Research map / Continual learning / Online and replay-based

Predicting the Susceptibility of Examples to Catastrophic Forgetting

ICML 2025 · Continual learning › Online and replay-based · approach: Theoretical analysis of forgetting · 1 citations (OpenAlex)

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.

Shows that examples learned faster are less prone to catastrophic forgetting and proposes Speed-Based Sampling (SBS), a replay buffer sampler that drops the quickest and slowest learned examples before uniform sampling, improving many replay CL methods.

Problem

Catastrophic forgetting in replay-based continual learning: uniform random sampling of the replay buffer treats all old-task examples equally, though some examples are far more susceptible to being forgotten.

Approach

Learning speed (fraction of epochs an example is classified correctly during training on its task) predicts forgetting: quickly learned examples are rarely forgotten, while slowly learned ones are. SBS uses this score to filter out the quickest q% and slowest s% examples before sampling the buffer uniformly, adding minimal overhead.

Key results

• SBS with tuned q and s improves final accuracy over uniform random sampling at buffer size 1k on CIFAR-100-20 (51.75 to 55.43) and CIFAR-10-5 (79.48 to 83.59).
• SBS achieves the highest mean accuracy in all six tested scenarios in the buffer-sampling comparison, e.g. 89.17 vs 82.4 random on CIFAR-10-5 with 10k buffer.
• Across replay methods (DER, DER++, X-DER, AGEM, ER, GEM, RPC, FDR), SBS consistently improves performance on CIFAR-100-5, CIFAR-10-5 and TinyImageNet-10 with a 500-example buffer.

Measured effects reported in the paper

Surprising findings

• Smaller buffers benefit from keeping quickly learned examples and discarding slow ones, whereas larger buffers shift toward retaining slowly learned examples.
• Learning speed correlates strongly with memorization: replay buffer size is strongly negatively correlated with the mean learning speed of remembered examples (r=-0.966).

Limitations

• Learning speed depends on the number of epochs, so it becomes coarse and noisy with few epochs and is not well-suited to single-epoch stream settings.
• Hyperparameters q and s require a grid search via an auxiliary RotNet task, adding cost for large datasets.
• Experiments are limited to image classification benchmarks with ResNet-18 and small datasets.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 TinyImageNet

Benchmarks: CIFAR-100-2 CIFAR-100-5 CIFAR-100-20 CIFAR-10-5 TinyImageNet-10 TinyImageNet-2

Metrics: mean final test accuracy across tasks standard error remembered-example learning speed correlation

Baselines: Random sampling Max Ent IPM GSS Herding LARS DER DER++ X-DER AGEM ER GEM

Builds on: experience replay accessibility score DER++ RotNet

Closest papers by meaning

• A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.91)
• Do Your Best and Get Enough Rest for Continual Learning (CVPR 2025 · similarity 0.91)
• Understanding the Forgetting of (Replay-based) Continual Learning via Feature Learning: Angle Matters (ICML 2025 · similarity 0.91)
• Unlocking the Power of Rehearsal in Continual Learning: A Theoretical Perspective (ICML 2025 · similarity 0.91)
• Mitigating Catastrophic Forgetting in Online Continual Learning by Modeling Previous Task Interrelations via Pareto Optimization (ICML 2024 · similarity 0.90)
• Layerwise Proximal Replay: A Proximal Point Method for Online Continual Learning (ICML 2024 · similarity 0.90)