Receptive Field Encoding Model for Dynamic Natural Vision

Encoding models are used to predict human brain activity in response to sensory stimuli. The purpose of these models is to explain how sensory information represent in the brain. Convolutional neural networks trained by images are capable of encoding magnetic resonance imaging data of humans viewing natural images. Considering the hemodynamic response function, these networks are capable of estimating the blood oxygen level dependence of subject viewing videos without any recurrence or feedback mechanism. For this purpose, feature map extracted from the convolutional neural network and the concept of receptive field has been used for the encoding model. The main assumption of this model is that activity in each voxel encodes a spatially localized region across multiple feature maps and for each voxel and this area are fixed for all feature maps. Contribution of each feature map in the activity of each voxel is determined by the corresponding weight.

In this study, three healthy volunteers watching a set of videos. This collection contains images that represent real-life visual experience. MRI and fMRI data are acquired on a 3 tesla MRI system phase-array surface coil.

Data revealed that human visual cortex has hierarchical structure. Earlier visual areas have a smaller receptive field size in and response to simple feature like edge, whereas higher visual areas have a larger receptive field size and response to more complex features, such as pattern.

This model of video stimuli has a higher interpretation capacity than the previous models.