Modern cognitive neuroscience has focused on mapping what information is processed both where and when in the brain as an initial step towards understanding how information is processed by the brain. Brain mapping is most commonly performed using the linear encoding model - a computational tool that can be used to align brain measurements with numerical stimulus representations that can be informative of the underlying neural processes. But as we move towards more naturalistic experimental settings and start to work with increasingly complex stimulus representations, it becomes important to test the boundaries of what we are able to learn using existing computational tools so we can adjust them accordingly. In this work, we examine two potential limitations of the linear encoding model. First, we look at a specific problem setting that shows that the existing framework does not always allow us to disentangle where specific stimulus-related information is processed in the brain. As an initial step towards addressing this issue, we propose a new framework that can be used to group together areas of the brain that are responsible for processing similar information. Second, we explore whether using a more expressive, nonlinear encoding model can allow us to better align the internal representational spaces of artificial neural networks, from which we can derive more complex and potentially more informative stimulus representations, with those of the brain. Overall, our proposed extensions to conventional encoding approaches can help neuroscientists unlock a richer and more complete space of brain mappings.

66 pages

Thesis Committee:
Leila Weibe (Chair)
Mi chael J. Tarr

Srinivasan Seshan, Head, Computer Science Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu