RI: Small: From a Machine Detector to a Machine Detective: Decisions and Queries with Uncertain and Incomplete Information

点击翻译按钮获取中文摘要

Much of machine learning is devoted to unrealistic, sterile settings where all relevant information for a prediction task has been pre-recorded and fed to one’s algorithm. This is incongruous with many real-world problems where predictions and decisions must be made with incomplete and muddied information. Moreover, the current machine learning paradigm is unprepared for the automated future where algorithmic agents are not just passively given information, but instead are able to actively obtain information from their environment as they make decisions. This project will extend past the current paradigm to develop a ‘machine detective’, a system that is capable of reasoning with incomplete instances and interacting with the environment to obtain new information, or new clues, on-the-fly as it is making decisions. The project shall not only increase machines’ predictive abilities in domains like health-care where recorded data is abounding with missing values, but shall also enable more efficient and robust predictions in interactive domains like computerized adaptive testing for education assessment and customer-service/troubleshooting chatbots where the machine can interact with a user (or the environment) to obtain new and relevant information.The work stemming from this award will serve as the underpinnings to intelligent agents that reason robustly about their decision-making process and weigh the cost of an incorrect prediction (e.g., a false positive or false negative) and the cost (e.g., in time, risk, or money) of obtaining additional information. This extends the typical paradigm in machine learning to give machines the ability to sequentially query for unobserved features to make more certain predictions through the following aims. First, the project develops methods that may infer with partially observed instances through novel generative models that learn conditional dependencies among features. Second, the learned dependencies are used as a foundation to learn non-greedy policies for acquiring informative unobserved features of a particular instance on hand. Lastly, the project builds models that may acquire data over a spatial-temporal domain, where features are indexed by spatial coordinates (e.g., when collecting information in the field), or are indexed by time (e.g., when collecting from sensors at specific moments).This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.