Record Details
Integrating learning and search for structured prediction
ScholarsArchive at Oregon State University
| Field | Value |
|---|---|
| Title | Integrating learning and search for structured prediction |
| Names |
Doppa, Janardhan Rao
(creator) Tadepalli, Prasad (advisor) Fern, Alan (advisor) |
| Date Issued | 2014-07-17 (iso8601) |
| Note | Graduation date: 2015 |
| Abstract | We are witnessing the rise of the data-driven science paradigm, in which massive amounts of data - much of it collected as a side-effect of ordinary human activity - can be analyzed to make sense of the data and to make useful predictions. To fully realize the promise of this paradigm, we need automated systems that can transform structured inputs to structured outputs. Examples include parsing a sentence, resolving coreferences of entity and event mentions in a piece of text, interpreting a visual scene, and translating from one language to another. Problems such as these are often referred to as structured prediction problems in the machine learning community. These prediction problems pose severe learning and inference challenges due to the huge number of possible outputs. This thesis explores how to integrate two fundamental branches of Artificial Intelligence, namely learning and search, to solve structured prediction tasks. We study a new framework for structured prediction called HC-Search, where we formulate the problem of structured prediction as an explicit search process in the combinatorial space of outputs. The system starts from a reasonably good initial solution and performs an heuristic search guided by a learned heuristic function H until a fixed number of alternative solutions has been generated or a fixed time limit is reached. It then evaluates each of these alternatives using a learned cost function C and returns the minimum-cost solution. There are three key learning challenges in this framework - Search space design: how can we automatically design an efficient search space over structured outputs?; Heuristic learning: how can we learn a heuristic function H for effectively guiding the search?; Cost function learning: how can we learn a cost function C that can accurately select the best output among the candidate outputs? We develop generic solutions for each of these learning challenges and an engineering methodology for applying this framework. We show that the HC-Search framework achieves results in a wide range of structured prediction problems that significantly exceed the best previous results. |
| Genre | Thesis/Dissertation |
| Topic | Structured Prediction |
| Identifier | http://hdl.handle.net/1957/50908 |
