Calibration under continual task shifts

Establish calibration methods that explicitly capture the heterogeneous, class-incremental-learning-specific calibration behavior induced by continual task shifts, which existing static-setting calibration approaches do not adequately address.

Background

Confidence calibration methods such as temperature scaling and its extensions have predominantly been developed for static, fixed-distribution classification. In class-incremental learning, the data and label distributions evolve as new tasks are introduced, while catastrophic forgetting causes confidence and accuracy to change at different rates across tasks. Consequently, a single global temperature, and static-setting sample-dependent calibration methods, may fail to account for the resulting task-specific calibration heterogeneity. The paper identifies this gap as unresolved and proposes MT-CC as a solution for the studied settings, but the broader problem of calibration under continual task shifts remains explicitly identified as unresolved.