Research
In my research, I've mainly been working in one area: the application of pattern recognition algorithms to ''read minds'' from functional MRI. My focus has been on dimensionality reduction methods and brain-mapping through the use of the ''searchlight'' technique.
Also, I like to see if computers can beat humans at games of chance. If you're feeling lucky, see if you can beat me in Rock, Paper, Scissors.
Dimensionality Reduction
Analyzing fMRI data is hard. To acquire an fMRI dataset, powerful electromagnets bombard your brain with energy and record the changes in signal resulting from changes in bloodflow through the brain's intricate network of tiny blood vessels. It records this measurement in slices; a typical slice consists of 64 x 64 pixels. A typical brain has somewhere between 30-40 slices, so the total dataset has tens of thousands of volume pixels, called voxels. Most pattern recognition algorithms balk at datasets with so many dimensions; what's more, typical fMRI datasets involve only a few hundred or thousand observations of each voxel at most.
In my undergraduate thesis at Princeton, I proposed a new method of unsupervised dimensionality reduction of fMRI data using a latent variable model. Using this model, we can replace the tens of thousands of voxels with a set of 40-60 ''neural components;'' we do this by using probabilistic inference to estimate values for the latent variables and parameters of the model. I'm currently working on better ways of fitting the model to data by applying the technique of variational inference.
You can download a PDF of my thesis here.
MVPA Mapping with Searchlights
In conventional brain mapping, analyses are performed on a per-voxel basis. However, we are interested in mapping qualities of the patterns of activity across the brain, which we can use to discriminate cognitive variables and predict a subject's behavior. (This approach is known as multi-voxel pattern analysis, or MVPA). We can locate the regions of the brain that contain classifiable information by exhaustively searching all possible spherical regions of neighboring voxels; this is like sliding around a ''searchlight,'' illuminating all possible brain regions and looking for multivariate discriminatibility.
At the Computational Memory Lab, I work as a developer for the Princeton Multi-Voxel Pattern Analysis Toolbox, an open-source Matlab suite that facilitates MVPA and searchlight analyses.