Methods Description v2.wxp

Project Abstract The prediction of relationships between a spatiotemporal fMRI brain activation signal and subjective human experience was carried out using three distinct methods: support vector machines (SVM), supervised principal components, and boosting principal components. The spatially normalized data in ANALYZE format are preprocessed by removing low frequency drift and high frequency by a band-pass Fast Fourier Transform filter. Subsequently, the three distinct methods are employed to generate predictions of Movie 1 features by using predictions and fMRI data from Movie 2 as a training dataset. The methods which predicted Movie 1 features most accurately for each feature were chosen to be utilized for the first submission. The effect of training on Movie 2 data alone or both Movie 1 and Movie 2 data together was assessed by training only on Movie 2 for the first submission and training on both Movies for the second. The third submission consisted of the better results for each feature and subject from the first two submissions. We found that none of the three methods were uniformly dominant across all subjects or all features. Each method predicted Movie 1 features very well (correlation > 0.5) for some features and poorly (correlation < 0.1) for others, however we are impressed with the information content of fMRI and each method's ability to extract that information. A novel approach to boosting supervised principal components and a genetic algorithm approach to optimizing support vector machine parameters are discussed in this paper.