Research map / Continual learning / Online and replay-based
Doubly Perturbed Task Free 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 DPCL, a task-free continual learning method that upper-bounds the TF-CL loss with input and classifier-weight adversarial perturbations, implemented via perturbed function interpolation, branched stochastic classifiers, and perturbation-induced memory management and learning-rate scheduling.
Problem
Task-free online continual learning suffers from catastrophic forgetting and poor plasticity because naive rehearsal ignores unseen future samples; the paper seeks a principled way to account for them.
Approach
Derives an upper bound for TF-CL with unknown future samples involving adversarial input and classifier-weight perturbations, then approximates it cheaply: perturbed function interpolation (feature-space noise plus mixup) for input, and multiple stochastic classifiers with weight averaging and variational inference for weights, plus a mutual-information-based memory update and adaptive learning ra
Key results
- On CIFAR100 disjoint (M=2K), DPCL reaches 45.27% ACC and 37.66% FM versus 41.49% ACC for the best baseline ODDL.
- On ImageNet-100 disjoint, DPCL achieves 30.92% ACC versus 27.68% for DRO.
- Under i-blurry CIFAR100 setup, DPCL reaches 50.22% ACC versus 47.04% for CLIB; ablations show each of PFI, BSC and PIMA contributes to the gain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DPCL instead of the strongest baseline by ACC | CIFAR100 disjoint · CIFAR100 (M=2K), disjoint | ACC: 41.49$\pm$ 1.38 -> 45.27$\pm$ 1.32 | Helped |
| use DPCL instead of the strongest baseline by FM | CIFAR100 disjoint · CIFAR100 (M=2K), disjoint | FM: 39.90$\pm$ 0.39 -> 37.66$\pm$ 1.18 | Helped |
| use DPCL instead of the strongest baseline by ACC | CIFAR100-SC disjoint · CIFAR100-SC (M=2K), disjoint | ACC: 41.93$\pm$ 0.42 -> 45.39$\pm$ 1.34 | Helped |
| use DPCL instead of the strongest baseline by FM | CIFAR100-SC disjoint · CIFAR100-SC (M=2K), disjoint | FM: 27.76$\pm$ 0.77 -> 26.57$\pm$ 1.63 | Helped |
| use DPCL instead of the strongest baseline by ACC | ImageNet-100 disjoint · ImageNet-100 (M=2K), disjoint | ACC: 27.68$\pm$ 1.23 -> 30.92$\pm$ 1.17 | Helped |
| use DPCL instead of the strongest baseline by FM | ImageNet-100 disjoint · ImageNet-100 (M=2K), disjoint | FM: 39.96$\pm$ 0.87 -> 37.33$\pm$ 1.53 | Helped |
| use DPCL instead of CLIB | CIFAR100 blurry · CIFAR100 (M=2K), blurry | ACC: 38.13$\pm$ 0.73 -> 47.58$\pm$ 2.75 | Helped |
| use DPCL instead of CLIB | CIFAR100 blurry · CIFAR100 (M=2K), blurry | FM: 4.69$\pm$ 0.99 -> 11.44$\pm$ 2.64 | Hurt |
| use DPCL instead of CLIB | CIFAR100 i-blurry · CIFAR100 (M=2K), i-blurry | ACC: 47.04$\pm$ 0.89 -> 50.22$\pm$ 0.39 | Helped |
| use DPCL instead of CLIB | CIFAR100 i-blurry · CIFAR100 (M=2K), i-blurry | FM: 11.69$\pm$ 2.12 -> 11.49$\pm$ 2.54 | Helped |
| add PFI (ablation) | DPCL on CIFAR100 disjoint · CIFAR100 (M=2K), disjoint | ACC: 40.85$\pm$ 1.68 -> 45.27$\pm$ 1.32 | Helped |
| add BSC (ablation) | DPCL on CIFAR100 disjoint · CIFAR100 (M=2K), disjoint | ACC: 41.75$\pm$ 1.43 -> 45.27$\pm$ 1.32 | Helped |
Surprising findings
- On CIFAR100-SC, plain ER outperforms several advanced baselines such as BiC, MIR and FS-DGPM.
- Adding PFI and BSC to BiC and CLIB improves them, showing the components are orthogonal to existing methods.
Limitations
- Theoretical bounds rely on strong assumptions (Lipschitz losses, binary-classification reduction, linear classifiers per class).
- Hyperparameters such as sigma_a and sigma_m were searched on CIFAR100 and then reused across datasets.
- Experiments are limited to image classification with ResNet backbones.
Experimental setup
Datasets: CIFAR100 CIFAR100-SC ImageNet-100
Benchmarks: CIFAR100 CIFAR100-SC ImageNet-100
Metrics: ACC FM
Baselines: ER EWC++ DER++ BiC MIR CLIB ER-CPR FS-DGPM DRO ODDL
Builds on: Noisy Feature Mixup DRO CLIB ER BiC
Closest papers by meaning
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.89)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.88)
- Addressing Loss of Plasticity and Catastrophic Forgetting in Continual Learning (ICLR 2024 · similarity 0.88)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.87)
- Resurrecting Old Classes with New Data for Exemplar-Free Continual Learning (CVPR 2024 · similarity 0.87)
- An Effective Dynamic Gradient Calibration Method for Continual Learning (ICML 2024 · 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.