CSAIL Event Calendar

I will describe a family of learning algorithms based on a new form of data-dependent regularization with emphasis on applications to semi-supervised learning.

While kernel-based learning algorithms, such as Support Vector Machines, show strong performance in fully supervised setting, they do not use the marginal distribution are thus unable to utilize unlabeled data. By using an additional regularizer, we obtain a new form of regularization directly reflecting the geometry of the marginal distribution. We show that the solution to the corresponding optimization problem admits an expansion in terms of kernel functions centered at both labeled and unlabeled points, leading to simple and efficient algorithms. Moreover, SVM and Regularized Least Squares, as well as several recently proposed graph-based transductive learning methods, can be obtained as special cases of our framework.

I will present some encouraging experimental results and discuss connections to clustering and supervised learning. This is joint work with Partha Niyogi and Vikas Sindhwani.