<ul><li>Study conducted on online and transductive online learning using Empirical Risk Minimization (ERM) and weak consistency oracles on instance subsets.</li><li>Comparison made with standard online models where learner has complete knowledge of the concept class. ERM oracle minimizes loss on a subset, weak consistency oracle provides binary signal for subset realizability.</li><li>Evaluation based on mistakes and oracle calls. Lower bounds in realizable and agnostic cases established for standard online setting with ERM access.</li><li>Transductive online model explored with known instance sequence but sequentially revealed labels. Optimal mistake bounds for general Littlestone classes achieved with weak consistency oracle calls.</li></ul>

Tradeoffs between Mistakes and ERM Oracle Calls in Online and Transductive Online Learning

Discover more