Sample Size Selection in Optimization Methods for Machine Learning

This paper presents a methodology for using varying sample sizes in batch-type optimization methods for large scale machine learning problems. The first part of the paper deals with the delicate issue of dynamic sample selection in the evaluation of the function and gradient. We propose a criterion for increasing the sample size based on variance … Read more

HOGWILD!: A Lock-Free Approach to Parallelizing Stochastic Gradient Descent

Stochastic Gradient Descent (SGD) is a popular algorithm that can achieve state-of-the-art performance on a variety of machine learning tasks. Several researchers have recently proposed schemes to parallelize SGD, but all require performance-destroying memory locking and synchronization. This work aims to show using novel theoretical analysis, algorithms, and implementation that SGD can be implemented *without … Read more

On the Use of Stochastic Hessian Information in Unconstrained Optimization

This paper describes how to incorporate stochastic curvature information in a Newton- CG method and in a limited memory quasi-Newton method for large scale optimization. The motivation for this work stems from statistical learning and stochastic optimization applications in which the objective function is the sum of a very large number of loss terms, and … Read more

An Improved Branch-and-Bound Method for Maximum Monomial Agreement

The NP-hard Maximum Monomial Agreement (MMA) problem consists of finding a single logical conjunction that best fits a weighted dataset of “positive” and “negative” binary vectors. Computing classifiers using boosting methods involves a maximum agreement subproblem at each iteration, although such subproblems are typically solved by heuristic methods. Here, we describe an exact branch and … Read more

Machine Learning for Global Optimization

In this paper we introduce the LeGO (Learning for Global Optimization) approach for global optimization in which machine learning is used to predict the outcome of a computationally expensive global optimization run, based upon a suitable training performed by standard runs of the same global optimization method. We propose to use a Support Vector Machine … Read more

Convergence and Convergence Rate of Stochastic Gradient Search in the Case of Multiple and Non-Isolated Extrema

The asymptotic behavior of stochastic gradient algorithms is studied. Relying on some results of differential geometry (Lojasiewicz gradient inequality), the almost sure point-convergence is demonstrated and relatively tight almost sure bounds on the convergence rate are derived. In sharp contrast to all existing result of this kind, the asymptotic results obtained here do not require … Read more

Convergence Rate of Stochastic Gradient Search in the Case of Multiple and Non-Isolated Minima

The convergence rate of stochastic gradient search is analyzed in this paper. Using arguments based on differential geometry and Lojasiewicz inequalities, tight bounds on the convergence rate of general stochastic gradient algorithms are derived. As opposed to the existing results, the results presented in this paper allow the objective function to have multiple, non-isolated minima, … Read more

Automated Tuning of Optimization Software Parameters

We present a method to tune software parameters using ideas from software testing and machine learning. The method is based on the key observation that for many classes of instances, the software shows improved performance if a few critical parameters have “good” values, although which parameters are critical depends on the class of instances. Our … Read more

Enclosing Machine Learning

This report introduces a new machine learning paradigm called enclosing machine learning for data mining. This novel method utilizes the virtues of human being’s cognition process and tries to imitate the two basic principles of cognition process from a macroscopical view, which are cognizing things of the same kind, recognizing things of a new kind … Read more