Date

2019

Document Type

Dissertation

Degree

Doctor of Philosophy

Department

Industrial Engineering

First Adviser

Takac, Martin

Abstract

The rising amount of data has changed the classical approaches in statistical modeling significantly. Special methods are designed for inferring meaningful relationships and hidden patterns from these large datasets, which build the foundation of a study called Machine Learning (ML). Such ML techniques have already applied widely in various areas and achieved compelling success. In the meantime, the huge amount of data also requires a deep revolution of current techniques, like the availability of advanced data storage, new efficient large-scale algorithms, and their distributed/parallelized implementation.There is a broad class of ML methods can be interpreted as Empirical Risk Minimization (ERM) problems. When utilizing various loss functions and likely necessary regularization terms, one could approach their specific ML goals by solving ERMs as separable finite sum optimization problems. There are circumstances where the nonconvex component is introduced into the ERMs which usually makes the problems hard to optimize. Especially, in recent years, neural networks, a popular branch of ML, draw numerous attention from the community. Neural networks are powerful and highly flexible inspired by the structured functionality of the brain. Typically, neural networks could be treated as large-scale and highly nonconvex ERMs.While as nonconvex ERMs become more complex and larger in scales, optimization using stochastic gradient descent (SGD) type methods proceeds slowly regarding its convergence rate and incapability of being distributed efficiently. It motivates researchers to explore more advanced local optimization methods such as approximate-Newton/second-order methods.In this dissertation, first-order stochastic optimization for the regularized ERMs in Chapter1 is studied. Based on the development of stochastic dual coordinate accent (SDCA) method, a dual free SDCA with non-uniform mini-batch sampling strategy is investigated [30, 29]. We also introduce several efficient algorithms for training ERMs, including neural networks, using second-order optimization methods in a distributed environment. In Chapter 2, we propose a practical distributed implementation for Newton-CG methods. It makes training neural networks by second-order methods doable in the distributed environment [28]. In Chapter 3, we further build steps towards using second-order methods to train feed-forward neural networks with negative curvature direction utilization and momentum acceleration. In this Chapter, we also report numerical experiments for comparing second-order methods and first-order methods regarding training neural networks. The following Chapter 4 purpose an distributed accumulative sample-size second-order methods for solving large-scale convex ERMs and nonconvex neural networks [35]. In Chapter 5, a python library named UCLibrary is briefly introduced for solving unconstrained optimization problems. This dissertation is all concluded in the last Chapter 6.

Share

COinS