2014 Theses Doctoral

# Selected machine learning reductions

Machine learning is a field of science aiming to extract knowledge from the data. Optimization lies in the core of machine learning as many learning problems are formulated as optimization problems, where the goal is to minimize/maximize an objective function. More complex machine learning problems are then often solved by reducing them to simpler sub-problems solvable by known optimization techniques. This dissertation addresses two elements of the machine learning system 'pipeline', designing efficient basic optimization tools tailored to solve specific learning problems, or in other words optimize a specific objective function, and creating more elaborate learning tools with sub-blocks being essentially optimization solvers equipped with such basic optimization tools. In the first part of this thesis we focus on a very specific learning problem where the objective function, either convex or non-convex, involves the minimization of the partition function, the normalizer of a distribution, as is the case in conditional random fields (CRFs) or log-linear models. Our work proposes a tight quadratic bound on the partition function which parameters are easily recovered by a simple algorithm that we propose. The bound gives rise to the family of new optimization learning algorithms, based on bound majorization (we developed batch, both full-rank and low-rank, and semi-stochastic variants), with linear convergence rate that successfully compete with state-of-the-art techniques (among them gradient descent methods, Newton and quasi-Newton methods like L-BFGS, etc.). The only constraint we introduce is on the number of classes which is assumed to be finite and enumerable. The bound majorization method we develop is simultaneously the first reduction scheme discussed in this thesis, where throughout this thesis by 'reduction' we understand the learning approach or algorithmic technique converting a complex machine learning problem into a set of simpler problems (that can be as small as a single problem). Secondly, we focus on developing two more sophisticated machine learning tools, for solving harder learning problems. The tools that we develop are built from basic optimization sub-blocks tailored to solve simpler optimization sub-problems. We first focus on the multi class classification problem where the number of classes is very large. We reduce this problem to a set of simpler sub-problems that we solve using basic optimization methods performing additive update on the parameter vector. Secondly we address the problem of learning data representation when the data is unlabeled for any classification task. We reduce this problem to a set of simpler sub-problems that we solve using basic optimization methods, however this time the parameter vector is updated multiplicatively. In both problems we assume that the data come in a stream that can even be infinite. We will now provide more specific description of each of these problems and describe our approach for solving them. In the multi class classification problem it is desirable to achieve train and test running times which are logarithmic in the label complexity. The existing approaches to this problem are either intractable or do not adapt well to the data. We propose a reduction of this problem to a set of binary regression problems organized in a tree structure and introduce a new splitting criterion (objective function) allowing gradient descent style optimization (bound optimization methods can also be used). A decision tree algorithm that we obtain differs from traditional decision trees in the objective optimized, and in how that optimization is done. The different objective has useful properties, such us it guarantees balanced and small-error splits, while the optimization uses an online learning algorithm that is queried and trained simultaneously as we pass over the data. Furthermore, we prove an upper-bound on the number of splits required to reduce the entropy of the tree leafs below small threshold. We empirically show that the trees we obtain have logarithmic depth, which implies logarithmic training and testing running times, and significantly smaller error than random trees. Finally, we consider the problem of unsupervised (clustering) learning of data representation, where the quality of obtained clustering is measured using a very simple, intuitive and widely cited clustering objective, k-means clustering objective. We introduce a family of online clustering algorithms by extending algorithms for online supervised learning, with access to expert predictors (which are basic sub-blocks of our learning system), to the unsupervised learning setting. The parameter vector corresponds to the probability distribution over the experts. Different update rules for the parameter vector depend on an approximation to the current value of the k-means clustering objective obtained by each expert, and model different levels of non-stationarity in the data. We show that when the experts are batch clustering algorithms with approximation guarantees with respect to the k-means clustering objective, applied to a sliding window of the data stream, our algorithms obtain approximation guarantees with respect to the k-means clustering objective. Thus simultaneously we address an open problem posed by Dasgupta for approximating k-means clustering objective on data streams. We experimentally show that our algorithms' empirical performance tracks that of the best clustering algorithm in its experts set and that our algorithms outperform widely used online algorithms.

## Subjects

## Files

- Choromanska_columbia_0054D_11872.pdf application/doc 2.64 MB Download File

## More About This Work

- Academic Units
- Electrical Engineering
- Thesis Advisors
- Jebara, Tony
- Degree
- Ph.D., Columbia University
- Published Here
- April 11, 2014