Reliably estimating the uncertainty of a prediction throughout the model lifecycle is crucial in many safety-critical applications. The most common way to measure this uncertainty is via the predicted confidence. While this tends to work well for in-domain samples, these estimates are unreliable under domain drift. Alternatively, a bias-variance decomposition allows to directly measure the predictive uncertainty across the entire input space. But, such a decomposition for proper scores does not exist in current literature, and for exponential families it is convoluted. In this work, we introduce a general bias-variance decomposition for proper scores and reformulate the exponential family case, giving rise to the Bregman Information as the variance term in both cases. This allows us to prove that the Bregman Information for classification measures the uncertainty in the logit space. We showcase the practical relevance of this decomposition on two downstream tasks. First, we show how to construct confidence intervals for predictions on the instance-level based on the Bregman Information. Second, we demonstrate how different approximations of the instance-level Bregman Information allow reliable out-of-distribution detection for all degrees of domain drift.
翻译:对模型生命周期中预测的不确定性进行可靠估计对于许多安全关键应用而言至关重要。测量这种不确定性的最常见方法是预测信心。虽然这对内部样本来说效果良好,但这些估计在域流中是不可靠的。或者,偏差分解分解可以直接测量整个输入空间的预测不确定性。但是,当前文献中没有适当的分数的分数,指数式家庭则处于混杂状态。在这项工作中,我们引入了一种对正确分数的一般偏差分分分分法,并重新配置指数式家庭案例,从而在两种情况下都产生Bregman信息的差异术语。这使我们能够证明用于分类的Bregman信息测量了登录空间的不确定性。我们展示了该分数分数在下游两个任务上的实际相关性。首先,我们展示了如何根据Bregman信息在实例一级进行预测时建立信任间隔。第二,我们展示了实例级Bregman信息的不同近似点如何允许可靠地进行离差检测所有水平的流度。