[home] [Personal Program] [Help]

---

Session: Poster Session 1 (Even numbers)
Session starts: Thursday 26 January, 16:00
Presentation starts: 16:00

---

Dora Gözükara (Radboud University)
Djamari Oetringer (Radboud University)
Umut Güçlü (Radboud University)
Linda Geerligs (Radboud University)

Abstract:
Studying the feature maps of artificial neural network (ANN) models and their relation to the brain have led to many new insights in computational neuroscience. ANN features are being used in a variety of ways, from predicting neural activity to understanding neural computation. Nevertheless, making these features comparable with brain data is far from a straightforward task. Conventionally, features from an entire ANN layer are grouped together, regardless of their spatial selectivity or size, and are compared with spatially selective brain areas, such as the early visual cortex. Consequently, a large amount of data is needed to train successful brain encoding models with large parameter spaces, and these models in turn need to learn which spatial location is relevant for the voxel to which they are being fit. In this work, we are building personalized encoding models that predict voxel timeseries data from a movie-viewing fMRI dataset using ANN features. We use eye-tracking and retinotopic data to build our personalized models that are not only specific to each participant, but also to each participant’s voxels. We do this by combining eye-tracking and voxel population receptive field data to sample only the relevant parts of the ANN feature map at only the relevant timepoints in the movie. We show that our personalized models make it possible to successfully train encoding models using limited data by reducing the model parameter space. We also present a new way to structure ANN features so that their comparison with fMRI data becomes more straightforward.