Optimization Algorithms for Machine Learning Designed for Parallel and Distributed Environments

This thesis proposes several optimization methods that utilize parallel algorithms for large-scale machine learning problems. The overall theme is network-based machine learning algorithms; in particular, we consider two machine learning models: graphical models and neural networks. Graphical models are methods categorized under unsupervised machine learning, aiming at recovering conditional dependencies among random variables from observed samples of a multivariable distribution. Neural networks, on the other hand, are methods that learn an implicit approximation to underlying true nonlinear functions based on sample data and utilize that information to generalize to validation data. The goal of finding the best methods relies on an optimization problem tasked with training such models. Improvements in current methods of solving the optimization problem for graphical models are obtained by parallelization and the use of a new update and a new step-size selection rule in the coordinate descent algorithms designed for large-scale problems. For training deep neural networks, we consider the second-order optimization algorithms within trust-region-like optimization frameworks. Deep networks are represented using large-scale vectors of weights and are trained based on very large datasets. Hence, obtaining second-order information is very expensive for these networks. In this thesis, we undertake an extensive exploration of algorithms that use a small number of curvature evaluations and are hence faster than other existing methods.