Abstract:
This report describes the application of an incremental cluster-wise regression analyses method to the EBC brain activity interpretation competition data. The method is particularly suitable in brain function studies where there is limited or no prior knowledge about the structure of the experimental design, as often is the case with natural sensory stimulations. The method (i) represents sensory stimulation data (feature signals) and brain activation data (voxel signals) as functional data, (ii) operates on clusters of voxel signals of arbitrary size, (iii) takes into account Haemodynamic differences between brain areas and across subjects and (iv) adopts an automatic “search-model-fit” strategy. The method has been used in the EBC competition to perform subject analysis as well as group analysis. The outcome of these analyses is not only the location of brain activations that highly correlate with the EBC Competition feature signals, but also regression functions, obtained from clusters of voxel signals that best correlate with the feature signal. These regression functions allow estimation of feature signals from new fMRI data. We have applied this method to estimate the EBC defined Base, Person and Location feature signals for Movie3, using Movie1, Movie2 fMRI and feature ratings as training data. Our first attempt only involved Base features, for which EBC computed correlation scores ranged between -0.17 (“Tools”) and 0.48 (“BodyParts”). The second attempt included Person features, for which strikingly only high correlations of up to 0.7 (“Jill”) were obtained. The results, especially for Person features, are promising given the limited amount of training data.
Sennay Ghebreab: http://staff.science.uva.nl/~ghebreab/
Sennay Ghebreab received an M.Sc in information systems in 1996, and a Ph.D. on the topic of medical image analysis in 2002, both from the University of Amsterdam. He moved to the Erasmus University Medical Center in 2002 to do postdoc research in the fields of visual learning from examples, model-based analysis of medical images, and content-based retrieval from large neuro-image databases. At the end of 2005 he returned to the Informatics Institute of the University of Amsterdam, where he is now a researcher participating in two national programs (MultimediaN and VL-e) on the topic of discovering and understanding human brain activity patterns arising from natural sensory stimulation. His research interests include cognitive vision, image analysis, pattern recognition, information retrieval, machine learning and data mining.
Pieter W. Adriaans: http://staff.science.uva.nl/~pietera/
Pieter Adriaans (1955) studied philosophy in Leiden, the Netherlands, under Nuchelmans and van Peursen. He was research assistant of Fresco for a while with the study of the philosophical estate of the well known Dutch philosopher and poet Johan Adreas Dér Mouw 1863-1919 as a special assignment. In 1983 he graduated and started to work as a software developer, and later service manager for Buro Microsoftware. In 1985 he became general manager of Compu'Disc and later general manager of Info'Products Informatica Diensten. He has been active in research in the areas of artificial intelligence and relational database systems since 1984. He and his business partner, Dolf Zantinge, founded Syllogic B.V. in 1989. In 1992 Adriaans received a PhD in computer science at the University of Amsterdam, where he is now a part-time professor in machine learning/artificial intelligence since 1998.
Arnold W.M. Smeulders: http://staff.science.uva.nl/~smeulder
Arnold W.M. Smeulders graduated from Technical University of Delft in physics in 1977 (MSc) and in 1982 from Leiden University in medicine (PhD) on the topic of visual pattern analysis. He is scientic director of the Intelligent Systems Lab Amsterdam, of the MultimediaN the Dutch public-private partnership, and of the ASCI national research school. He participates in the EU-Vision, DELOS and MUSCLE networks of excellence. He is fellow of International Association of Pattern Recognition. His research interest is in cognitive vision, content-based image retrieval, learning and tracking, and the picturelanguage question. He has written 300 papers in refereed journals and conferences and graduated 28 PhD-students. Currently he is an associated editor of the International Journal of Computer Vision and the IEEE Transactions on Multimedia.
