*=Equal Contributors
We consider online learning problems in the realizable setting, where there is a zero-loss solution, and propose new Differentially Private (DP) algorithms that obtain near-optimal regret bounds. For the problem of online prediction from experts, we design new algorithms that obtain near-optimal regret where is the number of experts. This significantly improves over the best existing regret bounds for the DP non-realizable setting which are . We also develop an adaptive algorithm for the small-loss setting with regret where is the total loss of the best expert…Apple Machine Learning Research