| Recent functional MRI devices allow to scan the whole brain with a resolution of 2 - 3 mm in 1 - 3 seconds. For special questions not requiring whole-brain scanning, both spatial and temporal sampling can be further increased to sub-millimeter and sub-second ranges. Furthermore, the combination of EEG and MEG with fMRI allows to integrate high-temporal resolution data in a common source (cortex) space. For example, we recently have demonstrated that EEG oscillations (power modulation) can be predicted from fMRI data. These technical and experimental developments are accompanied by an increased sophistication in data analysis techniques, including local and global multivariate machine learning tools aimed to decode distributed representations. In light of these technical and methodological developments, we present a framework for building large-scale neural network models exploiting specific spatio-temporal constraints in a novel way. In this framework, predicted dynamic activity patterns evolve directly on brain models derived from anatomical MRI data of a specific subject allowing to analyze empirial and predicted data with the same tools. The proposed framework will be demonstrated on a model of perceptual filling-in, which is based on empirically supported principles of surface perception and known architecture of early visual areas. The model explicitly simulates sub- and supra threshold activity, and, therefore, generates predictions not only for the activity distributions of spiking neurons, but also for fMRI. The model demonstrates the consequences of subthreshold neural spread of the BOLD signal for the activity distribution obtained with stimuli typically used to investigate perceptual filling-in, and it provides insight into the divergent data from human fMRI and neurophysiological experiments in animals. |
|