- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes Distribution Matching Replay (DMR) for task-agnostic, single-pass online class-incremental learning. DMR maintains class-wise statistics of all observed samples in a fixed random Fourier feature (RFF) space and scores removals and insertions by their effects on mean and covariance discrepancies between the buffer and historical summaries. Generally, DMR achieves satisfactory performance on some small-scale and middle-scale datasets.
Critical strengths
-
The problem studied is a very important topic: a small replay buffer is an imperfect proxy for the historical stream, and retaining compact statistics of discarded samples is a sensible way to improve selection under a single-pass constraint. The proposed method shows that more informative statistics about historical examples can be preserved.
-
The shared, fixed RFF map and additive statistics provide a coherent mechanism for historical distribution matching without retaining all past examples. Appendix D appropriately distinguishes sketch approximation from the separate RFF approximation and acknowledges adaptive-selection limitations.
Critical weaknesses
-
It seems that the proposed method can also address many typical problems of learning from streaming data. Therefore, can this method be applied to concept-drifted (or distributionally drifting) streaming data? If the streaming data has significant distribution shift, can this method still have satisfactory performance? This experiment can make the proposed method more general for online learning.
-
Some baseline methods are missing. For example, online coreset selection with a streaming k-center algorithm is not selected as a baseline method. Moreover, the baseline methods are quite old since they were proposed before 2024. Section 2 describes herding as matching class means and PICS as selecting visual-similarity clusters, yet neither appears in the comparison. Additionally, the setting is a continual learning setup, so more powerful continual learning baseline methods should be considered in experiments. It would be better if the authors can consider more recently proposed update and sampling strategies for reply buffers, especially in the continual learning domain.
-
How can the proposed method maintain examples of all seen classes? Will it happen that some minor classes in previously seen data are fully swept out? The experimental setup in this paper is unrealistic. The class imbalance problem sometimes is more troublesome in the setting of online learning or CIL, since in practice, the instances for each class are not equally important. But class-conditional matching does not fully specify the intended buffer distribution. Eq. (8) and Algorithm 1 compare marginal changes in each candidate's class-wise discrepancy, while the replacement quota is global and replay sampling is uniform over stored examples. Even perfect class-conditional matching does not match the joint distribution if buffer class proportions differ from the desired priors. No explicit class weights, minimum class occupancy, or allocation constraint is given. It would be better if the authors can explain these problems.
-
Kernel calibration may depend strongly on the beginning of the stream. Appendix H sets bandwidths from the buffer when it first reaches capacity, then freezes the RFF map to keep statistics comparable. This is consistent, but calibration occurs before many later classes are observed. Input scaling, augmentation, and the precise representation used by z(x) are also unclear, although they determine RBF geometry. Clarify these choices and test shuffled class orders and bandwidth perturbations, particularly on Tiny-ImageNet. The unresolved question is whether geometry calibrated on early classes remains informative for later classes.
Questions for authors
see weaknesses
6 comments
Response to Reviewer Ms92 (W2)
W2. Comparison with Additional Replay-Selection and Continual-Learning Methods
Additional selection baselines. We compare DMR against GCR [1], LWP [2] and CSReL [3] under the common ER learner. Both LWP and CSReL are 2025 methods; Low Weight Permutation (LWP) uses prediction-based representativeness, while CSReL measures sample importance via reducible loss. All comparisons in this section use the paper's dataset–buffer settings and corresponding training/evaluation protocols, with mean ± std over five runs. The table reports final average accuracy (%); DMR values are from the original Table 1.
| Dataset | Buffer | GCR (2022) | LWP (2025) | CSReL (2025) | DMR (ours) |
|---|---|---|---|---|---|
| Seq-MNIST | 50 | 60.7 ± 4.4 | 62.8 ± 1.1 | 65.4 ± 3.2 | 76.4 ± 1.4 |
| Seq-MNIST | 100 | 71.4 ± 3.1 | 74.8 ± 1.9 | 76.0 ± 1.5 | 82.1 ± 1.6 |
| Seq-CIFAR-10 | 200 | 27.2 ± 6.3 | 30.2 ± 2.2 | 29.8 ± 2.8 | 31.7 ± 1.0 |
| Seq-CIFAR-10 | 500 | 36.4 ± 4.3 | 40.2 ± 3.2 | 40.8 ± 1.4 | 43.9 ± 1.2 |
| Seq-CIFAR-100 | 1000 | 13.7 ± 0.8 | 16.7 ± 2.4 | 16.1 ± 0.4 | 18.3 ± 2.1 |
| Seq-CIFAR-100 | 2000 | 15.7 ± 0.8 | 19.6 ± 1.2 | 20.0 ± 2.1 | 20.9 ± 1.4 |
| Seq-Tiny-ImageNet | 1000 | 8.6 ± 1.1 | 9.6 ± 1.0 | 9.7 ± 0.3 | 10.1 ± 0.2 |
| Seq-Tiny-ImageNet | 2000 | 13.0 ± 0.3 | 13.5 ± 0.9 | 13.8 ± 0.4 | 14.1 ± 0.8 |
DMR has higher mean accuracy than all three compared baselines in all eight settings, and lower average forgetting as well (will be reported in the revised manuscript). It is worth noting that, although CSReL is a task-aware method, our proposed DMR manages to outperform it without leveraging task boundary/identity information.
