AbstractsComputer Science

Structured prediction describes problems which involve predicting multiple output variables with expressive and complex interdependencies and constraints. Learning over expressive structures (called structural learning) is usually time-consuming as exploring the structured space can be an intractable problem. The goal of this thesis is to present different techniques for structural learning, which learn by decomposing the problem space into simpler and tractable components. We will consider three different settings: fully supervised, unsupervised and semi-supervised, and discriminative latent variable setting, and present learning techniques for each case. For supervised structural learning, we describe a paradigm called Decomposed Learning (DecL) which decomposes the inference procedure during learning into small inference steps using additional application specific information. For unsupervised learning, we propose a family of Expectation Maximization [Dempster et al., 1977] algorithms called Unified Expectation Maximization (UEM) [Samdani et al., 2012a] that covers several seemingly divergent versions of EM e.g. hard EM. To efficiently add domain-specific declarative constraints into learning, we use a dual projected subgradient ascent algorithm which naturally decomposes the task into simpler components. In the discriminative latent variable scenario, we present a supervised latent variable model for clustering called the Latent Left-Linking Model (L3M) that can efficiently cluster items arriving in a streaming order. We decompose the learning process for L3M into small and efficient stochastic gradient descent steps that lead to rapid convergence.