Princeton University

School of Engineering & Applied Science

The Application of Optimization Techniques in Machine Learning

Haotiang Pang
Engineering Quadrangle B327
Thursday, August 10, 2017 - 11:00am to 12:30pm

The past decades have witnessed significantly progress in machine learning, and solving these problems requires the advancing in optimization techniques. High dimensional sparse learning has imposed a great computational challenge to large-scale data analysis. There are serious drawbacks to the existing methods for solving problems with a regularization parameter, as tuning the parameter for the desired solution is very inefficient. The parametric simplex method is applied to solve a broad class of sparse learning approaches. It uses the unknown weighting factor as the parameter and provides a powerful and efficient way to address these shortcomings. It has been demonstrated that this algorithm generates a pair of sparse dual and primal solution, both theoretically and in practice.
In addition, a convex optimization method named Inexact Peaceman-Rachford Splitting Method (IPRSM) is covered. It solves a convex minimization problem with linear constraints and a separable objective function. The inexact criteria used for solving the sub-problems are most popular in the literature and easily implementable.
Finally, a graph estimation method called the new graph Perceptron algorithm that performs on online binary classification problems is proposed. It is a new kernel based, mistake-driven online learning algorithm derived from online class action and extends to online graph estimation.