Research map / Continual learning / Online and replay-based
An Effective Dynamic Gradient Calibration Method for Continual Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Dynamic Gradient Calibration (DGC), an SVRG-style two-level gradient calibrator for continual learning that works with reservoir-sampling replay methods, and reports gains on split CIFAR/TinyImageNet benchmarks.
Problem
Catastrophic forgetting in continual learning under a limited replay buffer, where the buffer cannot hold all historical data so gradient estimates of the full historical objective are inaccurate.
Approach
Adapts stochastic variance reduction (SVRG) to continual learning by maintaining a recursively updated historical gradient term Γ' that approximates the full gradient over all past tasks without storing past data, so only one extra vector is stored; it plugs into ER, DER++, XDER and Dynamic ER.
Key results
- DGC-Dynamic ER reaches 89.90% FAIA on S-CIFAR10 with buffer 2000 versus 83.30% for Dynamic ER, an improvement above 6 points.
- DGC-DER++ improves over DER++ by 3.94% with 20 tasks on S-CIFAR100 versus 1.18% with 5 tasks.
- DGC-ER improves ER on S-CIFAR100 with buffer 2000 from 55.48% to 59.55% FAIA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DGC calibration to Dynamic ER | Dynamic ER · S-CIFAR10, buffer 2000 | FAIA (%): 83.30_{\pm0.93} -> 89.90_{\pm0.93} | Helped |
| Add DGC calibration to Dynamic ER | Dynamic ER · S-CIFAR100, buffer 2000 | FAIA (%): 64.57_{\pm2.02} -> 70.70_{\pm1.31} | Helped |
| Add DGC calibration to XDER | XDER · S-CIFAR100, buffer 500 | FAIA (%): 61.45_{\pm0.50} -> 62.70_{\pm0.44} | Helped |
| Add DGC calibration to XDER | XDER · S-CIFAR100, 20 tasks | FAIA (%): 60.45_{\pm0.46} -> 63.53_{\pm0.48} | Helped |
| Set alpha to 0 instead of the default 1e^{-3} (ablation) | DGC-ER · S-CIFAR100 | FAIA (%): 59.55_{\pm0.97} -> 55.48_{\pm1.52} | Hurt |
| Set alpha to 1e^{-2} instead of the default 1e^{-3} (ablation) | DGC-ER · S-CIFAR100 | FAIA (%): 59.55_{\pm0.97} -> 57.4_{\pm2.14} | Hurt |
| Set alpha to 1e^{-4} instead of the default 1e^{-3} (ablation) | DGC-ER · S-CIFAR100 | FAIA (%): 59.55_{\pm0.97} -> 59.14_{\pm0.91} | No clear effect |
| Set m to 1 instead of 200 (ablation) | DGC-ER · S-CIFAR100 | FAIA (%): 59.55_{\pm0.97} -> 49.11_{\pm5.30} | Hurt |
| Set m to 100 instead of 200 (ablation) | DGC-ER · S-CIFAR100 | FAIA (%): 59.55_{\pm0.97} -> 59.44_{\pm0.95} | No clear effect |
| Set m to 300 instead of 200 (ablation) | DGC-ER · S-CIFAR100 | FAIA (%): 59.55_{\pm0.97} -> 59.27_{\pm0.65} | No clear effect |
| Set m to 400 instead of 200 (ablation) | DGC-ER · S-CIFAR100 | FAIA (%): 59.55_{\pm0.97} -> 58.94_{\pm0.98} | No clear effect |
Surprising findings
- Plain ER outperforms DGC-ER at low memory footprints, with DGC's benefit appearing only as memory grows.
- DGC-XDER gains only 0.42% on S-CIFAR100 with 5 tasks, showing the benefit is small for some settings.
Limitations
- Calibration term relies on a stochastic approximation of the full historical gradient, so the unbiasedness holds only in expectation.
- Requires tuning of the mixing parameter alpha and the number of inner steps m; performance drops at low memory and when m=1.
- Theoretical guarantee assumes L-smooth and strongly convex losses, which deep networks do not satisfy.
Experimental setup
Datasets: S-CIFAR10 S-CIFAR100 S-TinyImageNet
Benchmarks: S-CIFAR10 S-CIFAR100 S-TinyImageNet
Metrics: FAIA FAA Final Forgetting AA_t in TFCL
Baselines: ER DER++ XDER Dynamic ER MOCA GSS GCR HAL ICARL AGEM AOP SSVRG
Builds on: SVRG experience replay DER++ XDER Dynamic ER streaming SVRG
Closest papers by meaning
- Learning to Continually Learn with the Bayesian Principle (ICML 2024 · similarity 0.90)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.89)
- Convergence and Implicit Bias of Gradient Descent on Continual Linear Classification (ICLR 2025 · similarity 0.89)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.89)
- LoRanPAC: Low-rank Random Features and Pre-trained Models for Bridging Theory and Practice in Continual Learning (ICLR 2025 · similarity 0.89)
- Optimal Rates in Continual Linear Regression via Increasing Regularization (NeurIPS 2025 · similarity 0.89)
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.