Research map / Continual learning / Class-incremental
Dense Network Expansion for Class Incremental Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 dense cross-task attention through TAB (ablation) | DNE · CIFAR100, Ns=10 | LA: 58.22 -> 68.04 | Helped |
| add dense cross-task attention through TAB | DNE · ImageNet100, Ns=10 | LA: +1.9% | Helped |
| increase heads per task from 1 to 4 (ablation) | DNE · CIFAR100, Ns=10 | LA: increases by 2% | Helped |
| increase heads per task from 1 to 4 (ablation) | DNE · ImageNet100, Ns=10 | LA: 1.28% | Helped |
| increase heads per task from 1 to 4 (ablation) | DNE · CIFAR100, Ns=25 | LA: 67.61 -> 71.47 | Helped |
| add cross-task attention to shallow layers (ablation) | DNE · CIFAR100, Ns=10 | LA: about 2% | Helped |
| add cross-task attention to deep layers (ablation) | DNE · CIFAR100, Ns=10 | LA: about 1% | Helped |
| use cross-task attention in MLP rather than in MHSA (ablation) | DNE · CIFAR100, Ns=10, k=1 | LA: 66.29 -> 68.04 | Helped |
| add cross-task attention in both MHSA and MLP instead of MLP alone (ablation) | DNE · CIFAR100, Ns=10, k=1 | LA: 68.04 -> 67.93 | Hurt |
| apply cross-task attention in both MLP linear layers instead of only one (ablation) | DNE · CIFAR100, Ns=10, k=1 | LA: 66.82 -> 68.04 | Helped |
| make Wq and Wk flexible across tasks, retaining shared Wv (ablation) | DNE TAB · CIFAR100, Ns=10, k=1 | LA: 68.04 -> 64.92 | Hurt |
| share Wv across tasks, retaining shared Wq and Wk (ablation) | DNE TAB · CIFAR100, Ns=10, k=1 | LA: 68.04 -> 64.44 | Hurt |
Closest papers by meaning
- Harnessing Neural Unit Dynamics for Effective and Scalable Class-Incremental Learning (ICML 2024 · similarity 0.86)
- Task-Agnostic Guided Feature Expansion for Class-Incremental Learning (CVPR 2025 · similarity 0.86)
- eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI 2024 · similarity 0.85)
- MIND: Multi-Task Incremental Network Distillation (AAAI 2024 · similarity 0.85)
- Class Incremental Learning via Likelihood Ratio Based Task Prediction (ICLR 2024 · similarity 0.85)
- Replay-and-Forget-Free Graph Class-Incremental Learning: A Task Profiling and Prompting Approach (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.