Skip to Main content Skip to Navigation
Theses

A multi-source perspective on inter-subject learning. Contributions to neuroimaging.

Abstract : Inter-subject learning is a family of learning problems encountered in the analysis of data recorded in human subjects where we need to perform predictions on data recorded from a subject that was not available at training time. The most usual problem that uses inter-subject learning is to ask whether an unknown individual is healthy or sick, i.e to design a computer-aided diagnosis tool. In this thesis, we argue that such inter-subject learning questions should be addressed within the multi-source learning framework, and we formalize it as such in the context of neuroimaging studies. Indeed, each subject is a different source of data, with data samples that potentially live in different feature spaces and that are drawn from different probability distributions. The multi-source setting therefore constitutes an extenstion of the domain adaptation problem where a single source of training data is available. We then introduce three original contributions motivated by inter-subject learning questions in neuroimaging. The result of our first contribution is a method that is able to perform reliable inter-subject predictions from fMRI data using fine-scale spatial patterns defined within a region of interest. Because of the strong inter-subject variability present at such fine scale, the original feature spaces are different across subjects. Our contribution consists in designing a common space for the patterns of all subjects using graphical representations of the patterns together with a graph kernel that implicitly projects the samples into a reproducing kernel hilbert space. We show that this approach is effective through the increased accuracy achieved on an inter-subject prediction task designed to study the functional organization of the human auditory cortex. Our second contribution is a new method that enables to detect local differences in cortical shape across groups of anatomical MRI scans. The objects used to detect such differences are, yet again, graphical reprentations, this time designed from the spatial organization of the sulcal pits – the deepest points of cortical sulci. Using a graph kernel designed for these objects allows to project them into a reproducing kernel hilbert space and to quantify the differences between groups through the performances of a classifier trained to recognize these groups. A non-parametric spatial inference method is then proposed to perform the detection of cortical zones where the differences are statistically significant. We validate this method by showing that it detects cortical asymmetries and gender differences using a large database of healthy subjects. The third contribution of this thesis is a multi-source domain adaptation technique. Our method builds upon the kernel mean matching, a distribution matching procedure that estimates importance weights for the training samples so that the weighted source distribution matches more closely the target distribution than the unweighted one. We introduce an extension of the kernel mean matching for the multi-source case, i.e when the training samples are drawn from several sources of data. We present preliminary results of this framework on a inter-subject prediction task used to analyse data from a magneto-encephalography experiment.
Complete list of metadata

Cited literature [201 references]  Display  Hide  Download

https://hal.archives-ouvertes.fr/tel-01251384
Contributor : Sylvain Takerkart <>
Submitted on : Wednesday, January 27, 2016 - 11:12:04 PM
Last modification on : Friday, November 27, 2020 - 3:08:49 AM
Long-term archiving on: : Thursday, April 28, 2016 - 11:30:33 AM

Licence


Distributed under a Creative Commons Attribution - NonCommercial - NoDerivatives 4.0 International License

Identifiers

  • HAL Id : tel-01251384, version 3

Collections

Citation

Sylvain Takerkart. A multi-source perspective on inter-subject learning. Contributions to neuroimaging.. Machine Learning [cs.LG]. Aix-Marseille Université, 2015. English. ⟨tel-01251384v3⟩

Share

Metrics

Record views

537

Files downloads

587