Self-Certifying Classification by Linearized Deep Assignment
This work addresses the need for reliable risk certification in machine learning, particularly for graph-based classification, though it is incremental by building on existing PAC-Bayes frameworks.
The authors tackled the problem of certifying out-of-sample risk for deep stochastic classifiers on graph data, achieving tight risk bounds that align with empirical test errors and improve computational efficiency over prior methods.
We propose a novel class of deep stochastic predictors for classifying metric data on graphs within the PAC-Bayes risk certification paradigm. Classifiers are realized as linearly parametrized deep assignment flows with random initial conditions. Building on the recent PAC-Bayes literature and data-dependent priors, this approach enables (i) to use risk bounds as training objectives for learning posterior distributions on the hypothesis space and (ii) to compute tight out-of-sample risk certificates of randomized classifiers more efficiently than related work. Comparison with empirical test set errors illustrates the performance and practicality of this self-certifying classification method.