Research map / Continual learning / Class-incremental
AdaPrior: Bayesian-Inspired Adaptive Prior Correction for Long-Tailed Continual Learning
S Divakar Bhat, Amit Popat More, Mudit Soni, Bhuvan Aggarwal
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add post-hoc AdaPrior correction after training with AdaPrior Loss (Full AdaPrior) (ablation) | AdaPrior Loss · CIFAR100-LT shuffled LFH 5 tasks | average incremental accuracy: 40.97 -> 43.31 | Improved |
| add AdaPrior Loss to LUCIR+CE (ablation) | LUCIR+CE · ImageNet-subset-LT shuffled LFH 10 tasks | average incremental accuracy: 48.98 -> 58.21 | Improved |
| add post-hoc correction on top of AdaPrior Loss in ordered ImageNet-subset LFH (ablation) | AdaPrior Loss · ImageNet-subset ordered LFH 10 tasks | average incremental accuracy: 61.34 -> 61.23 | No clear effect |
| estimate priors dynamically (AdaPrior post-hoc) instead of static class frequency (logN) (ablation) | LUCIR+CE · CIFAR100-LT shuffled 5 tasks | average incremental accuracy: 42.37 (logN) vs 42.08 (AdaPrior post-hoc) | No clear effect |
| estimate priors dynamically (AdaPrior post-hoc) instead of static class frequency (logN) (ablation) | LUCIR+CE · ImageNet-subset-LT shuffled 10 tasks | average incremental accuracy: 55.03 -> 56.27 | Improved |
| remove mixup from the training pipeline (ablation) | Full AdaPrior · CIFAR100-LT shuffled 5 tasks | average incremental accuracy: 43.31 -> 41.83 | Hurt |
| add AdaPrior post-hoc calibration | LUCIR+CE · ImageNet-subset-LT shuffled 5 tasks | mean ECE: 0.1500 -> 0.1322 | Improved |
| set EMA momentum gamma near 0.05 (ablation) | AdaPrior · hyperparameter study | accuracy | Improved |
| raise post-hoc correction weight alpha to about 0.6-0.7 (ablation) | AdaPrior post-hoc · hyperparameter study | accuracy | Mixed |
| apply post-hoc AdaPrior to PODNet backbone without tuning | PODNet · CIFAR100-LT and ImageNet-LT | accuracy: +1-3% | Improved |
| reduce exemplars per class from 20 to 5 (ablation) | AdaPrior · ImageNet-subset-LT LFH | accuracy | No clear effect |
Closest papers by meaning
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.89)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.88)
- SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning (ICLR 2025 · similarity 0.88)
- Adaptive Retention & Correction: Test-Time Training for Continual Learning (ICLR 2025 · similarity 0.88)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.87)
- SAFE: Slow and Fast Parameter-Efficient Tuning for Continual Learning with Pre-Trained Models (NeurIPS 2024 · similarity 0.87)
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.
