Research map / Continual learning / Online and replay-based
CODE-CL: Conceptor-Based Gradient Projection for Deep Continual Learning
Marco P. E. Apolinario, Sakshi Choudhary, Kaushik Roy
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 |
|---|---|---|---|
| free K shared directions (K=80) vs none (K=0) (ablation) | CODE-CL · 5-Datasets | ACC (%): 91.67 -> 93.32 | Improved |
| free K shared directions (K=80) vs none (K=0) (ablation) | CODE-CL · 5-Datasets | BWT (%): -1.36 -> -0.25 | Improved |
| increase free dimensions K beyond 20 (ablation) | CODE-CL · Split CIFAR100, Split miniImageNet | ACC | Improved |
| increase aperture alpha (ablation) | CODE-CL · Split CIFAR100, Split miniImageNet | BWT | Improved |
| increase aperture alpha from 4 to 16 (ablation) | CODE-CL · 5-Datasets | ACC (%): 93.32 -> 93.46 | No clear effect |
| raise threshold epsilon from 0.5 to 0.8 (ablation) | CODE-CL · Split CIFAR100 | ACC (%): 77.21 -> 75.71 | Hurt |
| lower threshold epsilon from 0.5 to 0.2 (ablation) | CODE-CL · Split MiniImageNet | ACC (%): 68.83 -> 68.61 | No clear effect |
| conceptor-based pseudo-orthogonal projection with selective release of shared directions vs SGP | Split CIFAR100 · Split CIFAR100 | relative FWT: 4.74 -> 5.92 | Improved |
| CODE-CL vs SGP on overlapping classes | OL-CIFAR100 · OL-CIFAR100 | ACC (%): 75.00 -> 76.89 | Improved |
| CODE-CL vs TRGP and CUBER on memory and time | Split CIFAR100 · Split CIFAR100 | GPU memory and runtime | Improved |
| CODE-CL vs HAT/TRGP on forgetting | Split CIFAR100 · Split CIFAR100 | BWT (%): -0.1 -> -1.1 | No clear effect |
Later papers that compared with this method
1 later accepted paper reports CODE-CL as a baseline in their results tables; 0 of them report a better number on most of the results they share with it.
- LANCE: Low Rank Activation Compression for Efficient On-Device Continual Learning (ECCV 2026 · better on 0 of 2 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TRGP — TRGP: Trust Region Gradient Projection for Continual Learning (ICLR 2022) · better on 3 of 3 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
- Prompt Gradient Projection for Continual Learning (ICLR 2024 · similarity 0.90)
- CLR: Channel-wise Lightweight Reprogramming for Continual Learning (ICCV 2023 · similarity 0.90)
- Visual Prompt Tuning in Null Space for Continual Learning (NeurIPS 2024 · similarity 0.89)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.89)
- BiLoRA: Almost-Orthogonal Parameter Spaces for Continual Learning (CVPR 2025 · similarity 0.88)
- Continuous Subspace Optimization for Continual Learning (NeurIPS 2025 · 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.
