Research map / Continual learning / Online and replay-based
Transfer without Forgetting
Matteo Boschini, Lorenzo Bonicelli, Angelo Porrello, Giovanni Bellitto, Matteo Pennisi, Simone Palazzo, Concetto Spampinato, Simone Calderara
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 |
|---|---|---|---|
| add L_FP feature-preservation loss on top of L_CL (ablation) | TwF · Split CIFAR-100 (pretrain Tiny ImageNet), Class-IL, buffer 500 | final average accuracy (%): 44.01 -> 55.50 | Improved |
| add L_FP on top of L_CL (ablation) | TwF · Split CIFAR-10 (pretrain CIFAR-100), Class-IL, buffer 500 | final average accuracy (%): 75.79 -> 83.29 | Improved |
| add L_FP on top of L_CL (ablation) | TwF · Split CUB-200 (pretrain ImageNet), Class-IL, buffer 400 | final average accuracy (%): 56.53 -> 59.06 | Improved |
| add L_AUX on top of L_CL and L_FP (ablation) | TwF · Split CIFAR-100 (pretrain Tiny ImageNet), Class-IL, buffer 500 | final average accuracy (%): 55.50 -> 56.83 | Improved |
| add L_AUX on top of L_CL and L_FP (ablation) | TwF · Split CIFAR-10 (pretrain CIFAR-100), Class-IL, buffer 5120 | final average accuracy (%): 89.53 -> 89.55 | No clear effect |
| add L_CL (replay classification loss) to L_FP (ablation) | TwF · Split CIFAR-10 (pretrain CIFAR-100), Class-IL, buffer 500 | final average accuracy (%): 62.63 -> 83.65 | Improved |
| add replay buffer to TwF with only L_FP and L_AUX (ablation) | TwF · Split CUB-200 (pretrain ImageNet), Class-IL, buffer 1000 | final average accuracy (%): 38.43 -> 59.67 | Improved |
| use TwF feature preservation instead of oEwC with no replay (ablation) | TwF vs oEwC · Split CIFAR-100 (pretrain Tiny ImageNet), Class-IL | final average accuracy (%): 10.95 -> 49.74 | Improved |
| replace DER++ and iCaRL baselines by TwF under dissimilar pretraining (SVHN) | TwF vs DER++ · Split CIFAR-100 (pretrain SVHN), Class-IL, buffer 500 | final average accuracy (%): 36.46 -> 43.56 | Improved |
| add EwC regularizer on pretraining weights to DER++ | DER++ · Split CIFAR-100 (pretrain Tiny ImageNet) | final average accuracy (%) | Improved |
| dedicate part of the buffer to pretraining images in DER++ | DER++ · Split CIFAR-100 (pretrain Tiny ImageNet), buffer 2000 | final average accuracy (%) | Improved |
| use CO2L instead of ER on Split CUB-200 | CO2L · Split CUB-200 (pretrain ImageNet), Class-IL, buffer 400 | final average accuracy (%): 45.82 -> 8.96 | 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 7 of 7 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
- Detect, Decide, Unlearn: A Transfer-Aware Framework for Continual Learning (ICLR 2026 · similarity 0.90)
- Is Forgetting Less a Good Inductive Bias for Forward Transfer? (ICLR 2023 · similarity 0.90)
- Disentangling and mitigating the impact of task similarity for continual learning (NeurIPS 2024 · similarity 0.88)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.88)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.88)
- Learning to Prompt Knowledge Transfer for Open-World Continual Learning (AAAI 2024 · 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.
