Research map / Continual learning / Class-incremental
Task-Adaptive Saliency Guidance for Exemplar-free 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 TASS, a plug-and-play saliency supervision framework (boundary-guided mid-level saliency drift regularization, low-level auxiliary saliency/boundary distillation, and saliency noise injection) that reduces attention drift and improves exemplar-free class incremental learning on CIFAR-100, Tiny-ImageNet, and ImageNet-Subset.
Problem
Exemplar-free class incremental learning suffers catastrophic forgetting because the model's saliency/attention drifts toward new-task features and background regions when no past exemplars can be stored.
Approach
Guide the network's attention with task-agnostic low-level saliency and boundary maps from a stationary CSNet teacher, supervise Grad-CAM saliency to avoid dilated object boundaries at three backbone stages, and inject elliptical saliency noise that the network learns to denoise, so saliency stays stable across tasks without exemplars.
Key results
- On CIFAR-100 with 10 tasks, SSRE+TASS reaches 57.93% last accuracy vs 55.01% for SSRE (about 3% gain), and TASS raises MUC from 15.47 to 30.13 on Tiny-ImageNet 10 tasks.
- TASS applied to PASS improves CIFAR-100 10-task accuracy from 49.0 to 54.5 in the ablation (5.5 points).
- Ablation shows average IoU between DINO attention and student saliency rises from 50.1 to 82.3 on task 4 with SSRE+TASS versus SSRE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add TASS to MUC | MUC · CIFAR-100, 5 tasks | last accuracy: 38.45 -> 49.17 (+10.72) | Helped |
| Add TASS to MUC | MUC · Tiny-ImageNet, 20 tasks | top-1 accuracy: 9.14 -> 27.70 (+18.56) | Helped |
| Add TASS to IL2A | IL2A · CIFAR-100, 10 tasks | last accuracy: 45.32 -> 53.24 (+7.92) | Helped |
| Add TASS to PASS | PASS · Tiny-ImageNet, 20 tasks | top-1 accuracy: 32.78 -> 42.57 (+9.79) | Helped |
| Add TASS to SSRE | SSRE · CIFAR-100, 10 tasks | last accuracy: 55.01 -> 57.93 (+2.92) | Helped |
| Add low-level multi-task supervision (ablation) | PASS + TASS · CIFAR-100, 10 tasks | top-1 accuracy: 49.0 -> 51.2 | Helped |
| Add dilated boundary supervision after low-level multi-task supervision (ablation) | PASS + TASS · CIFAR-100, 10 tasks | top-1 accuracy: 51.2 -> 53.0 | Helped |
| Add saliency noise injection after low-level and boundary supervision (ablation) | PASS + TASS · CIFAR-100, 10 tasks | top-1 accuracy: 53.0 -> 54.5 | Helped |
| Add low-level multi-task supervision (ablation) | SSRE + TASS · CIFAR-100, 10 tasks | top-1 accuracy: 55.0 -> 56.2 | Helped |
| Add dilated boundary supervision after low-level multi-task supervision (ablation) | SSRE + TASS · CIFAR-100, 10 tasks | top-1 accuracy: 56.2 -> 57.3 | Helped |
| Add saliency noise injection after low-level and boundary supervision (ablation) | SSRE + TASS · CIFAR-100, 10 tasks | top-1 accuracy: 57.3 -> 57.9 | Helped |
| Add TASS to PRAKA | PRAKA · CIFAR-100, 10 tasks | last accuracy: 59.20 -> 60.04 | Helped |
Surprising findings
- Simply distilling attention between tasks, as in prior saliency-distillation work, was reported to fail to offer plasticity and to be susceptible to attention forgetting.
Limitations
- Uses auxiliary saliency knowledge from a pretrained CSNet, which the authors note may raise concerns about unfair comparison with pure EFCIL.
- Gains depend on the base EFCIL method; the work relies on Grad-CAM and a fixed low-level teacher.
Experimental setup
Datasets: CIFAR-100 Tiny-ImageNet ImageNet-Subset ImageNet-Full
Benchmarks: CIFAR-100 Tiny-ImageNet ImageNet-Subset ImageNet-Full
Metrics: average top-1 accuracy last top-1 accuracy average forgetting IoU
Baselines: SSRE PASS IL2A MUC EWC LwF-MC iCaRL EEIL LUCIR RRR FeTrIL PRAKA
Builds on: SSRE PASS Grad-CAM LwM RRR
Closest papers by meaning
- Task-recency bias strikes back: Adapting covariances in Exemplar-Free Class Incremental Learning (NeurIPS 2024 · similarity 0.87)
- Elastic Feature Consolidation For Cold Start Exemplar-Free Incremental Learning (ICLR 2024 · similarity 0.86)
- Saliency-driven Experience Replay for Continual Learning (NeurIPS 2024 · similarity 0.85)
- Salient Frequency-aware Exemplar Compression for Resource-constrained Online Continual Learning (AAAI 2025 · similarity 0.85)
- SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation (ICLR 2024 · similarity 0.85)
- Advancing Cross-domain Discriminability in Continual Learning of Vision-Language Models (NeurIPS 2024 · similarity 0.84)
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.