Research map / Continual learning / Online and replay-based
The Inter-batch Diversity of Samples in Experience Replay for Continual 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 studying inter-batch sample diversity in experience replay via a Batch Cosine Distance metric and a Random Batch Sampling proof-of-concept that selects the highest-scoring of several random replay batches, evaluated on Split MNIST and Split CIFAR-10.
Problem
Experience replay in continual learning suffers from catastrophic forgetting; existing replay sample selection focuses on individual sample informativeness and ignores the cumulative diversity of samples across a batch and a sequence of batches.
Approach
Introduces Batch Cosine Distance (BCD), a metric measuring changes in hidden representations within a batch before and after a model update, and uses it to rank random candidate replay batches (Random Batch Sampling), selecting the highest-scoring one.
Key results
- On CIFAR-10 with memory 200, RBS reached 27.74% average accuracy vs 24.39% for MIR and 22.54% for ER.
- On CIFAR-10 with memory 1000, RBS reached 42.15% vs 40.09% for MIR and 28.53% for ER.
- Diverse Features Sampling (DFS) improved average accuracy over reservoir sampling with random and MIR buffer selection in most CIFAR-10 settings (e.g., 31.09 vs 29.59 at memory 500 for random).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DFS instead of reservoir sampling for memory selection | Experience Replay with random replay or MIR · CIFAR-10, memory 200–1000 | A5 accuracy (%): 23.84 ± 0.90 -> 24.31 ± 0.69; 29.59 ± 0.73 -> 31.09 ± 1.04; 35.62 ± 0.99 -> 35.78 ± 1.52; 25.70 ± 0.98 -> 26.67 ± 0.95; | Helped |
| Select replay batches using BCD-ranked random batch sampling (RBS) instead of ER | Experience Replay · CIFAR-10, memory 200–1000 | A5 accuracy (%): 22.54 ± 2.00 -> 27.74 ± 3.45; 26.51 ± 2.76 -> 36.79 ± 2.73; 28.53 ± 4.17 -> 42.15 ± 2.08 | Helped |
| Select replay batches using BCD-ranked random batch sampling (RBS) instead of MIR | Experience Replay · CIFAR-10, memory 200–1000 | A5 accuracy (%): 24.39 ± 1.85 -> 27.74 ± 3.45; 32.06 ± 3.51 -> 36.79 ± 2.73; 40.09 ± 3.52 -> 42.15 ± 2.08 | Helped |
| Select replay batches using BCD-ranked random batch sampling (RBS) instead of ER | Experience Replay · MNIST, memory 200–800 | A5 accuracy (%): 79.50 ± 5.00 -> 81.81 ± 2.70; 84.91 ± 3.61 -> 87.01 ± 3.25; 86.22 ± 2.29 -> 89.33 ± 2.49 | Helped |
| Select replay batches using BCD-ranked random batch sampling (RBS) instead of MIR | Experience Replay · MNIST, memory 200–800 | A5 accuracy (%): 80.86 ± 4.09 -> 81.81 ± 2.70; 85.09 ± 5.61 -> 87.01 ± 3.25; 89.47 ± 1.94 -> 89.33 ± 2.49 | Mixed |
Surprising findings
- On MNIST, the advantage of the proposed method over MIR narrows or disappears at larger memory sizes (e.g., 89.33 vs 89.47 at memory 800).
Limitations
- Random Batch Sampling is a proof-of-concept and the paper does not systematically develop sequence-level diversity strategies.
- DFS has high computational complexity and tends to accumulate outliers, per the author.
- Experiments are limited to small benchmarks (Split MNIST and Split CIFAR-10) with small networks.
Experimental setup
Datasets: MNIST CIFAR-10
Benchmarks: Split MNIST Split CIFAR-10
Metrics: average accuracy A5 average forgetting F5
Baselines: Experience Replay MIR Random sampling Reservoir sampling
Builds on: Experience Replay Maximally Interfered Retrieval Diverse Features Sampling
Closest papers by meaning
- Online Bias Correction for Task-Free Continual Learning (ICLR 2023 · similarity 0.89)
- New Insights on Reducing Abrupt Representation Change in Online Continual Learning (ICLR 2022 · similarity 0.89)
- Online Continual Learning with Maximal Interfered Retrieval (NeurIPS 2019 · similarity 0.88)
- Predicting the Susceptibility of Examples to Catastrophic Forgetting (ICML 2025 · similarity 0.87)
- Summarizing Stream Data for Memory-Constrained Online Continual Learning (AAAI 2024 · similarity 0.87)
- Layerwise Proximal Replay: A Proximal Point Method for Online Continual Learning (ICML 2024 · similarity 0.86)
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.