Understanding Generalization via Leave-One-Out Conditional Mutual Information
This work addresses the fundamental problem of generalization theory for researchers in statistical learning, offering incremental theoretical insights with practical implications for algorithm analysis.
The paper tackles the problem of understanding generalization in machine learning by introducing leave-one-out conditional mutual information (CMI) variants, which control mean generalization error for bounded loss functions and provide explicit connections to classical leave-one-out error estimates, leading to upper and lower bounds on risk that converge within a factor of two for certain cases. As a result, they apply this to analyze the one-inclusion graph algorithm, matching optimal bounds for VC classes in the realizable setting and addressing an open challenge.
We study the mutual information between (certain summaries of) the output of a learning algorithm and its $n$ training data, conditional on a supersample of $n+1$ i.i.d. data from which the training data is chosen at random without replacement. These leave-one-out variants of the conditional mutual information (CMI) of an algorithm (Steinke and Zakynthinou, 2020) are also seen to control the mean generalization error of learning algorithms with bounded loss functions. For learning algorithms achieving zero empirical risk under 0-1 loss (i.e., interpolating algorithms), we provide an explicit connection between leave-one-out CMI and the classical leave-one-out error estimate of the risk. Using this connection, we obtain upper and lower bounds on risk in terms of the (evaluated) leave-one-out CMI. When the limiting risk is constant or decays polynomially, the bounds converge to within a constant factor of two. As an application, we analyze the population risk of the one-inclusion graph algorithm, a general-purpose transductive learning algorithm for VC classes in the realizable setting. Using leave-one-out CMI, we match the optimal bound for learning VC classes in the realizable setting, answering an open challenge raised by Steinke and Zakynthinou (2020). Finally, in order to understand the role of leave-one-out CMI in studying generalization, we place leave-one-out CMI in a hierarchy of measures, with a novel unconditional mutual information at the root. For 0-1 loss and interpolating learning algorithms, this mutual information is observed to be precisely the risk.