K-center, herding and PICS. The original Table 1 already includes DSS, a 2-approximate k-center selection method evaluated under our common online protocol. As for herding and PICS, they rely on task boundary to perform candidate set selection or visual-similarity cluster construction. Consequently, neither of these methods can directly satisfy our task-agnostic, single-pass access constraints. That said, we do acknowledge that herding is indeed an important and widely adopted sample selection strategy. Hence, we additionally evaluate the original iCaRL-style herding. The results reveal that DMR achieves higher mean accuracy (+2.5 to +8.0 pp) and lower forgetting (-6.7 to -2.3 pp) than herding in all eight settings; this task-aware comparison will appear together with CSReL in the revised appendix.
Additional learners. We integrate DMR into PCR [4] and LOR [5], replacing only the buffer-update policy while preserving each learner's objective and training procedure, using paired runs. LOR is a recent method proposed in 2026; it explores lookahead model states along different weighted combinations of current-batch and replay gradients, then optimizes a first-order Log-Sum-Exp surrogate that emphasizes the worst-performing sampled directions.
| Dataset | Buffer | PCR (2023) | PCR + DMR | LOR (2026) | LOR + DMR |
|---|---|---|---|---|---|
| Seq-MNIST | 50 | 82.4 ± 4.8 | 85.2 ± 1.9 | 86.5 ± 1.9 | 88.0 ± 1.0 |
| Seq-MNIST | 100 | 87.3 ± 2.5 | 90.6 ± 2.2 | 90.9 ± 0.5 | 92.1 ± 0.3 |
| Seq-CIFAR-10 | 200 | 42.9 ± 4.5 | 44.1 ± 3.9 | 56.7 ± 0.9 | 58.9 ± 1.0 |
| Seq-CIFAR-10 | 500 | 50.8 ± 4.0 | 52.9 ± 2.4 | 60.7 ± 1.6 | 62.1 ± 1.6 |
| Seq-CIFAR-100 | 1000 | 25.8 ± 1.0 | 29.2 ± 1.2 | 24.5 ± 1.2 | 26.3 ± 0.4 |
| Seq-CIFAR-100 | 2000 | 32.9 ± 0.9 | 34.1 ± 1.5 | 26.9 ± 1.2 | 28.1 ± 0.8 |
| Seq-Tiny-ImageNet | 1000 | 15.6 ± 2.0 | 18.7 ± 1.9 | 19.3 ± 1.0 | 22.9 ± 0.5 |
| Seq-Tiny-ImageNet | 2000 | 19.6 ± 0.9 | 21.8 ± 0.3 | 24.9 ± 1.6 | 26.5 ± 0.9 |
DMR improves mean accuracy in all 16 comparisons (PCR: +1.2 to +3.4 pp; LOR: +1.2 to +3.6 pp), and reduces average forgetting in all 16, extending the original Table 2 to additional mechanisms and a 2026 learner.
Response to Reviewer Ms92 (W3)
W3. Class Preservation and Class-Imbalanced Streams
Allocation and class priors. We agree that matching alone does not guarantee matching the joint distribution , which also depends on . However, historical class proportions are not necessarily the optimal replay allocation under a limited memory budget. For example, classes with limited intra-class variation may be adequately represented by fewer exemplars, leaving more capacity for classes with greater diversity. DMR evaluates candidate replacements through changes in class-conditional discrepancy, allowing allocation to respond to per-class matching errors rather than enforcing historical-frequency quotas.
Class preservation. All DMR experiments in the submission and rebuttal used two safeguards: an incoming class absent from the buffer () receives insertion score , prioritizing admission when slots are available; deleting a class's last buffered sample () incurs removal cost . Since removal minimizes cost, represented classes are preserved when the deletion quota satisfies , where is the number of deletions, the buffer size, and the number of represented classes. This condition holds in all tested configurations. The submitted version's Appendix E, including Algorithm 1 and its empty-class convention, describes an obsolete earlier implementation; it differs from the current implementation only in the safeguards, and all experiments used the implmentation with safeguards. We apologize for this inconsistency and will revise it to reflect the safeguards used in the experiments.
