Research map / Continual learning / Class-incremental
T-CIL: Temperature Scaling using Adversarial Perturbation for Calibration in Class-Incremental Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes T-CIL, a post-hoc temperature scaling method for class-incremental learning that calibrates old and new task confidence using adversarially perturbed memory exemplars, with perturbation magnitude set on a new-task validation set only.
Problem
Post-hoc confidence calibration in class-incremental learning is hard because memory limits prevent keeping a validation set of old-task data, and old-task exemplars are already overfit from training.
Approach
Adversarially perturb memory exemplars with a single magnitude (found by binary search matching the new-task optimal temperature) but direct perturbations toward the closest class for old-task exemplars and the farthest class for new-task exemplars, so the perturbed set yields a temperature that calibrates both old and new tasks.
Key results
- On CIFAR-100 with ER, T-CIL reduces ECE from 16.28 (best TS baseline) to 5.74, a 65.2% reduction in calibration error versus the best post-hoc baseline.
- Combined with DER, T-CIL reduces ECE by 80.5% on CIFAR-100 (22.38 to 4.37) with accuracy essentially unchanged (69.98%).
- Accuracy on CIFAR-100 is unchanged by T-CIL (56.25% vs 56.25% for TS), showing post-hoc calibration does not reduce accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use T-CIL post-hoc calibration instead of vanilla ER calibration | ER · CIFAR-10, CIFAR-100, and Tiny-ImageNet | ECE (%): 28.16± 0.33 -> 17.70± 2.60; 27.66± 0.29 -> 5.74±0.53; 32.56± 0.36 -> 8.12± 0.38 | Helped |
| Use T-CIL post-hoc calibration instead of vanilla calibration | EEIL · CIFAR-10, CIFAR-100, and Tiny-ImageNet | ECE (%): 15.48± 0.75 -> 10.49± 2.34; 21.96± 0.29 -> 10.30±1.10; 29.69± 0.36 -> 15.58± 0.76 | Helped |
| Use T-CIL post-hoc calibration instead of vanilla calibration | WA · CIFAR-10, CIFAR-100, and Tiny-ImageNet | ECE (%): 19.10± 0.50 -> 15.61± 0.23; 8.89± 0.64 -> 3.87± 0.52; 10.97± 0.41 -> 11.43± 1.00 | Mixed |
| Use T-CIL post-hoc calibration instead of vanilla calibration | DER · CIFAR-10, CIFAR-100, and Tiny-ImageNet | ECE (%): 21.81± 0.46 -> 12.70± 1.35; 22.38± 0.37 -> 4.37± 0.59; 39.00± 1.72 -> 6.91± 0.86 | Helped |
| Replace T-CIL target selection with random target classes (ablation) | T-CIL with ER · CIFAR-100 | ECE (%): 5.74±0.53 -> 15.09± 0.54 | Hurt |
| Select the closest target class for both old- and new-task exemplars (ablation) | T-CIL with ER · CIFAR-100 | ECE (%): 5.74±0.53 -> 16.39± 0.73 | Hurt |
| Select the farthest target class for both old- and new-task exemplars (ablation) | T-CIL with ER · CIFAR-100 | ECE (%): 5.74±0.53 -> 14.52± 0.45 | Hurt |
Surprising findings
- Using memory exemplars directly for temperature optimization yields abnormally small temperatures because the model is near-perfectly accurate on its own training exemplars.
- DER achieves the best accuracy on CIFAR-100 and Tiny-ImageNet but has the highest calibration error among the compared CIL techniques.
Limitations
- Assumes the new-task validation set is available and that old tasks have lower accuracy than the new task
- Calibration gains are mixed for WA on Tiny-ImageNet, where T-CIL slightly increases ECE over vanilla WA
- Evaluated only on small image classification benchmarks with ResNet backbones
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Tiny-ImageNet
Benchmarks: CIFAR-10 CIFAR-100 Tiny-ImageNet
Metrics: top-1 accuracy ECE AECE
Baselines: Vanilla TS ETS IRM PerturbTS
Builds on: temperature scaling FGSM Experience Replay DER WA
Closest papers by meaning
- Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · similarity 0.88)
- Consistency-Guided Temperature Scaling Using Style and Content Information for Out-of-Domain Calibration (AAAI 2024 · similarity 0.87)
- On the Limitations of Temperature Scaling for Distributions with Overlaps (ICLR 2024 · similarity 0.86)
- On Temperature Scaling and Conformal Prediction of Deep Classifiers (ICML 2025 · similarity 0.85)
- Enhancing Robustness in Incremental Learning with Adversarial Training (AAAI 2025 · similarity 0.85)
- Divide et Calibra: Multiclass Local Calibration via Vector Quantization (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.