Rajan Patel
Emory University
Prediction of dynamic experiences with neural networks and fMRI
Methods | Poster | Slides
Rajan Patel
Project Abstract
We utilized feed forward neural networks and the preprocessed fMRI data provided by the Pittsburgh Brain Activity Interpretation Competition committee to predict groups of feature vectors for each subject. We grouped features into several sets and for each subject, trained one neural network to predict each set of features. We further preprocessed the fMRI data and for each set of features, functionally clustered intra-cranial voxels that exhibited high mutual information to the corresponding set of features. The centroid at each time point of each cluster of voxels served as the input for the neural network. Training, testing, and validation of each neural network were done on the combined data from the first two runs. The outputs from each neural network were smoothed using a moving average filter. Finally, we discuss how post-processing of the Instructions prediction vector allowed us to exceed a correlation of 0.99 with the Instructions feature for run 3.
Rajan Patel, PhD works at Google, Inc. as a search quality analyst and is an adjunct assistant professor of Biostatistics at Emory University. His main interest in the area of functional neuroimaging is the development of statistical methods to better understand functional connectivity of the brain.
Links of Interest:
