Biological vision
I am developing methods to characterize the response of visual cells from natural data. The Volterra model is a generic model that has been widely used to characterize visuals cells. I developed the Volterra Relevant Space Technique (VRST) to allow the estimation of high-dimensional Volterra models from natural data (paper, and code). This method uses a nice mathematical trick to allow the estimation of the large number of Volterra parameters.

The VRST assummes that the cell response depends on a low-dimensional subspace of the images. Several methods have been proposed to estimate this subspace (Spike Triggered Average, Spike Triggered Covariance, Maximally Informative Dimensions, ...), but none of these methods can exploit the full two-dimensional, and spatio-temporal, information in the inputs, or cannot be used with natural stimuli. So I developed the extended Projection Pursuit (ePPR) algorithm, that overcomes these limitations, and allows to estimate the two-dimensional and spatio-temporal image subspace on which the response of the cell depends. ePPR can use arbitrary, including natural, stimuli. In a collaboration with Jon Touryan and Gidon Felsen, we are using ePPR to characterize the responses of the simple and complex cells described here. The manuscript reporting our results is under review.

Brain Imaging
To understand how the brain works it is important to have good observations. During the first year of my PhD I worked at the Biomedical Imaging Research Lab looking at images of brain activity with Positron Emission Tomography (PET) and Magnetoencephalography (MEG). We used nice mathematical tools, and nice computing, to understand beautiful data. Our Internet2-based 3D PET Image Reconstruction using a PC cluster paper nicely shows this.

Information Retrieval
The phenomenon of the Internet has always fascinated me. My first experience analyzing data dealt with Internet data. I developed a technique that learned from a training collection of HTML documents the optimal way to combine several HTML ranking heuristics into a single similarity measure for HTML documents. This was my first conference paper.