MIT EECS - Basis Technology Undergraduate Research and Innovation Scholar
Deep learning methods for bioinformatics
Tommi S. Jaakkola
Deep learning models have recently been used for various auditory and visual tasks, giving state of the art results. Their success is primary based on how they learn to sparsely represent the data. Although several other sparse coding methods have already been used in biomedical applications, the adaptation of deep learning methods is still lacking. Our contributions include developing architectures and algorithms for deep networks adapted to the specific problem of inferring phenotype from a given genotype. Specifically, the deep learning method combines the signals from hundreds of different SNPs correlated with the phenotype to predict the phenotype.
I worked on a boosting approach to matrix factorization with Prof. Tommi Jaakkola. Matrix factorization is a collaborative filtering method used for the application of recommendation systems. I worked on industrial level recommendations systems during my summer internship at foursquare, a service that provides personalized venue (eg. restaurants, parks) recommendations to users. I have also participated in the International Olympiad of Informatics in Bulgaria (2009) and Canada (2010).