EAGER: New Algorithms for Feature-Efficient Learning

点击翻译按钮获取中文摘要

The main goal of this exploratory research project is to invent theoretically sound and practical machine-learning algorithms designed to perform well under various limitations involving access to data features during deployment. This tackles a major difficulty encountered in many machine-learning applications: in running an algorithm, accessing features of the data can be time consuming or costly. For example, in medical diagnosis, features of patients may correspond to results of medical tests, which can take significant time to run, carry enormous cost, and even impose heath risks. Current machine-learning techniques are ill-equipped to tackle such impediments. This project involves approaches that incorporate feature-efficient optimization into the training phase of machine-learning algorithms and also the creation of new frameworks for reducing both error rates and costs associated with acquiring features. Successful developments in feature-efficient algorithms create an important advance for application areas ranging from medical diagnosis to query-answering on the World Wide Web. Additional facets of this project include incorporating its research findings into graduate courses and broadening participation in research.This project investigates new models for jointly optimizing feature costs, prediction time, and classification error rates, to create feature-efficient predictors. Techniques for this exploratory project include solving original optimization problems, creating novel machine-learning reductions, and analyzing the problem via statistical query oracles. Another aspect of this work is to tackle a budgeted learning formalization by moving the feature-cost optimization into the training phase of budgeted boosting classifiers and support vector machines.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.