Research map / Continual learning / Online and replay-based
Online Task-Free Continual Generative and Discriminative Learning via Dynamic Cluster Memory
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 Dynamic Cluster Memory (DCM), a model-free, label-free memory management method for online task-free continual learning that builds, fills, and prunes memory clusters using a knowledge discrepancy measure, and applies it to DDPM image generation and classification.
Problem
Memory-based online task-free continual learning (OTFCL) relies on supervised signals for sample selection, which prevents its use in unsupervised learning such as continual image generation with a fixed memory budget.
Approach
A nonparametric Knowledge Discrepancy Measure (square-error or Jensen-Shannon on forward-diffused samples) drives dynamic creation of memory clusters on novelty, cluster-wise sample storage by similarity, and graph-based pruning of overlapping clusters, all without model feedback or labels.
Key results
- DCM-SE achieves average FID 54.87 across Split MNIST/Fashion/SVHN/CIFAR10 versus 89.05 for CGKD-GAN and 84.72 for R-DDPM (lower is better).
- On Split CIFAR100 classification, DCM-SE reaches 22.03 ± 0.79 and DCM-JS 22.98 ± 0.74 accuracy, above CoPE (21.62) and CNDPM (20.10).
- The extended Dynamic-JS variant reaches 98.52% on Split MNIST and 55.94% on Split CIFAR10, the best reported results in Table 4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Dynamic Cluster Memory with square-error KDM for DDPM memory management | DDPM · Split MNIST, Split Fashion, Split SVHN, and Split CIFAR10 | FID: 28.57 vs 71.67; 46.65 vs 82.23; 61.52 vs 65.18; 82.74 vs 106.18 | Helped |
| Use Dynamic Cluster Memory with Jensen–Shannon KDM for DDPM memory management | DDPM · CelebA-3DChair, CelebA-CACD, CelebA-ImageNet, and Split MINIImageNet | FID: 40.45 vs 132.12; 48.38 vs 78.00; 110.60 vs 186.76; 146.98 vs 169.26 | Helped |
| Use DCM-SE or DCM-JS memory management for DDPM | DDPM · CelebA-HQ, CACD, and FFHQ at 128 × 128 × 3 and 256 × 256 × 3 | FID: 89.23 vs 132.65; 57.19 vs 142.66; 90.80 vs 157.03; 75.18 vs 168.52; 110.21 vs 236.98; 129.38 vs 254.32 | Helped |
| Use DCM-SE memory management for classification | Classifier · Split MNIST | accuracy: 95.12 ± 0.32 vs 94.07 ± 0.13 | Helped |
| Extend DCM-JS with network expansion (Dynamic-JS) (ablation) | DCM-JS classifier · Split CIFAR10 | accuracy: 49.89± 1.23 -> 55.94 ± 0.58 | Helped |
| Extend DCM-JS with network expansion (Dynamic-JS) (ablation) | DCM-JS classifier · Split CIFAR100 | accuracy: 22.98 ± 0.74 -> 26.27 ± 0.63 | Helped |
| Vary the expansion threshold λ (ablation) | DCM-SE · Split MNIST | FID | No clear effect |
| Use square-error KDM instead of Jensen–Shannon KDM (ablation) | DCM memory management · Split CIFAR10 classification | accuracy: 50.02± 1.12 vs 49.89 ± 1.23 | Mixed |
Surprising findings
- Label-free memory management (DCM) matches or beats label-dependent memory methods in classification despite not using supervised signals.
- DCM-SE with only six memory clusters performs well on Split MNIST.
Limitations
- Image generation results are mixed: DCM-JS is worse than DCM-SE on several high-resolution datasets (e.g., CACD at 256x256 and FFHQ), and DCM-JS underperforms on CelebA-3DChair.
- Memory cluster count and expansion threshold λ require hyperparameter selection; a small λ causes frequent expansion and memory overload.
- Evaluation relies on FID with a fixed memory budget of 2,000 samples and a limited set of baselines.
Experimental setup
Datasets: MNIST Fashion-MNIST SVHN CIFAR10 CIFAR100 CelebA CACD 3DChair ImageNet MINI-ImageNet FFHQ
Benchmarks: Split MNIST Split Fashion Split SVHN Split CIFAR10 Split CIFAR100 CelebA-CACD CelebA-ImageNet CelebA-3DChair Split MINIImageNet CelebA-HQ FFHQ
Metrics: FID classification accuracy
Baselines: CGKD-GAN CGKD-WAE CNDPM LTS LGM R-VAE R-DDPM MeRGANs GEM iCARL MIR GSS
Builds on: DDPM Improved DDPM Reservoir sampling CGKD CNDPM
Closest papers by meaning
- Task-Free Continual Generation and Representation Learning via Dynamic Expansionable Memory Cluster (AAAI 2024 · similarity 0.85)
- Online Task-Free Continual Learning via Dynamic Expansionable Memory Distribution (CVPR 2025 · similarity 0.83)
- Recall-Oriented Continual Learning with Generative Adversarial Meta-Model (AAAI 2024 · similarity 0.83)
- Continual Unsupervised Generative Modelling via Online Optimal Transport (AAAI 2025 · similarity 0.83)
- Federated Continual Learning Goes Online: Uncertainty-Aware Memory Management for Vision Tasks and Beyond (ICLR 2025 · similarity 0.83)
- Dynamic Expansion Diffusion Learning for Lifelong Generative Modelling (AAAI 2025 · similarity 0.82)
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.