Date of Award
Doctor of Philosophy (PhD)
Large scale machine learning requires tradeoffs. Commonly this tradeoff has led practitioners to choose simpler, less powerful models, e.g. linear models, in order to process more training examples in a limited time. In this work, we introduce parallelism to the training of non-linear models by leveraging a different tradeoff--approximation. We demonstrate various techniques by which non-linear models can be made amenable to larger data sets and significantly more training parallelism by strategically introducing approximation in certain optimization steps.
For gradient boosted regression tree ensembles, we replace precise selection of tree splits with a coarse-grained, approximate split selection, yielding both faster sequential training and a significant increase in parallelism, in the distributed setting in particular. For metric learning with nearest neighbor classification, rather than explicitly train a neighborhood structure we leverage the implicit neighborhood structure induced by task-specific random forest classifiers, yielding a highly parallel method for metric learning. For support vector machines, we follow existing work to learn a reduced basis set with extremely high parallelism, particularly on GPUs, via existing linear algebra libraries.
We believe these optimization tradeoffs are widely applicable wherever machine learning is put in practice in large scale settings. By carefully introducing approximation, we also introduce significantly higher parallelism and consequently can process more training examples for more iterations than competing exact methods. While seemingly learning the model with less precision, this tradeoff often yields noticeably higher accuracy under a restricted training time budget.
Roger Chamberlain, Robert Pless