To better understand why machine learning works, we cast learning problems as searches and characterize what makes searches successful. We prove that any search algorithm can only perform well on a narrow subset of problems, and show the effects of dependence on raising the probability of success for searches. We examine two popular ways of understanding what makes machine learning work, empirical risk minimization and compression, and show how they fit within our search framework. Leveraging the "dependence-first" view of learning, we apply this knowledge to areas of unsupervised time series segmentation and automated hyperparameter optimization, developing new algorithms with strong empirical performance on real-world problem classes.
Cosma R. Shalizi (Chair)
Milos Hauskrecht (University of Pittsburgh)