You are here
Emanuele Olivetti: Presentation/Poster at BNP Conference
Brain imaging experiments on humans are frequently used in confirmatory neuroscience studies where different scientific hypotheses about how the brain works are compared. In the most common case the subject is provided a sequence of i.i.d. stimuli from two classes while the concurrent brain activity is recorded. One hypothesis postulates that the brain area of interest to the study is differentially involved in processing those stimuli, while the other hypothesis postulates that no difference occurs (e.g. Knops et al 2009, Van Gerven 2009). Then the experiment is meant to collect evidence to support one of the two hypotheses.
In the field of neuroimaging data analysis, the best practice to address such a confirmatory data analysis problem is to train a classifier on part of the high-dimensional class-labelled recordings and then to test whether the classifier is able to correctly predict the classes of stimulus on the remaining part of the data (Pereira et al. 2008). The ability to learn the discrimination problem is interpreted as evidence of differential brain activity in the area of interest.
In this work we propose an alternative solution that does not involve the use of classifiers. We recast the hypothesis testing problem as a high-dimensional two-sample test, where each sample is the set of brain activity recordings related to one of the two classes of stimulus. Differently from most of the current literature on high-dimensional two-sample tests (Gretton et al. 2012), we adopt the Bayesian framework and we propose the use of a nonparametric approach. We discuss the motivation for both the choices and we show experimental evidence of the effectiveness of this approach on magnetoencephalography (MEG) data from a covert spatial attention task.