Extracting task-relevant low dimensional representations under data sparsity
Seyedmehdi Orouji, Megan Peters, University of California Irvine, United States
Posters 2 Poster
Pacific Ballroom H-O
Fri, 26 Aug, 19:30 - 21:30 Pacific Time (UTC -7)
Functional magnetic resonance imaging (fMRI) is used to study higher order visual and cognitive processing in humans. However, fMRI data is noisy and contains task-irrelevant information, which hamstrings our capacity to identify meaningful representations. One possible avenue is to find cleaner, lower dimensional representations that reflect more task-relevant information. However, fMRI datasets typically have many more features (voxels) than samples available, so state of the art machine learning models often overfit. To remedy this, we propose the task-relevant autoencoder via classifier enhancement (TRACE). We benchmarked TRACE on the MNIST dataset and verified that it can find features that are more task-relevant compared to a standard autoencoder (AE). Critically, we then tested TRACE’s performance under increasing sparsity of samples. We found that TRACE’s performance is superior both quantitatively and qualitatively in comparison to AE, and that TRACE is more robust than AE under extreme data paucity. These results suggest TRACE as a viable method for extracting ‘cleaner’ low-dimensional representations from fMRI data.