ABSTRACT

Functional magnetic resonance imaging (fMRI) measures brain signals at a resolution of approximately 1-3 mm and thus is the spatially most accurate noninvasive measurement technique that can be applied in human brain research. From its beginning, fMRI data was commonly analyzed by a mass-univariate approach based on the general linear model (GLM) and statistical parametric mapping (SPM; Friston et al., 1995). Univariate statistical tests are performed and the results are plotted at each position of the brain individually. Usually, the statistical maps are spatially smoothed to allow for inhomogeneities in the anatomy of individual subjects. The GLM/SPM has been very successful in locating brain regions, the activation of which is modulated by a specic mental operation. However, the brain often encodes mental contents using distributed populations of neurons (Georgopoulos et al., 1986; Tanaka, 1996) and their correlations are important (Averbeck et al., 2006). For example, the sum of the preferred movement directions of individual motor cortex cells, the so-called population vector, precisely codes movement direction (Georgopoulos et  al., 1986). Similarly, the object, a monkey, is currently viewing can be decoded by applying classiers to multiple unit recordings in the temporal lobe (Hung et al., 2005). The standard GLM approach is blind to such patterned activity. Thus, it has frequently remained unclear in human neuroimaging how and where the information about specic mental contents is processed and represented in the brain. The question of how the external world is represented in the brain, that is, the question of neural coding, has been nicely illustrated by Ernst Mach (1886) in his famous drawing of his room through his own eye. Two basic types of multivariate code are often distinguished, a sparse code and a distributed code (see Figure 24.1; Haynes, 2009).