- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes AIR, an online class-incremental learning method for PTMs with adapters. The method has two components: Historical Geometry Anchoring (HGA), which uses the previous classifier directions as anchors for old-class features, and Mixed-Geometry Calibration (MGC), which replaces classifier weights after each task with class-mean features computed from the replay memory. The paper argues that these two components reduce representation drift and classifier-feature misalignment. Experiments are reported on CIFAR-100, ImageNet-R, ImageNet-A, and CUB-200 under several replay budgets.
Critical strengths
- The paper is generally clear and easy to follow. Framing continual learning as a mismatch between an evolving representation and the classifier is intuitive.
- The method is relatively simple and practical. It works with a frozen pretrained backbone, lightweight adapters, and standard replay, without introducing a complicated optimization procedure.
- The empirical results are strong under the authors' evaluation protocol, and the ablations are useful in separating the contributions of HGA and MGC.
Critical weaknesses
-
I am concerned about the novelty of the method, particularly MGC. The central operation in MGC is to compute the mean feature of each class after an incremental stage and directly replace the corresponding classifier weight with this prototype. This is essentially the same classifier construction used in MOS: MOS explicitly computes class prototypes after an incremental stage and uses the resulting prototypes as the weights of a cosine classifier. Here, this operation is presented under the name “mixed-geometry calibration”, but I do not see a clear distinction from this prior idea. This is especially important because the ablation suggests that MGC accounts for most of the improvement. For example, on CIFAR-100 with M=1000, Adapter+ER obtains 66.23/55.88 CAA/FAA, adding MGC alone gives 90.34/84.93, while full AIR obtains 91.11/85.62. Thus, almost the entire gain in this setting comes from the prototype-based classifier. MOS is also the strongest baselines in Table 1, which makes the relationship to MOS particularly relevant.
-
I also do not find HGA particularly novel. Freezing the classifier weights associated with previously learned classes while learning the new ones is a common strategy in CIL. In AIR, these classifier weights are replaced by class-mean prototypes through MGC, so HGA can be interpreted as anchoring the evolving representation to frozen prototypes from the previous task. This provides a different geometric interpretation, but I do not find the underlying mechanism clearly distinguishable from the standard practice of freezing the classifier weights for previously learned classes.
-
I have concerns about the fairness of the baseline comparison. Many of the methods in Tables 1 and 2 were originally proposed for batch CIL rather than strict online learning with a single pass over each example. The paper adapts these methods to a one-epoch online protocol. These methods normally require multiple epochs to optimize their parameters and may therefore be significantly disadvantaged when forced into a one-pass setting. In contrast, AIR's strongest component, MGC, replaces the classifier weights with prototypes and therefore does not suffer from the same SGD convergence issue. I would like to see more comparisons with prototype-based methods that do not have this limited-optimization problem.
-
There is a comparison in Table 2 with another prototype-based method, RanPAC. However, I have reason to question the validity of this comparison. RanPAC uses a frozen representation with a random projection and incrementally accumulates statistics for its classifier. Because these statistics are additive, the method can be applied sample-by-sample in an online setting, and the resulting classifier would be the same as in the batch CIL setting. The only part that is not directly compatible with the online protocol is RanPAC’s initial Phase-1 feature adaptation on the first task, which should be removed. Importantly, the original RanPAC paper already reports this “No Phase 1” setting. In Table 1 of the RanPAC paper, it obtains 89.0% on CIFAR-100, 71.8% on ImageNet-R, 58.2% on ImageNet-A, and 88.7% on CUB-200 with the same ViT-B/16 backbone. Therefore, I do not think the RanPAC values reported in Table 2, which are taken from Ma’sum et al., are valid. These results are also better than AIR at M=1000, despite using no replay data. This makes me question whether the empirical advantage reported for AIR is as strong as the current comparison suggests.
