Research map / Continual learning / Online and replay-based
Growing a Brain with Sparsity-Inducing Generation for Continual Learning
Hyundong Jin, Gyeong-hyeon Kim, Chanho Ahn, Eunwoo Kim
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 |
|---|---|---|---|
| transform old parameters with hypernetwork-generated sparse parameters instead of keeping them fixed (ablation) | GrowBrain · ImageNet+FG class-incremental | accuracy: 45.3 -> 60.1 | Improved |
| transform old parameters with hypernetwork-generated sparse parameters instead of keeping them fixed (ablation) | GrowBrain · UCF-101 class-incremental | accuracy: 40.4 -> 64.7 | Improved |
| add sparsity-inducing generation (lambda_sp > 0) (ablation) | GrowBrain · ImageNet+FG class-incremental | accuracy: 58.1 -> 60.1 | Improved |
| add sparsity-inducing generation (lambda_sp > 0) (ablation) | GrowBrain · UCF-101 class-incremental | accuracy: 64.0 -> 64.7 | Improved |
| add channel-adaptive segments and matrix decomposition (scalability) (ablation) | GrowBrain · ImageNet+FG class-incremental | number of layer embeddings: 3262 -> 133 | Improved |
| add channel-adaptive segments and matrix decomposition (scalability) (ablation) | GrowBrain · ImageNet+FG class-incremental | accuracy: 59.8 -> 60.1 | No clear effect |
| raise lambda_sp from 0.1 to 1.0 (ablation) | GrowBrain · ImageNet+FG class-incremental | accuracy and parameters: 15.7M -> 9.4M params, +0.7% accuracy | Improved |
| use parameter isolation with transformation instead of regularization methods (LwF, EWC-On) | GrowBrain · ImageNet+FG class-incremental | accuracy: gaps of 48.8% and 31.7% vs EWC-On and LwF | Improved |
| use GrowBrain instead of LwF on ActivityNet class-incremental T=5 | GrowBrain · ActivityNet class-incremental (T=5, Avg) | accuracy: 42.8 vs 44.1 (LwF) | Hurt |
| use all previous parameter sets (PackNet) on dissimilar domains | PackNet · ImageNet+FG class-incremental | accuracy: drops of 12.2% and 9.0% after WikiArt and Sketch | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- iCaRL — iCaRL: Incremental Classifier and Representation Learning (CVPR 2017) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Continual learning with hypernetworks (ICLR 2020 · similarity 0.90)
- A continual learning survey: Defying forgetting in classification tasks (TPAMI 2021 · similarity 0.90)
- Recall-Oriented Continual Learning with Generative Adversarial Meta-Model (AAAI 2024 · similarity 0.88)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.88)
- Task-Free Continual Generation and Representation Learning via Dynamic Expansionable Memory Cluster (AAAI 2024 · similarity 0.87)
- Task-Free Dynamic Sparse Vision Transformer for Continual Learning (AAAI 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.
