Research map / Continual learning / Online and replay-based
An Empirical Study of Example Forgetting during Deep Neural Network Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Remove the least-forgotten examples in order (ablation) | ResNet18 · CIFAR-10 | test generalization performance: 30% of the dataset can be removed | No clear effect |
| Remove the least-forgotten examples in order (ablation) | ResNet18 · CIFAR-10 | test generalization performance: up to 35% can be removed with marginal degradation (less than 0.2%) | Hurt |
| Remove examples selected at random (ablation) | ResNet18 · CIFAR-10 | test generalization performance | Hurt |
| Remove examples in order of increasing forgetting statistics (ablation) | Classifier · MNIST | test generalization performance: 80% of MNIST | No clear effect |
| Remove examples in order of increasing forgetting statistics (ablation) | Classifier · permutedMNIST | test generalization performance: 50% of permutedMNIST | No clear effect |
| Remove examples in order of increasing forgetting statistics (ablation) | ResNet18 · CIFAR-100 | test performance: all unforgettable examples (~$8\%$ of the training set) | No clear effect |
| Remove chunks with increasingly higher forgetting statistics (ablation) | CIFAR-10 classifier · CIFAR-10 | generalization performance: $5{,}000$ examples | Mixed |
| Change labels for 20% of training examples (ablation) | CIFAR-10 · CIFAR-10 with noisy labels | forgetting events: 20% | Hurt |
| Add increasing additive Gaussian pixel noise (ablation) | CIFAR-10 training images · CIFAR-10 | number of unforgettable examples: $\lambda \in \{0.5, 1, 2, 10\}$ | Hurt |
| Use ResNet18's example ordering instead of WideResNet's own ordering (ablation) | WideResNet · CIFAR-10 | generalization performance: 30% of the dataset removed | No clear effect |
| Partition examples by forgetting statistics into f0 and fN rather than using a random split (ablation) | CIFAR-10 continual-learning setup · CIFAR-10 | forgetting on the other partition: $5k$ examples | Mixed |
Closest papers by meaning
- Characterizing Datapoints via Second-Split Forgetting (NeurIPS 2022 · similarity 0.93)
- Fortuitous Forgetting in Connectionist Networks (ICLR 2022 · similarity 0.90)
- Predicting the Susceptibility of Examples to Catastrophic Forgetting (ICML 2025 · similarity 0.90)
- The Importance of Being Lazy: Scaling Limits of Continual Learning (ICML 2025 · similarity 0.88)
- On the Diminishing Returns of Width for Continual Learning (ICML 2024 · similarity 0.88)
- Towards Understanding Catastrophic Forgetting in Two-layer Convolutional Neural Networks (ICML 2025 · similarity 0.88)
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.