Research map / Continual learning / Online and replay-based
DiL: Discrete-anchored Representation Alignment for Semi-Supervised Continual Learning
Nanyi Wang, Chaojie Chen, Zuoqi Tang, Jinxiang Lai, Xingcai Wu, Qi Wang
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 |
|---|---|---|---|
| replace direct full-prototype distillation (DPD) with channel-selective DCD (ablation) | iCaRL&Fix · CIFAR10-30 (5 tasks) | Last accuracy: 64.78 -> 68.2 | Improved |
| add confusion-matrix-guided block assignment (Mconf) in CACE (ablation) | iCaRL&Fix+DCD+CACE · CIFAR10-30 (5 tasks) | Last accuracy: 69.26 -> 71.55 | Improved |
| add CACE to DCD (ablation) | iCaRL&Fix · CIFAR10-30 (5 tasks) | Last accuracy: 68.2 -> 69.26 | Improved |
| add DCD distillation on top of CACE+Mconf (ablation) | iCaRL&Fix · CIFAR10-30 (5 tasks) | Last accuracy: 64.96 -> 71.55 | Improved |
| use DCD instead of feature KD or logit KD (ablation) | DiL · CIFAR10-30 (5 tasks) | accuracy on unlabeled data | Improved |
| replace CACE with random or clustering-based anchor initialization, or remove CACE (ablation) | DiL · CIFAR100-20 (10 tasks) | base-class accuracy and center drift | Hurt |
| use DiL instead of USP with small replay buffer 250 (ablation) | iCaRL&Fix · CIFAR10-30 (5 tasks) | Last accuracy: 59.93 -> 65.06 | Improved |
| use DiL instead of USP with 20K buffer on iCaRL&Fix, CIFAR100-125 | iCaRL&Fix · CIFAR100-125, buffer 20K | Last accuracy: 55.17 -> 54.04 | Hurt |
| use DiL instead of USP with 20K buffer on ImageNet100-100 (iCaRL&Fix) | iCaRL&Fix · ImageNet100-100, buffer 20K | Last accuracy: 51.73 -> 68.06 | Improved |
| use DiL instead of USP with DER&Fix on CIFAR10-30 | DER&Fix · CIFAR10-30 (5 tasks) | Last accuracy: 73.65 -> 73.32 | No clear effect |
| use DiL instead of USP on 11-task CUB-200 | UaD-CIE · CUB-200 11-task few-shot | Avg base-novel accuracy gap: above 10% -> 3.18% | Improved |
Closest papers by meaning
- Dynamic Sub-graph Distillation for Robust Semi-supervised Continual Learning (AAAI 2024 · similarity 0.89)
- Persistence Homology Distillation for Semi-supervised Continual Learning (NeurIPS 2024 · similarity 0.87)
- Continual Learning on a Diet: Learning from Sparsely Labeled Streams Under Constrained Computation (ICLR 2024 · similarity 0.87)
- Non-exemplar Online Class-Incremental Continual Learning via Dual-Prototype Self-Augment and Refinement (AAAI 2024 · similarity 0.86)
- Provable Contrastive Continual Learning (ICML 2024 · similarity 0.85)
- Cut out and Replay: A Simple yet Versatile Strategy for Multi-Label Online Continual Learning (ICML 2025 · 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.
