Top-label calibration (Spotlight #7)

We study the problem of post-hoc calibration for multiclass classification, with an emphasis on histogram binning. Multiple works have focused on calibration with respect to the confidence of just the predicted class (or 'top-label'). We find that the popular notion of confidence calibration [Guo et al., 2017] is not sufficiently strong -- there exist predictors that are not calibrated in any meaningful way but are perfectly confidence calibrated. We propose a closely related (but subtly different) notion, top-label calibration, that accurately captures the intuition and simplicity of confidence calibration, but addresses its drawbacks. We formalize a histogram binning (HB) algorithm that reduces top-label multiclass calibration to the binary case, prove that it has clean theoretical guarantees without distributional assumptions, and perform a methodical study of its practical performance. Some prediction tasks require stricter notions of multiclass calibration such as class-wise or canonical calibration. We formalize appropriate HB algorithms corresponding to each of these goals. In experiments with deep neural nets, we find that our principled versions of HB are often better than temperature scaling, for both top-label and class-wise calibration. Code for this work will be made publicly available at https://github.com/aigen/df-posthoc-calibration.