Research map / Continual learning / Class-incremental
Dual-Consistency Model Inversion for Non-Exemplar Class Incremental 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 DCMI, a model-inversion method for non-exemplar class incremental learning that synthesizes old-class images with both semantic (prototype alignment) and domain (adversarial alignment with new-class real images) consistency, plus prototypical routing for task-unbiased inference.
Problem
Non-exemplar class incremental learning suffers catastrophic forgetting of old classes; inverted synthetic old-class images have a large domain gap from real data, which hurts distillation and old-class recognition, and classifiers show task-recency bias.
Approach
Generate old-class synthetic images via a conditional generator with embeddings superposed by prototype similarity, enforcing semantic consistency through prototype-similarity cross-entropy in feature space and domain consistency by adversarially aligning synthetic images with real new-class images; at inference, prototype matching selects a task head (prototypical routing).
Key results
- On CIFAR-100, DCMI reaches 67.9%, 66.8%, and 64.0% average accuracy under 5, 10, and 20 phases, surpassing the previous best non-exemplar results by 1.3%, 1.0%, and 2.2%.
- On Tiny-ImageNet, DCMI outperforms SOPE by 1.1%, 1.0%, and 0.6% under 5, 10, and 20 phases.
- Ablation on CIFAR-100 shows that combining synthetic and real data for distillation with both semantic and domain consistency gives the best result (67.9% AN, 7.8% FN at P=5).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DCMI instead of the strongest listed baseline | NECIL with ResNet-18 · CIFAR-100 P=5, P=10, and P=20 | average accuracy: 66.6 -> 67.9; 65.8 -> 66.8; 61.8 -> 64.0 | Helped |
| Use DCMI instead of SOPE | NECIL with ResNet-18 · Tiny-ImageNet P=5, P=10, and P=20 | average accuracy: 53.7 -> 54.8; 52.9 -> 53.9; 51.9 -> 52.5 | Helped |
| Use DCMI instead of the strongest listed baseline | NECIL with ResNet-18 · ImageNet-Subset P=5, P=10, and P=20 | average accuracy: 69.6 -> 70.5; 69.2 -> 70.0; 62.5 -> 65.5 | Helped |
| Use DCMI instead of FeTrILfc | NECIL with ResNet-18 · ImageNet-Full P=10 E=20 | average accuracy: 64.4 -> 61.9 | Hurt |
| Add prototypical routing (ablation) | DCMI · CIFAR-100 P=5, P=10, and P=20 | average accuracy: 58.1 -> 67.9; 56.8 -> 66.8; 49.5 -> 64.0 | Helped |
| Add self-supervised learning to the initial task (ablation) | DCMI · CIFAR-100 P=5, P=10, and P=20 | average accuracy: 65.1 -> 67.9; 64.5 -> 66.8; 60.9 -> 64.0 | Helped |
| Add classifier-logit knowledge distillation (LCKD) (ablation) | DCMI · CIFAR-100 P=5, P=10, and P=20 | average accuracy: 22.6 -> 67.9; 10.0 -> 66.8; 10.6 -> 64.0 | Helped |
| Add feature knowledge distillation (LF KD) (ablation) | DCMI · CIFAR-100 P=5, P=10, and P=20 | average accuracy: 67.2 -> 67.9; 65.1 -> 66.8; 56.3 -> 64.0 | Helped |
| Use synthetic KD data with both semantic and domain consistency rather than semantic consistency alone (ablation) | DCMI · CIFAR-100 P=5 | average accuracy: 28.9 -> 67.3 | Helped |
| Use synthetic KD data with both semantic and domain consistency rather than domain consistency alone (ablation) | DCMI · CIFAR-100 P=5 | average accuracy: 54.6 -> 67.3 | Helped |
| Use real and synthetic data for KD rather than real data only (ablation) | DCMI · CIFAR-100 P=5 | average accuracy: 67.6 -> 67.9 | Helped |
Surprising findings
- Synthetic data alone with only semantic or only domain consistency yields poor performance (e.g., 28.9% AN at P=5 with synthetic only and no consistency constraints).
Limitations
- Forgetting increases as the number of tasks grows, with representational distance rising across tasks.
- Relies on self-supervised pretraining on the initial task and on the old model for prototypes and inversion.
Experimental setup
Datasets: CIFAR-100 Tiny-ImageNet ImageNet-Subset ImageNet-Full
Benchmarks: CIFAR-100 Tiny-ImageNet ImageNet-Subset ImageNet-Full
Metrics: average accuracy average forgetting
Baselines: iCaRL EEIL UCIR PODNet EWC LwF-MC MUC SDC ABD PASS IL2A SSRE
Builds on: ABD PASS DeepInversion SupCon LabelAug
Closest papers by meaning
- Prospective Representation Learning for Non-Exemplar Class-Incremental Learning (NeurIPS 2024 · similarity 0.87)
- Model Inversion with Layer-Specific Modeling and Alignment for Data-Free Continual Learning (NeurIPS 2025 · similarity 0.86)
- Dynamic Integration of Task-Specific Adapters for Class Incremental Learning (CVPR 2025 · similarity 0.85)
- Non-exemplar Online Class-Incremental Continual Learning via Dual-Prototype Self-Augment and Refinement (AAAI 2024 · similarity 0.85)
- FCS: Feature Calibration and Separation for Non-Exemplar Class Incremental Learning (CVPR 2024 · similarity 0.84)
- DS-AL: A Dual-Stream Analytic Learning for Exemplar-Free Class-Incremental Learning (AAAI 2024 · 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.