Research map / Continual learning / Class-incremental
Long-Tail Class Incremental Learning via Independent Sub-prototype Construction
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 a learnable sub-prototype space (basis vectors shared across classes, with re-sampling) plus a reminiscence space (per-class Gaussian and correlation statistics for distillation) to tackle long-tail class incremental learning without replay.
Problem
Long-tail class incremental learning, where models must learn from imbalanced, evolving data streams without catastrophic forgetting, especially for minority classes with few samples.
Approach
Builds a learnable sub-prototype space whose basis vectors are shared across classes to integrate knowledge and re-sample features that mitigate intra-task imbalance, and a reminiscence space storing per-class feature distributions and correlation statistics to distill against forgetting, all without storing replay exemplars.
Key results
- On Shuffled LT-CIL CIFAR100 (rho=0.01, 5 tasks), the method reaches 40.2% average accuracy, about 3% above LUCIR+LWS (37.2%) without replay data.
- On Ordered LT-CIL CIFAR100 (rho=0.01, 5 tasks), the method reaches 44.8%, a 2.5% gain over LUCIR+LWS (42.3%).
- Ablation on CIFAR100 shows the sub-prototype and reminiscence constraints together raise accuracy from 34.7% to 40.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full two-parallel-space method instead of LUCIR+LWS | LT-CIL · CIFAR100, Shuffled LT-CIL, ρ = 0.01, 5 tasks | Average Accuracy: 37.2 -> 40.2 | Helped |
| Use the full two-parallel-space method instead of LUCIR+LWS | LT-CIL · ImageNet-Subset, Ordered LT-CIL, ρ = 0.01, 10 tasks | Average Accuracy: 49.6 -> 52.6 | Helped |
| Add sub-prototype construction loss Lcon (ablation) | Classification-loss-only model · CIFAR100, Shuffled LT-CIL, ρ = 0.01, 5 tasks | Accuracy: 11.3 -> 10.4 | Hurt |
| Add feature re-sampling loss Lf (ablation) | Lcls + Lcon model · CIFAR100, Shuffled LT-CIL, ρ = 0.01, 5 tasks | Accuracy: 10.4 -> 34.7 | Helped |
| Add sub-prototype distillation loss Ldis1 (ablation) | Lcls + Lcon + Lf model · CIFAR100, Shuffled LT-CIL, ρ = 0.01, 5 tasks | Accuracy: 34.7 -> 36.3 | Helped |
| Add classifier distillation loss Ldis2 (ablation) | Lcls + Lcon + Lf model · CIFAR100, Shuffled LT-CIL, ρ = 0.01, 5 tasks | Accuracy: 34.7 -> 38.9 | Helped |
| Add both reminiscence-space constraints Ldis1 and Ldis2 (ablation) | Lcls + Lcon + Lf model · CIFAR100, Shuffled LT-CIL, ρ = 0.01, 5 tasks | Accuracy: 34.7 -> 40.2 | Helped |
Surprising findings
- Constructing the sub-prototype space alone slightly lowers accuracy (11.3% to 10.4%) before re-sampling training recovers it.
Limitations
- Relies on hyperparameters (k, h, lambda values) whose sensitivity is only in the supplementary material.
- Evaluated only on CIFAR100 and ImageNet-Subset with ResNet18 backbone.
- Minority-class accuracy remains lower than BalPoE and LUCIR+LWS in task 0.
Experimental setup
Datasets: CIFAR100 ImageNet-Subset
Benchmarks: CIFAR100 ImageNet-Subset
Metrics: average accuracy
Baselines: LDAM BalPoE MDCS EWC LwF SDC PASS IL2A SAVC iCaRL TwF SCoMMER
Builds on: LUCIR SDC PASS IL2A
Closest papers by meaning
- Gradient Reweighting: Towards Imbalanced Class-Incremental Learning (CVPR 2024 · similarity 0.90)
- Defying Imbalanced Forgetting in Class Incremental Learning (AAAI 2024 · similarity 0.90)
- Make Domain Shift a Catastrophic Forgetting Alleviator in Class-Incremental Learning (AAAI 2025 · similarity 0.89)
- Knowledge Memorization and Rumination for Pre-trained Model-based Class-Incremental Learning (CVPR 2025 · similarity 0.89)
- Non-exemplar Online Class-Incremental Continual Learning via Dual-Prototype Self-Augment and Refinement (AAAI 2024 · similarity 0.88)
- Adaptive Discovering and Merging for Incremental Novel Class Discovery (AAAI 2024 · similarity 0.88)
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.