Classifying under computational resource constraints: anytime classification using probabilistic estimators

摘要: In many online applications of machine learning, the computational resources available for classification will vary from time to time. Most techniques are designed operate within constraints minimum expected and fail utilize further when they available. We propose a novel anytime algorithm, averaged probabilistic estimators (AAPE), which is capable delivering strong prediction accuracy with little CPU utilizing additional increase accuracy. The idea run an ordered sequence very efficient Bayesian (single improvement steps) until runs out. Theoretical studies empirical validations reveal that by properly identifying, ordering, invoking ensembling single steps, AAPE able accomplish accurate whenever it interrupted. It also output class probability estimates beyond simple 0/1-loss classifications, as well adeptly handle incremental learning.