Research map / Continual learning / Class-incremental
Elastic Feature Consolidation For Cold Start Exemplar-Free 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 Elastic Feature Consolidation (EFC), an exemplar-free class-incremental learning method that regularizes feature drift with an Empirical Feature Matrix pseudo-metric, plus an asymmetric prototype replay loss (PR-ACE) and EFM-weighted prototype updates, targeting the Cold Start setting.
Problem
Exemplar-free class-incremental learning when the first task is small (Cold Start), where the backbone must remain plastic yet previous-task knowledge must be retained without stored samples, causing feature drift and task-recency bias.
Approach
Derive an analytic Empirical Feature Matrix (a second-order, KL-based pseudo-metric in feature space) that regularizes feature drift only along directions important to previous tasks, combined with Gaussian prototype rehearsal via an asymmetric cross-entropy loss and EFM-weighted prototype drift compensation.
Key results
- On CIFAR-100 Cold Start 10-step, EFC reaches 58.58% per-step accuracy versus 53.91% for the best baseline (LwF) and 51.20% for FeTrIL.
- On ImageNet-Subset Warm Start 10-step, EFC achieves 68.85% per-step accuracy, about 5 points above FeTrIL (63.56%).
- On ImageNet-1K Cold Start 10-step, EFC reaches 42.62% per-step accuracy versus 34.28% for FeTrIL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace FeTrIL with EFC | CIFAR-100 Cold Start · 10-step and 20-step class-incremental learning | per-step incremental accuracy: 34.94 ± 0.46 -> 43.62 ± 0.70; 23.28 ± 1.24 -> 32.15 ± 1.33 | Helped |
| Replace FeTrIL with EFC | Tiny-ImageNet Cold Start · 10-step and 20-step class-incremental learning | per-step incremental accuracy: 30.97 ± 0.90 -> 34.10 ± 0.77; 25.70 ± 0.61 -> 28.69 ± 0.40 | Helped |
| Replace LwF or FeTrIL with EFC | ImageNet-Subset Cold Start · 10-step and 20-step class-incremental learning | per-step incremental accuracy: 37.71 ± 2.53 -> 47.38 ± 1.43; 26.63 ± 1.45 -> 35.75 ± 1.74 | Helped |
| Replace FeTrIL with EFC | ImageNet-1K Cold Start · 10-step and 20-step class-incremental learning | average incremental accuracy: 48.93 ± 0.15 -> 56.52 ± 0.10; 40.02 ± 0.15 -> 49.80 ± 0.20 | Helped |
| Add prototype updates using the EFM (ablation) | EFC · CIFAR-100 Cold Start 10 Step | per-step incremental accuracy: 39.28 ± 0.97 -> 43.62 ± 0.70 | Helped |
| Add prototype updates using the EFM (ablation) | EFC · CIFAR-100 Cold Start 20 Step | per-step incremental accuracy: 27.28 ± 1.48 -> 32.15 ± 1.33 | Helped |
| Add prototype updates using the EFM (ablation) | EFC · Tiny-ImageNet Cold Start 10 Step | per-step incremental accuracy: 31.23 ± 0.88 -> 34.10 ± 0.77 | Helped |
| Add prototype updates using the EFM (ablation) | EFC · Tiny-ImageNet Cold Start 20 Step | per-step incremental accuracy: 25.12 ± 0.23 -> 28.69 ± 0.40 | Helped |
| Add prototype updates using the EFM (ablation) | EFC · ImageNet-Subset Cold Start 10 Step | per-step incremental accuracy: 39.99 ± 2.20 -> 47.38 ± 1.43 | Helped |
| Add prototype updates using the EFM (ablation) | EFC · ImageNet-Subset Cold Start 20 Step | per-step incremental accuracy: 29.62 ± 2.60 -> 35.75 ± 1.74 | Helped |
| Use symmetric prototype loss instead of PR-ACE with EFM regularization (ablation) | EFC · CIFAR-100 Cold Start | accuracy | Hurt |
| Replace feature distillation with EFM regularization (ablation) | EFC · CIFAR-100 Cold Start | accuracy | Helped |
Surprising findings
- Feature distillation with the symmetric prototype loss essentially degenerates to a method similar to FeTrIL, i.e., freezing the backbone after the first task.
- Feature distillation and feature-drift-based methods such as FeTrIL, SSRE and PASS fall below LwF in Cold Start scenarios.
Limitations
- Stored class means may drift from real class means, potentially causing forgetting in long task streams.
- Estimating covariance drift remains open; storage of class covariance matrices grows linearly with number of classes.
Experimental setup
Datasets: CIFAR-100 Tiny-ImageNet ImageNet-Subset ImageNet-1K
Benchmarks: CIFAR-100 Tiny-ImageNet ImageNet-Subset ImageNet-1K
Metrics: per-step incremental accuracy average incremental accuracy
Baselines: EWC LwF PASS Fusion FeTrIL SSRE
Builds on: Elastic Weight Consolidation feature distillation PASS SDC (semantic drift compensation) Asymmetric cross entropy (PODNet-style asymmetric loss)
Closest papers by meaning
- Task-recency bias strikes back: Adapting covariances in Exemplar-Free Class Incremental Learning (NeurIPS 2024 · similarity 0.90)
- Data-Free Reservoir Features for Efficient Long-Horizon Cold-Start Continual Learning (NeurIPS 2026 · similarity 0.88)
- Resurrecting Old Classes with New Data for Exemplar-Free Continual Learning (CVPR 2024 · similarity 0.87)
- Compress to One Point: Neural Collapse for Pre-Trained Model-Based Class-Incremental Learning (AAAI 2025 · similarity 0.86)
- Gradual Divergence for Seamless Adaptation: A Novel Domain Incremental Learning Method (ICML 2024 · similarity 0.86)
- Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning (CVPR 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.