Research map / Continual learning / Online and replay-based
Continual Learning with Scaled Gradient Projection
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 |
|---|---|---|---|
| Allow importance-scaled gradient updates along past-task gradient bases instead of blocking those directions | GPM on Split CIFAR-100 · Split CIFAR-100 | ACC (%): 72.48 $\pm$ 0.40 -> 76.05 $\pm$ 0.43 | Helped |
| Allow importance-scaled gradient updates along past-task gradient bases instead of blocking those directions | GPM on Split miniImageNet · Split miniImageNet | ACC (%): 60.41 $\pm$ 0.61 -> 62.83 $\pm$ 0.33 | Helped |
| Allow importance-scaled gradient updates along past-task gradient bases instead of blocking those directions | GPM on CIFAR-100 Superclass · CIFAR-100 Superclass | ACC (%): 57.72 -> 59.05 | Helped |
| Use a larger scale coefficient alpha (ablation) | SGP on CIFAR-100 Superclass · CIFAR-100 Superclass | ACC | Hurt |
| Use a larger scale coefficient alpha (ablation) | SGP on Split miniImageNet · Split miniImageNet | ACC | Hurt |
| Use a larger scale coefficient alpha (ablation) | SGP on Split CIFAR-100 · Split CIFAR-100 | ACC | Mixed |
| Apply gradient projection after Adam's adaptive update rather than passing projected gradients into Adam (ablation) | GPM and SGP on continual Atari learning · Six sequential Atari games | accumulated normalized reward | Helped |
| Use SGP instead of GPM | Continual learning across six Atari games · Six sequential Atari games | accumulated normalized reward: 12% | Helped |
| Use SGP instead of BLIP | Continual learning across six Atari games · Six sequential Atari games | accumulated normalized reward: 36% | Helped |
| Add importance storage for gradient bases | GPM memory on image-classification tasks · Image-classification benchmarks | extra memory: 0.1% | Hurt |
Closest papers by meaning
- An Effective Dynamic Gradient Calibration Method for Continual Learning (ICML 2024 · similarity 0.91)
- Addressing Loss of Plasticity and Catastrophic Forgetting in Continual Learning (ICLR 2024 · similarity 0.90)
- Convergence and Implicit Bias of Gradient Descent on Continual Linear Classification (ICLR 2025 · similarity 0.89)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.89)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.88)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · 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.