You are here
Emanuele Olivetti: Presentation/Talk at University of Maastrict
1) Two-Sample Tests vs. Decoding
Assessing whether the patterns of brain activity systematically differ when the subject is presented with different sets of stimuli is called "brain decoding". The most common solution to this problem is based on testing whether a classifier can accurately predict the type of stimulus from brain data. In this work we present a novel approach to the brain decoding problem which does not require any classifier. The proposed method is based on a kernel-based high-dimensional two-sample test recently proposed in the machine learning literature. This frequentist hypothesis test tries to determine whether the set of brain recordings related to one kind of stimulus, i.e. the first sample, and the ones related to the other kind of stimulus, i.e. the second sample, are drawn from the same probability distribution or not. In this work we illustrate the advantages of this novel approach together with experimental evidence of its efficacy on magnetoencephalographic (MEG) data from a Face, House and Body discrimination task. Additionally we present some preliminary steps of our new investigation on a Bayesian two-sample test based on a Gaussian process prior.
2) Hypothesis Testing with Classifiers
Machine learning is increasingly adopted in neuroimaging-based neuroscience studies. The paradigm of predicting the stimuli provided to the subject from the concurrent brain activity is known as brain decoding and accurate predictions support the hypothesis that the brain activity encodes those stimuli. Testing the hypothesis whether the classifier is able to decode the stimuli or not is a crucial step of this approach. In this talk we will introduce the three main frameworks for testing hypothesis, i.e. Fisherian, Neyman-Pearson and Bayesian, and for each of them we will show examples on how to test classifiers. Then we will discuss the issues of using cross-validation when testing classifiers. We will conclude by proposing a novel Bayesian test of independence that is meant to address many of the shortcomings of the traditional test for classifiers.