More Than Meets the fMRI: Representational Similarities between Real and Artificially Generated fMRI Data
Pabitra Sharma, Sveekruth Sheshagiri Pai, Indian Institute of Science, Bangalore, India
Posters 2 Poster
Pacific Ballroom H-O
Fri, 26 Aug, 19:30 - 21:30 Pacific Time (UTC -8)
The Generative Adversarial Network (GAN) is a class of deep learning models used to create synthetic media such as the notorious "deepfakes" seen in recent years. We trained a GAN on a parcellated resting state fMRI data subset from the Human Connectome Project (HCP) to study how well certain functionally relevant features are preserved in their artificial counterpart. In doing so, we discovered anti-correlations between some of the brain regions that were not present in the original data, but have previously been reported in the literature. However, strong inter-hemispherical functional connectivity was preserved in both. Independent Component Analysis (ICA) revealed the presence of common resting state networks between the two. Of particular importance, we were able to recover the default mode network (DMN), a hallmark of resting state fMRI.