A Taylor polynomial expansion line search for large-scale optimization

Abstract

In trying to cope with the Big Data deluge, the landscape of distributed computing has changed. Large commodity hardware clusters, typically operating in some form of MapReduce framework, are becoming prevalent for organizations that require both tremendous storage capacity and fault tolerance. However, the high cost of communication can dominate the computation time in large-scale optimization routines in these frameworks. This thesis considers the problem of how to efficiently conduct univariate line searches in commodity clusters in the context of gradient-based batch optimization algorithms, like the staple limited-memory BFGS (LBFGS) method. In it, a new line search technique is proposed for cases where the underlying objective function is analytic, as in logistic regression and low rank matrix factorization. The technique approximates the objective function by a truncated Taylor polynomial along a fixed search direction. The coefficients of this polynomial may be computed efficiently in parallel with far less communication than needed to transmit the high-dimensional gradient vector, after which the polynomial may be minimized with high accuracy in a neighbourhood of the expansion point without distributed operations. This Polynomial Expansion Line Search (PELS) may be invoked iteratively until the expansion point and minimum are sufficiently accurate, and can provide substantial savings in time and communication costs when multiple iterations in the line search procedure are required.

Three applications of the PELS technique are presented herein for important classes of analytic functions: (i) logistic regression (LR), (ii) low-rank matrix factorization (MF) models, and (iii) the feedforward multilayer perceptron (MLP). In addition, for LR and MF, implementations of PELS in the Apache Spark framework for fault-tolerant cluster computing are provided. These implementations conferred significant convergence enhancements to their respective algorithms, and will be of interest to Spark and Hadoop practitioners. For instance, the Spark PELS technique reduced the number of iterations and time required by LBFGS to reach terminal training accuracies for LR models by factors of 1.8--2. Substantial acceleration was also observed for the Nonlinear Conjugate Gradient algorithm for MLP models, which is an interesting case for future study in optimization for neural networks. The PELS technique is applicable to a broad class of models for Big Data processing and large-scale optimization, and can be a useful component of batch optimization routines.