen@imm.dtu.dk, Phone, +45 45253889.

Project 6: Bayesian Signal Detection: Theory project

This project concerns theoretical investigations of Bayesian detection.

Possible directions for the analyses

The case of mis-specified priors. How does the Bayesian signal detector operate as the prior deviates from the true parameter distribution?.
Analytical and computer experiments.
Variational and sampling based methods. For some Bayesian problems it is impossible to do the integrals and other approximate methods must be invoked. So-called variational Bayes methods have been developed for systems with exponential family pdf's and conjugate priors

Additional material

Teacher: Lars Kai Hansen, e-mail: lkhansen@imm.dtu.dk, Phone, +45 45253889.

Project 7: Bayesian GLM for analysis of fMRI data

This project concerns detection of fMRI signal using a Bayesian General Linear Model (GLM). The project will include specification of priors and comparisons of the properties of a Bayesian GLM to a standard GLM (parameters obtained with Maximum Likelihood).

Possible subjects:

Model selection: Use the posterior probabilities to determine the number of harmonics in the signal and the most likely model.
Robustness: Investigate robustness of MAP parameters compared to ML parameters.
Priors: Investigate the prior distribution (of parameters and/or noise).

Additional material

Teacher: Kristoffer Hougaard Madsen, e-mail: khm@imm.dtu.dk, Phone, +45 45253894.

Project 8: Comparing supervised and unsupervised methods for analysis of fMRI data

This project concerns unsupervised detection of fMRI signals and comparisons to supervised methods. The project will include construction of simulated fMRI imaging data. Examples: Clustering, Principal Component Analysis and Independent Component Analysis.

Possible data applications:

Signal detection in BOLD fMRI
Filtering of noise contributions in BOLD fMRI data including movement and physiologic confounds.

Additional material

Teacher: Kristoffer Hougaard Madsen, e-mail: khm@imm.dtu.dk, Phone, +45 45253894.

Project 9: Bayesian Neural Networks for Segmentation of fMRI Brain Scan data

Brain scan analysis using fMRI is an important and non-invasive technqiue. In order to assist doctors with diagnosis support advanced neural network techniques can be used. Here we will use the network to establish the relation of fMRI signal modalities and known labeled conditions of the brain.

The purpose of the project is to system for segmentation of fMRI signals into relevant classes using a Bayesian neural network classifier.

Detailed description:

Study of the data set and feature extraction from multiple signal modalities. Including feeding signals from an region of interest around desried voxel.
Setting up a MacKay Bayesian Neural network simulator for classification
Test, optimization and evaluation. This involves estimating confusion matrices and comparison with simple classification schemes such as Gaussian or linear classifiers.
Building a Matlab demo.
Extending the training software to handle soft-label training (non multual exlusive labels).

Additional material

Teacher: Jan Larsen, e-mail: jl@imm.dtu.dk, Phone: +45 45253923.