DMR also has a potential implicit balancing tendency. For , removing feature changes the class mean by ; adding one changes it by . Thus, under comparable conditions, changes to smaller classes can be larger. This potentially makes their samples costlier to remove, although candidate scores also depend on feature geometry and the current discrepancy.
Original benchmark buffer diagnostics. Across all eight settings, DMR had zero missing classes in all 40 final-task snapshots and lower task-averaged class-count coefficient of variation (CV) than RS. RS lost classes on Seq-Tiny-ImageNet () while DMR did not. Representative results are as follows; full results will appear in the revision.
| Setting | Final missing, RS / DMR | Mean CV across tasks, RS | Mean CV across tasks, DMR |
|---|---|---|---|
| Seq-CIFAR-100, | 0 / 0 | 0.214 ± 0.021 | 0.152 ± 0.008 |
| Seq-CIFAR-100, | 0 / 0 | 0.151 ± 0.009 | 0.105 ± 0.002 |
| Seq-Tiny-ImageNet, | 1.60 ± 1.15 / 0 | 0.326 ± 0.023 | 0.182 ± 0.006 |
| Seq-Tiny-ImageNet, | 0 / 0 | 0.229 ± 0.002 | 0.120 ± 0.009 |
Class-imbalanced streams. We evaluate task-wise class-imbalanced Seq-CIFAR-100, , with . Training counts decay exponentially within each 10-class task; the test set is unchanged. RS and DMR share five paired streams and seeds. Minority denotes the least frequent 30% of classes.
| IF | Method | Final Acc. | Fgt. | Mean CV across tasks ↓ | Final missing | Minority coverage |
|---|---|---|---|---|---|---|
| 3 | RS | 9.6 ± 0.9 | 13.9 ± 1.0 | 0.403 ± 0.009 | 0 | 100% |
| 3 | DMR | 12.8 ± 2.0 | 11.2 ± 1.0 | 0.161 ± 0.005 | 0 | 100% |
| 10 | RS | 4.7 ± 1.2 | 16.8 ± 0.7 | 0.731 ± 0.023 | 1.60 ± 1.82 | 94.67 ± 6.06% |
| 10 | DMR | 8.0 ± 1.4 | 10.8 ± 0.6 | 0.389 ± 0.084 | 0 | 100% |
For IF=3/10, DMR improves accuracy by 3.2/3.3 pp and reduces forgetting by 2.7/6.0 pp, retaining all seen classes in the final buffer across the tested seeds.
An explicitly balanced variant of DMR. We also tested DMR with approximately uniform feasible class quotas for removal/insertion. Its accuracy is lower than original DMR in all eight settings (e.g., 15.7 ± 1.3% vs. 18.3 ± 2.1% on Seq-CIFAR-100, ). This suggests that rigid uniform quotas do not improve DMR in these settings.
Response to Reviewer Ms92 (W1)
We sincerely thank the reviewer for the constructive suggestions on distribution shifts, baselines, class preservation, and kernel calibration.
W1. Robustness to Distribution Shift and Concept Drift
We evaluate substantial distribution shifts using a heterogeneous, class-disjoint OCIL stream. This tests shifts across image domains and class semantics. As for within-class temporal drift, systematically adapting to such nonstationarity is more closely related to Online Domain-Incremental Learning (ODIL) and lies beyond the class-disjoint setting evaluated here. Therefore, as discussed in Section 5 of the paper, we leave this as a direction for future work.
Cross-dataset experiment. We introduce Seq-MultiDataset, comprising five sequential tasks from distinct datasets: CIFAR-10 (Natural Objects and Scenes) → FashionMNIST (Clothing and Footwear) → SVHN (Real-world Street House Numbers) → KMNIST (Cursive Japanese Characters) → EuroSAT-RGB (Satellite Imagery of Land Use and Terrain). Each task contains 10 disjoint classes with 500 training samples per class, yielding unified 50-way classification. Images are resized to 32×32×3 (replicating grayscale channels) and normalized with mean and standard deviation of 0.5. Reservoir sampling (RS) and DMR use the same ER learner under single-pass, task-agnostic training with K=500, differing only in buffer updates. Results are mean ± std over five seeds; overall accuracy equally weights the class-macro test accuracies of the five datasets to address the imbalance in test-set sizes across these datasets.
DMR improves overall final accuracy by 3.6 pp and reduces average forgetting by 2.0 pp. Mean accuracy improves on all five constituent datasets, supporting DMR's effectiveness under the evaluated cross-dataset distribution/semantic shifts.