On the Sample Complexity of Learning from a Sequence of Experiments
We analyze the sample complexity of a new problem: learning from a sequence of experiments. In this problem, the learner should choose a hypothesis that performs well with respect to an infinite sequence of experiments, and their related data distributions. In practice, the learner can only perform m experiments with a total of N samples drawn from those data distributions. By using a Rademacher complexity approach, we show that the gap between the training and generation error is O((m/N)^0.5). We also provide some examples for linear prediction, two-layer neural networks and kernel methods.
READ FULL TEXT