30 likes | 126 Views
Learning Updatable Classifiers from Remote Data. Does your learning algorithm assume direct access to data? In practice: Data is too large to ship to a centralized location In certain domains, such as personalized medicine, privacy regulations may explicitly prohibit direct access to data.
E N D
Learning Updatable Classifiers from Remote Data • Does your learning algorithm assume direct access to data? • In practice: • Data is too large to ship to a centralized location • In certain domains, such as personalized medicine,privacy regulations may explicitly prohibit direct access to data. • Even in cases where data can be shipped to a centralized location, the local copy of the dataset may quickly become out of date due to frequent updates to the data. • Needed: Approaches for learning from data without direct access,in settings where the data can be accessed only through statistical queries
General Framework • Query Formulation:poses the relevant statistical queries • Hypothesis Generation: uses the resulting statistics to update or refine a partial model (and if necessary, further invoke the statistical query component)