T-CIL: Temperature Scaling using Adversarial Perturbation for Calibration in Class-Incremental Learning

TL;DR

This work tackles the mismatch between model confidence and actual accuracy in class-incremental learning (CIL) by introducing T-CIL, a post-hoc calibration method that does not require old-task validation data. T-CIL uses adversarial perturbations of memory exemplars with a task-aware direction policy and a magnitude tuned from the new-task validation set to optimize temperature scaling across both old and new tasks. The method integrates with existing CIL techniques (e.g., ER, EEIL, WA, DER) and consistently achieves substantial reductions in calibration error with minimal impact on accuracy. By enabling reliable confidence estimates under memory constraints, T-CIL enhances the practicality and safety of deployed continual learning systems.

Abstract

We study model confidence calibration in class-incremental learning, where models learn from sequential tasks with different class sets. While existing works primarily focus on accuracy, maintaining calibrated confidence has been largely overlooked. Unfortunately, most post-hoc calibration techniques are not designed to work with the limited memories of old-task data typical in class-incremental learning, as retaining a sufficient validation set would be impractical. Thus, we propose T-CIL, a novel temperature scaling approach for class-incremental learning without a validation set for old tasks, that leverages adversarially perturbed exemplars from memory. Directly using exemplars is inadequate for temperature optimization, since they are already used for training. The key idea of T-CIL is to perturb exemplars more strongly for old tasks than for the new task by adjusting the perturbation direction based on feature distance, with the single magnitude determined using the new-task validation set. This strategy makes the perturbation magnitude computed from the new task also applicable to old tasks, leveraging the tendency that the accuracy of old tasks is lower than that of the new task. We empirically show that T-CIL significantly outperforms various baselines in terms of calibration on real datasets and can be integrated with existing class-incremental learning techniques with minimal impact on accuracy.

Figures (9)

• Figure 1: Overview of our proposed framework, T-CIL, a post-hoc calibration framework for class-incremental learning when a validation set only from the new task is available with memory. Our method leverages exemplars from memory by applying adversarial perturbations, whose direction and magnitude are determined based on feature distance and new task validation data, respectively. The temperature is then optimized using these perturbed exemplars.
• Figure 2: Temperatures after training each task using ER method on the CIFAR-100 (left), and Tiny-ImageNet (right) datasets. For each task, we show two different temperatures: (1) optimized on a validation set containing only new-task data, and (2) the optimal temperature obtained on the test set across all tasks.
• Figure 3: Task-wise accuracy comparison after completing training on the final (10th) task of CIFAR-100 dataset using Experience Replay (ER) (left), and Dynamically Expandable Representation (DER) yan2021dynamically (right).
• Figure 4: A perturbation direction policy of T-CIL visualized on a two-dimensional feature space with decision boundaries. For old-task data, the target class is selected based on the closest distance in feature space, while for the new-task data, the farthest class is selected based on maximum distance in the feature space.
• Figure 5: The optimal perturbation magnitude differences ($\Delta\epsilon^*$) on the CIFAR-100 test sets after training each task, comparing the gap between old and new tasks versus new tasks only. Results are presented for the perturbation direction policies of T-CIL and three alternative policies.