| dc.description.abstract | While the visual system is a relatively well-studied region of the brain, the interactions between its neural
populations are poorly understood. In this project, we proposed an approach to model these interactions by
combining encoding and decoding modelling in a natural scene paradigm. We utilized an open-source fMRI
dataset, which was recorded using natural scene images as stimuli to model the response amplitude of each
voxel to these images. To achieve this, we combined encoding and decoding modelling: we created a 2D representational geometry of each ROI, which depicts the difference between the responses to each image. We
fitted a circular Gaussian function to each voxel’s responses to all images on the 2D representational spaces
using two strategies: we fitted our response functions using the representational geometry of all other ROIs
for one model and to the one of a voxel’s own ROI for another model. The fitted Gaussian function then gave
us the preferred position and size of the response function of each voxel. We used these variables to obtain
the predicted response amplitude of each voxel to each image by passing the 2D representational space of a
voxel’s ROI through our fitted Gaussian functions. From this, we calculated the amount of variance our models can explain in each voxel. Then, we found which 2D representational space could best explain each ROI.
We found a circular relationship between the early regions of the visuals system (V1-V3) and that TO-1 was
the ROI that could best explain most later regions. Additionally, our models’ predictive performances were
the highest in later regions, indicating that voxels in these regions have object recognition responses. Furthermore, we compared the distances between each pair of voxels’ preferred position on these 2D space and their
position on the cortical surface. We found that these two distances are moderately correlated for every ROI,
unravelling a structural organization across the cortical surface. Finally, our proposed methodology allowed
us to study the brain’s responses without continuous and low-dimensionality parameters, showing great potential for further study of the brain’s response to more complex and naturalistic stimuli. | |
