Research map / Continual learning / Online and replay-based
Meta-Consolidation for Continual Learning
Joseph K J, Vineeth N Balasubramanian
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 |
|---|---|---|---|
| train VAE meta-distribution and consolidate (vs skipping) (ablation) | MERLIN · Split CIFAR-100 | average accuracy: 15.7 -> 43.5 | Improved |
| train VAE meta-distribution and consolidate (vs skipping) (ablation) | MERLIN · Permuted MNIST | average accuracy: 29.9 -> 85.5 | Improved |
| ensemble 30 sampled models instead of 1 (ablation) | MERLIN · Split Mini-ImageNet | average accuracy: 30.1 -> 39.1 | Improved |
| ensemble 100 instead of 30 models (ablation) | MERLIN · Split MNIST | average accuracy: 90.6 -> 91.6 | No clear effect |
| increase chunk size from 300 to 2000 (ablation) | MERLIN · Split Mini-ImageNet | average accuracy: 40.5 -> 37.8 | Hurt |
| task-aware instead of task-agnostic inference (ablation) | MERLIN · Split MNIST | average accuracy: 90.7 -> 97.4 | Improved |
| task-aware instead of task-agnostic inference (ablation) | MERLIN · Split CIFAR-10 | average accuracy: 82.9 -> 82.3 | No clear effect |
| increase exemplar memory from 100 to 2000 (ablation) | MERLIN · Split CIFAR-10 | average accuracy: 77.9 -> 88.4 | Improved |
| add auxiliary classification loss to ELBO (ablation) | MERLIN · Split Mini-ImageNet | average accuracy: 40.1 -> 40.8 | No clear effect |
| add auxiliary classification loss to ELBO (ablation) | MERLIN · Split CIFAR-100 | average accuracy: 43.5 -> 42.5 | Mixed |
| use MERLIN meta-consolidation instead of GEM/iCaRL/GSS (ablation) | MERLIN · Split CIFAR-100 | average accuracy: 40.6 -> 43.5 | Improved |
| MERLIN vs iCaRL on forgetting | MERLIN · Split MNIST | forgetting: 1.7 -> 6.4 | Hurt |
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.
- iCaRL — iCaRL: Incremental Classifier and Representation Learning (CVPR 2017) · better on 4 of 4 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
- Learning to Continually Learn with the Bayesian Principle (ICML 2024 · similarity 0.87)
- Reinforced Continual Learning (NeurIPS 2018 · similarity 0.87)
- Meta Continual Learning Revisited: Implicitly Enhancing Online Hessian Approximation via Variance Reduction (ICLR 2024 · similarity 0.86)
- Recall-Oriented Continual Learning with Generative Adversarial Meta-Model (AAAI 2024 · similarity 0.85)
- Improving Continual Learning Performance and Efficiency with Auxiliary Classifiers (ICML 2025 · similarity 0.85)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · 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.
