Research map / Continual learning / Online and replay-based
Online Task-Free Continual Learning via Dynamic Expansionable Memory Distribution
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 DEMD, a dynamic expandable memory of Gaussian memory distributions with expansion, augmentation and reduction steps for online task-free unsupervised/supervised continual generative modeling.
Problem
Catastrophic forgetting in online task-free continual learning, where task and class boundaries are unavailable, especially for continual generative modeling with limited memory.
Approach
A dual memory (short-term LIFO buffer plus a dynamically expandable set of Gaussian memory distributions in latent space) where new distributions are added when cosine distance of incoming samples exceeds a threshold, and redundant distributions are removed by symmetric KL discrepancy.
Key results
- On Split MNIST, Split Fashion, Split SVHN, Split CIFAR10 the DEMD average FID is 49.11, versus 53.30 for the best baseline DCM-JS (class-incremental, Table 1).
- On CelebA-3DChair and CelebA-CACD, DEMD FID is 38.16 and 45.28, the best among compared methods (Table 3).
- Under class-imbalanced setting, DEMD average FID is 51.81 versus 54.73 for DCM-SE (Table 2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DEMD instead of DCM-SE | Class-incremental generative modeling · Split MNIST | FID: 28.57 -> 23.62 | Helped |
| Use DEMD instead of DCM-JS | Class-incremental generative modeling · Split Fashion | FID: 43.38 -> 40.16 | Helped |
| Use DEMD instead of DCM-SE | Class-incremental generative modeling · Split SVHN | FID: 61.52 -> 59.42 | Helped |
| Use DEMD instead of DCM-JS | Class-incremental generative modeling · Split CIFAR10 | FID: 76.58 -> 73.27 | Helped |
| Use DEMD instead of DCM-SE | Imbalanced-class generative modeling · Split MNIST | FID: 29.16 -> 26.95 | Helped |
| Use DEMD instead of DCM-SE | Imbalanced-class generative modeling · Split Fashion | FID: 46.91 -> 43.27 | Helped |
| Use DEMD instead of DCM-SE | Imbalanced-class generative modeling · Split SVHN | FID: 60.58 -> 57.92 | Helped |
| Use DEMD instead of DCM-SE | Imbalanced-class generative modeling · Split CIFAR10 | FID: 82.28 -> 79.13 | Helped |
| Use DEMD instead of DCM-SE | Domain-incremental generative modeling · CelebA-3DChair | FID: 40.45 -> 38.16 | Helped |
| Use DEMD instead of DCM-JS | Domain-incremental generative modeling · CelebA-CACD | FID: 48.38 -> 45.28 | Helped |
| Use DEMD instead of DCM-SE | Few-shot continual generative modeling · S-MINIImageNet | FID: 146.98 -> 140.02 | Helped |
| Use DEMD instead of DCM-SE | High-resolution image generation · CelebA-HQ 128 × 128 × 3 | FID: 89.23 -> 86.71 | Helped |
Surprising findings
- Dynamic expansion baseline CGKD-GAN underperforms memory-based DCM methods, attributed to frozen experts being unable to learn new samples and weaker generators.
Limitations
- Dynamic expansion models require considerable storage and computation, as noted by the authors' related-work discussion
- Expansion threshold lambda is a hand-set hyperparameter (0.053 on Split MNIST)
Experimental setup
Datasets: MNIST Fashion-MNIST SVHN CIFAR10 CACD MiniImageNet CelebA 3DChair ImageNet CelebA-HQ FFHQ
Benchmarks: Split MNIST Split Fashion Split SVHN Split CIFAR10 CelebA-3DChair CelebA-CACD S-MINIImageNet CelebA-HQ FFHQ
Metrics: FID
Baselines: MeRGANs LTS LGM DCM-SE DCM-JS CGKD-GAN CGKD-WAE
Builds on: DCM CGKD VAE
Closest papers by meaning
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.90)
- Dynamic Expansion Diffusion Learning for Lifelong Generative Modelling (AAAI 2025 · similarity 0.90)
- Task-Free Continual Generation and Representation Learning via Dynamic Expansionable Memory Cluster (AAAI 2024 · similarity 0.89)
- Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.89)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.89)
- Mitigating Catastrophic Forgetting in Online Continual Learning by Modeling Previous Task Interrelations via Pareto Optimization (ICML 2024 · similarity 0.89)
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.