A two-stage approach to estimating voxel-specific encoding models improves prediction of hemodynamic responses to natural images
-
1
Radboud University Nijmegen, Donders Institute for Brain, Cognition and Behaviour, Netherlands
Unsupervised feature learning has become an alternative approach to descriptive and explanatory modeling of single voxel responses to natural images in functional magnetic resonance imaging. In this approach, features are first learned from low-resolution and small natural image patches since learning generally requires a large amount of memory and computation power. Features are then nonlinearly extracted from stimuli and finally linearly regressed on stimulus-evoked single voxel responses. While this approach was recently shown to improve prediction of human brain activity in response to natural images, it has two problems. First, features that are learned from low-resolution and small natural image patches might not be adapted to statistical regularities that alter single voxel responses. Second, features that are extracted from stimulus regions that are outside single voxel receptive fields might be notoriously redundant, and overfitting might occur. In this study, we introduce a two-stage approach to solve these problems. In the first stage, a general encoding model is estimated and used to simulate single voxel responses to point stimuli. Single voxel receptive fields are estimated by fitting two-dimensional Gaussian functions to simulated single voxel responses. In the second stage, new voxel-specific encoding models are estimated as follows: For each voxel, features are (i) learned from high-resolution natural image patches that are of the same size as its estimated receptive field, (ii) nonlinearly extracted from stimulus regions that are within its estimated receptive field and (iii) linearly regressed on its stimulus-evoked responses. Concretely, features are learned using sparse coding, nonlinearly extracted using convolution and compressive nonlinearity, and linearly regressed using ridge regression. Note that different features are learned for each voxel. We validate the two-stage approach by predicting single voxel responses to natural images and identifying natural images from stimulus-evoked multiple voxel responses. We show that encoding and decoding performance of the voxel-specific encoding models is significantly higher than that of the general encoding model. These results demonstrate that the two-stage approach improves modeling of single voxel responses to natural images in functional magnetic resonance imaging.
Keywords:
deep learning,
functional magnetic resonance imaging,
natural image statistics,
Neural coding,
Unsupervised Feature Learning
Conference:
Neuroinformatics 2014, Leiden, Netherlands, 25 Aug - 27 Aug, 2014.
Presentation Type:
Poster, not to be considered for oral presentation
Topic:
Neuroimaging
Citation:
Güçlü
U,
Knechten
M and
Van Gerven
M
(2014). A two-stage approach to estimating voxel-specific encoding models improves prediction of hemodynamic responses to natural images.
Front. Neuroinform.
Conference Abstract:
Neuroinformatics 2014.
doi: 10.3389/conf.fninf.2014.18.00049
Copyright:
The abstracts in this collection have not been subject to any Frontiers peer review or checks, and are not endorsed by Frontiers.
They are made available through the Frontiers publishing platform as a service to conference organizers and presenters.
The copyright in the individual abstracts is owned by the author of each abstract or his/her employer unless otherwise stated.
Each abstract, as well as the collection of abstracts, are published under a Creative Commons CC-BY 4.0 (attribution) licence (https://creativecommons.org/licenses/by/4.0/) and may thus be reproduced, translated, adapted and be the subject of derivative works provided the authors and Frontiers are attributed.
For Frontiers’ terms and conditions please see https://www.frontiersin.org/legal/terms-and-conditions.
Received:
27 Apr 2014;
Published Online:
04 Jun 2014.
*
Correspondence:
Mr. Umut Güçlü, Radboud University Nijmegen, Donders Institute for Brain, Cognition and Behaviour, Nijmegen, Netherlands, umuguc@gmail.com