Modern deep classifiers trained by gradient descent typically need many passes over the data to become accurate, and even then they are often poorly calibrated. The standard remedy, post-hoc temperature scaling, adjusts calibration after training but cannot address the underlying issue, according to the paper's abstract.

The authors propose a variational filtering approach for Bayesian last layers, using an independent binary probit surrogate. This suggests a way to treat the final layer probabilistically during training rather than patching calibration afterward. The abstract does not yet include experimental results, so the reported benefits are limited to the problem statement and the proposed method.