A Generative-Discriminative Approach to Human Brain Mapping
Abstract
During everyday behaviours, the brain shows complex spatial patterns of activity. These activity maps are very replicable within an individual, but vary significantly across individuals, even though they are evoked by the same behaviour. It is unknown how differences in these spatial patterns relate to differences in behavior or function. More fundamentally, the structural, developmental, and genetic factors that determine the spatial organisation of these brain maps in each individual are unclear. Here we propose a new quantitative approach for uncovering the basic principles by which functional brain maps are organized. We propose to take an generative-discriminative approach to human brain mapping, with the fundamental idea that, if we understand the underlying principles that organises brain activity maps, we should be able to generate sets of artificial maps that are indistinguishable from real ones. Different generative models are tested by a series of adversarial, classifier models, ranging from linear classifiers based on specific marginal statistics, to a full convolutional neural network. We apply our new framework to a collection of measured finger activity maps measures with fMRI in the human sensori-motor cortex (N=50). To account for characteristics of the brain maps that depend on the specific measurement process (spatial resolution of fMRI, signal-to-noise, etc) we supplemented the generative process with a measurement model. Initial results clearly demonstrate that the matching of simple marginal statistics (covariance and smoothness of activity pattern) is sufficient to fool the human eye, but not a more systematic machine learning approach. The proposed evaluation framework therefore opens up a pathway for discovering specific characteristics of brain activity maps that are important to explain function or individual differences.