We provide a new information-theoretic generalization error bound that is exactly tight (i.e., matching even the constant) for the canonical quadratic Gaussian (location) problem. Most existing bounds are order-wise loose in this setting, which has raised concerns about the fundamental capability of information-theoretic bounds in reasoning the generalization behavior for machine learning. The proposed new bound adopts the individual-sample-based approach proposed by Bu et al., but also has several key new ingredients. Firstly, instead of applying the change of measure inequality on the loss function, we apply it to the generalization error function itself; secondly, the bound is derived in a conditional manner; lastly, a reference distribution is introduced. The combination of these components produces a KL-divergence-based generalization error bound. We show that although the latter two new ingredients can help make the bound exactly tight, removing them does not significantly degrade the bound, leading to an asymptotically tight mutual-information-based bound. We further consider the vector Gaussian setting, where a direct application of the proposed bound again does not lead to tight bounds except in special cases. A refined bound is then proposed for decomposable loss functions, leading to a tight bound for the vector setting.
翻译:暂无翻译