Research map / Continual learning / Class-incremental
A Theoretical Bridge Between Long-Tailed Recognition and Continual Learning
Mahdiyar Molahasani, Michael Greenspan, Ali Etemad
Read the paperCheck your idea against 137,217 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.
Theoretically links long-tailed recognition to continual learning by showing long-tailed training drives parameters into an O(1/sqrt(IF)) neighborhood of the dominant-class solution.
Problem
Understanding how training on long-tailed class distributions relates to continual learning, which has not been previously connected to long-tailed recognition theoretically.
Approach
Shows that training on a long-tailed dataset drives model parameters into an O(1/sqrt(IF)) neighborhood of the dominant-class solution under uniform and exponentially decaying cardinality distributions, and builds on this to bridge LTR and CL.
Key results
- Long-tailed training drives parameters into an O(1/sqrt(IF)) neighborhood of the dominant-class solution under both uniform and exponentially decaying cardinality distributions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| learn Head then Tail sequentially with SGP instead of joint training | ResNet-32 baseline · CIFAR100-LT IF=100 | accuracy (%): 38.3 -> 50.7 | Improved |
| use a stronger CL method (SGP) instead of LwF in CLTR (ablation) | CLTR · CIFAR100-LT IF=100 | accuracy (%): 45.1 -> 50.7 | Improved |
| use TPL as CL method in CLTR instead of LwF (ablation) | CLTR · CIFAR10-LT IF=100 | accuracy (%): 76.3 -> 84.7 | Improved |
| use TPL as CL method in CLTR instead of EWC (ablation) | CLTR · ImageNet-LT | top-1 accuracy (%): 48.9 -> 53.9 | Improved |
| apply CLTR (SGP) versus weight-balancing WD+Max | CLTR · CIFAR100-LT IF=100 | accuracy (%): 53.4 -> 50.7 | Hurt |
| use DualPrompt CL on CLIP instead of full fine-tuning | CLIP ViT-B/32 · CIFAR100-LT IF=100 | accuracy (%): 73.4 -> 79.2 | Improved |
| use CL method instead of naive sequential training (ablation) | CLTR (DualPrompt) · CIFAR100-LT IF=100 (CLIP) | accuracy (%): 48.0 -> 79.2 | Improved |
| add CL regularization to Head-then-Tail training (ablation) | CLTR (LwF/EWC) · CIFAR100-LT | upper-bound inequality | Improved |
| use more partitions (N=4) with stronger CL methods, N=2 with weaker | CLTR · LTR benchmarks | accuracy (%) | No clear effect |
| prompt learning with CLIP instead of full fine-tuning | CLIP ViT-B/32 · CIFAR100-LT IF=100 | accuracy (%): 73.4 -> 60.4 | Hurt |
Closest papers by meaning
- TailCon: Mitigating Tail Signal Erosion through Memory Consolidation for Long-Tailed Recognition (NeurIPS 2026 · similarity 0.90)
- LoRanPAC: Low-rank Random Features and Pre-trained Models for Bridging Theory and Practice in Continual Learning (ICLR 2025 · similarity 0.89)
- Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition (NeurIPS 2024 · similarity 0.88)
- Forgetting, Ignorance or Myopia: Revisiting Key Challenges in Online Continual Learning (NeurIPS 2024 · similarity 0.88)
- DiffuLT: Diffusion for Long-tail Recognition Without External Knowledge (NeurIPS 2024 · similarity 0.87)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 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.
