My work consists in using machine learning methods towards forward and reverse modeling of fMRI brain activity following sensory stimulation.
The main approaches I take to create predictive models from and to BOLD fMRI brain imaging data lie in regularized empirical risk minimization methods, often with non-smooth convex regularizers, which lead to convex optimization problems for which iterative algorithms can be devised.
Recently I have had success in forward modelling brain activity from features extracted from convolutional networks. This project was one of the reasons to create sklearn-theano, an open-source software package which makes the use of powerful convolutional nets very easy. It has also sparked my general interest in this type of learning architecture.