Research map / Continual learning / Class-incremental
Task-recency bias strikes back: Adapting covariances in Exemplar-Free 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 AdaGauss, an exemplar-free class-incremental learning method that adapts both means and covariances of Gaussian class distributions across tasks via an adapter network and prevents dimensionality collapse with an anti-collapse loss, with projected distillation.
Problem
Exemplar-free class-incremental learning suffers from drift of class covariances when the feature extractor is updated, and from dimensionality collapse that makes covariance matrices non-invertible and causes task-recency bias.
Approach
Train an adapter network to map old class Gaussians into the new latent space (adapting both mean and covariance), and add an anti-collapse loss on the Cholesky diagonal of minibatch covariance to keep features full-rank, combined with feature distillation through a learnable projector.
Key results
- On ImagenetSubset (T=10) trained from scratch, AdaGauss reaches 51.1 A_last and 65.0 A_inc, improving over EFC by 3.7 points in A_last.
- On CUB200 (T=10) with ImageNet-pretrained ResNet18, AdaGauss improves A_last over EFC by 4.8 points (55.8 vs 51.0).
- Removing the anti-collapse loss and using covariance shrinkage of 0.5 drops ImagenetSubset (T=10) average accuracy by 4.0 points relative to the full method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| adapt both old-class means and covariances rather than adapting means only | AdaGauss versus EFC · ImagenetSubset (T=10) | A_last: +3.7 percentage points | Helped |
| adapt both old-class means and covariances rather than adapting means only | AdaGauss versus EFC · ImagenetSubset (T=20) | A_last: +6.8% points | Helped |
| replace Bayes classification and covariance use with NMC and no covariance matrix (ablation) | AdaGauss · CIFAR-100 (T=10) and ImagenetSubset (T=10) | A_last: 9.7% and 9.6% points lower | Hurt |
| replace Bayes classification with NMC while retaining full covariance (ablation) | AdaGauss · CIFAR-100 (T=10) and ImagenetSubset (T=10) | A_last | Hurt |
| use diagonal instead of full covariance matrices (ablation) | AdaGauss · CIFAR-100 (T=10) and ImagenetSubset (T=10) | A_last: 5.0% and 3.9% | Hurt |
| remove mean adaptation (ablation) | AdaGauss · CIFAR-100 (T=10) and ImagenetSubset (T=10) | A_last: 21.5% and 27.2% | Hurt |
| remove covariance adaptation (ablation) | AdaGauss · CIFAR-100 (T=10) and ImagenetSubset (T=10) | A_last: 3.2% and 3.1% | Hurt |
| replace anti-collapse loss with covariance shrinkage (0.5) (ablation) | AdaGauss · CIFAR-100 (T=10) and ImagenetSubset (T=10) | A_last: 5.9% and 4.0% | Hurt |
| use projected distillation instead of logit distillation | AdaGauss · CIFAR100, ImagenetSubset, and CUB200 (T=10) | A_last: 1.4%, 0.9%, and 4.0% points | Helped |
| use anti-collapse loss during feature training (ablation) | AdaGauss · ImagenetSubset (T=10) | A_last: Without L_AC and shrink, results were not included due to inability to invert covariance matrices. | Helped |
| increase anti-collapse regularization strength from β=1 (ablation) | AdaGauss · ImagenetSubset | A_last | Hurt |
| decrease anti-collapse regularization strength to β=0.1 (ablation) | AdaGauss · ImagenetSubset | A_last | Hurt |
Surprising findings
- Adapting only means without covariance adaptation (EFC-style) yields little gap, but the EFC approach does not reduce covariance distance versus no adaptation.
Limitations
- Cross-entropy only separates classes from the current task, so old classes may overlap with each other and with new classes, causing forgetting.
- With very few examples per class, high-dimensional covariance is hard to estimate, requiring a bottleneck layer that may limit representational strength.
- On half-dataset settings AdaGauss is slightly worse than the most recent baselines using default hyperparameters.
Experimental setup
Datasets: CIFAR100 TinyImageNet ImagenetSubset CUB200 FGVCAircraft StanfordCars
Benchmarks: CIFAR100 TinyImageNet ImagenetSubset CUB200 FGVCAircraft StanfordCars
Metrics: average accuracy (A_last) average incremental accuracy (A_inc) KL divergence to ground-truth distributions training and inference time
Baselines: EWC LwF PASS IL2A SSRE FeTrIL FeCAM DS-AL EFC
Builds on: EFC FACIL FeCAM PASS knowledge distillation
Closest papers by meaning
- Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · similarity 0.90)
- Elastic Feature Consolidation For Cold Start Exemplar-Free Incremental Learning (ICLR 2024 · similarity 0.90)
- Semantic Shift Estimation via Dual-Projection and Classifier Reconstruction for Exemplar-Free Class-Incremental Learning (ICML 2025 · similarity 0.88)
- eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI 2024 · similarity 0.88)
- Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning (CVPR 2024 · similarity 0.87)
- Class Incremental Learning via Likelihood Ratio Based Task Prediction (ICLR 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.