Linear and Nonlinear Dimensionality Reduction in fMRI Data for Picture-Sentence Classification

fMRI data is represented in a space with very high dimensionality. Because of this, classifiers such as SVM and Naive Bayes may overfit this data. Dimensionality reduction methods are intended to extract features from data in a high dimensional space. Training a classifier on data in a lower dimension may improve the true error of the classifier beyond the performance obtained by training in a higher dimensional space. We experimented with the PCA[1], Isomap[2], and LLE[4] methods for extracting features from fMRI data of subjects viewing pictures and sentences. Additionally we experimented with an extension to the LLE algorithm, which improved classification.