Morten Mørup

Post Doc. - Informatics and Mathematical Modeling,

Technical University of Denmark, Bld. 321/118, 2800 Kgs. Lyngby, phone +45 4525 3900

email: mm@imm.dtu.dk or morten.morup@gmail.com

Research

My primary research interest is in the field of Machine Learning and data mining, attempting to form methods that can blindly extract the relevant information in data. My main focus is in biomedical applications in the field of Neuro-Informatics and Bio-Informatics.

Below is some excerpts from recent work.

 

Shift invariant Decomposition

When obtaining data there is often a propagation delay between the signals and sensors rendering instantaneous blind source separation models such as ICA/NMF invalid. Thus, we have researched bi-linear and multi-linear models that can handle propagation delay. In particular we have demonstrated how delay modeling can improve component identification and  alleviate degeneracy encountered in multi-linear models

ERPWAVELAB for multi-channel time-frequency
Through my work on time-frequency analysis of event related wavelet transformed data I applied multi-way techniques such as the PARAFAC and TUCKER model in combination with non-negative decomposition techniques. As to my knowledge no toolbox was publicly available for multi-channel time-frequency analysis. Consequently, I created ERPWAVELAB in collaboration with Sidse M. Arnfred at Hvidovre Hospital. The toolbox is under ongoing development and is freely available at www.ERPWAVELAB.org.

Double Deconvolution for Music Transcription
Each instrument of music has a more or less unique harmonic frequency structure that to some extent can be considered purely shifted in the log-frequency domain when the various keys of the instruments are played. Consequently, in collaboration with Mikkel N. Schmidt a model that can account for these basic assumption to separate and transcribe music was formed. The model we formed turned out to be a tensor double deconvolution for which we devised two algorithms to analyze the absolute log-spectrogram of music. Some of this work can also be found at www.intelligentsound.org.

Tensor double deconvolution for music transcription and separation

Graphical user interface of  ERPWAVELAB