Research map / Continual learning / Class-incremental
Prospective Representation Learning 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 Prospective Representation Learning (PRL), which squeezes base-phase embeddings to reserve space for future classes and uses prototype-guided latent-space alignment to fit new classes into that space, improving non-exemplar class-incremental learning on four benchmarks.
Problem
Non-exemplar class-incremental learning (NECIL) must discriminate old and new classes without any stored old-class samples; existing methods resolve old/new conflicts only after new tasks arrive, and base-phase training spreads embeddings over the full space, leaving little room for later classes.
Approach
Address conflicts prospectively: in the base phase, a preemptive embedding squeezing (PES) loss tightens intra-class clusters and reserves inter-class space; in incremental phases, a prototype-guided representation update (PGRU) uses a learned projector to make new-class features orthogonal to stored old prototypes in a latent space and aligns the current encoder to that latent space.
Key results
- On CIFAR-100 (P=5) PRL reaches 71.26% average incremental accuracy vs 70.44% for the prior SOTA NAPA-VQ, and on TinyImageNet (P=5) 58.12% vs 53.69% for SOPE.
- On ImageNet-Subset PRL averages a 3.6% gain over suboptimal methods; on ImageNet-1K (P=10) it reaches 62.74% vs 61.53% for POLO.
- Plugged into IL2A and PASS on CIFAR-100, PRL adds on average about 2.7% and 2.3% respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PES to the baseline (ablation) | NECIL baseline · CIFAR-100 (P=5) | average incremental accuracy: 69.25 -> 70.57 | Helped |
| Add PES to the baseline (ablation) | NECIL baseline · CIFAR-100 (P=10) | average incremental accuracy: 68.52 -> 69.64 | Helped |
| Add PES to the baseline (ablation) | NECIL baseline · CIFAR-100 (P=20) | average incremental accuracy: 65.93 -> 67.58 | Helped |
| Add PES to the baseline (ablation) | NECIL baseline · TinyImageNet (P=5) | average incremental accuracy: 55.04 -> 57.08 | Helped |
| Add PES to the baseline (ablation) | NECIL baseline · TinyImageNet (P=10) | average incremental accuracy: 54.15 -> 55.84 | Helped |
| Add PES to the baseline (ablation) | NECIL baseline · TinyImageNet (P=20) | average incremental accuracy: 51.65 -> 53.58 | Helped |
| Add PGRU to the baseline (ablation) | NECIL baseline · CIFAR-100 (P=5) | average incremental accuracy: 69.25 -> 70.36 | Helped |
| Add PGRU to the baseline (ablation) | NECIL baseline · CIFAR-100 (P=10) | average incremental accuracy: 68.52 -> 69.23 | Helped |
| Add PGRU to the baseline (ablation) | NECIL baseline · CIFAR-100 (P=20) | average incremental accuracy: 65.93 -> 67.17 | Helped |
| Add PGRU to the baseline (ablation) | NECIL baseline · TinyImageNet (P=5) | average incremental accuracy: 55.04 -> 56.79 | Helped |
| Add PGRU to the baseline (ablation) | NECIL baseline · TinyImageNet (P=10) | average incremental accuracy: 54.15 -> 56.05 | Helped |
| Add PGRU to the baseline (ablation) | NECIL baseline · TinyImageNet (P=20) | average incremental accuracy: 51.65 -> 53.16 | Helped |
Surprising findings
- The base-phase PES constraint contributes more than the incremental-phase PGRU module (ablation: PES alone gives 70.57 vs PGRU alone 70.36 on CIFAR-100 P=5).
Limitations
- The number and distribution of unknown future classes cannot be predicted, so how much base-class space to reserve is left open.
- Sensitivity to hyperparameters lambda and gamma increases with more incremental phases (20 phases).
Experimental setup
Datasets: CIFAR-100 TinyImageNet ImageNet-Subset ImageNet-1K
Benchmarks: CIFAR-100 TinyImageNet ImageNet-Subset ImageNet-1K
Metrics: average incremental accuracy top-1 accuracy per phase
Baselines: PASS SSRE SOPE POLO PRAKA NAPA-VQ SDC LwF_MC
Builds on: PRAKA PASS IL2A SSRE
Closest papers by meaning
- FCS: Feature Calibration and Separation for Non-Exemplar Class Incremental Learning (CVPR 2024 · similarity 0.89)
- Dual-Consistency Model Inversion for Non-Exemplar Class Incremental Learning (CVPR 2024 · similarity 0.87)
- Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.87)
- Compress to One Point: Neural Collapse for Pre-Trained Model-Based Class-Incremental Learning (AAAI 2025 · similarity 0.86)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.86)
- eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI 2024 · similarity 0.86)
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.