-
The main motivation of AIR is explicitly geometric: the authors argue that old-class features drift away from their historical directions and that the classifier becomes misaligned with the updated feature space. However, the experiments primarily measure classification accuracy and forgetting. I would expect a paper whose main claim is that it maintains geometric consistency to directly measure this phenomenon. For example, the authors could report cosine similarity between old-class features and historical classifier directions, classifier-to-prototype angular mismatch, or the amount of representation drift across tasks.
Questions for authors
- How exactly is the replay reservoir maintained? The paper states that standard reservoir sampling is used. Since the smallest memory is M=1000 while the datasets contain at most 200 classes, why not guarantee retained examples for each observed class? This would still satisfy the strict online protocol and would prevent the missing-class failures reported for EASE-R and MOS-R.
3 comments
Response to Weaknesses and Questions (2/2)
W5
Direct geometric evidence, M=5000. We have added geometric diagnostics throughout learning process. The metrics are computed from test-set class centroids: H measures the cosine similarity between old-class centroids and the corresponding classifier directions from the preceding stage, while D is the difference between the mean correct-class and incorrect-class cosine similarities over all seen classes. The results are averaged for the 2-10 tasks. The corresponding results will be included in the appendix.
| CIFAR-100 | ImageNet-R | ImageNet-A | CUB-200 | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Variant | FAA | H | D | FAA | H | D | FAA | H | D | FAA | H | D |
| Adapter+ER | 56.16 | 0.515 | 0.506 | 27.11 | 0.270 | 0.255 | 19.31 | 0.112 | 0.195 | 31.49 | 0.365 | 0.360 |
| HGA only | 63.22 | 0.631 | 0.614 | 37.43 | 0.423 | 0.373 | 26.99 | 0.192 | 0.180 | 33.80 | 0.426 | 0.388 |
| MGC only | 87.73 | 0.562 | 0.704 | 73.43 | 0.441 | 0.599 | 55.21 | 0.525 | 0.536 | 82.85 | 0.574 | 0.737 |
| AIR | 88.94 | 0.693 | 0.796 | 77.13 | 0.596 | 0.700 | 61.77 | 0.657 | 0.562 | 86.09 | 0.667 | 0.795 |
The results directly demonstrate AIR's effectiveness in maintaining geometric consistency.
Q1
We apply reservoir sampling after each update: the first examples are stored; for the example (), we sample uniformly from and replace buffer entry only if . This gives each example an equal probability of retention but does not guarantee coverage of every class. Applying class-balanced sampling only to EASE-R and MOS-R would make the comparison unfair. We therefore rerun both methods and AIR with class-balanced sampling at M=1000, as reported below.
| CIFAR-100 | ImageNet-R | ImageNet-A | CUB-200 | |||||
|---|---|---|---|---|---|---|---|---|
| Method | FAA | CAA | FAA | CAA | FAA | CAA | FAA | CAA |
| EASE-R | 84.29 | 89.55 | 54.02 | 63.87 | 41.43 | 55.03 | 79.52 | 88.29 |
| MOS-R | 84.47 | 89.53 | 66.04 | 73.83 | 42.57 | 53.13 | 75.33 | 83.08 |
| AIR | 86.62 | 91.34 | 69.69 | 76.03 | 47.42 | 60.60 | 79.43 | 87.80 |
Although EASE-R has a small accuracy advantage on CUB-200, AIR achieves substantially higher CAA and FAA on the other three datasets.
I thank the authors for their detailed rebuttal and additional experiments. I appreciate their efforts to clarify the methodological contributions, provide stronger baseline comparisons, and include direct geometric evidence. These additions strengthen the empirical evaluation and address several of my concerns. However, I remain unconvinced about the methodological novelty of MGC and HGA relative to existing approaches. While the new results are valuable, they do not sufficiently change my overall assessment, and I therefore maintain my original score.
Response to Weaknesses and Questions (1/2)
Thank you for your feedback. We address your questions in detail below.
W1
MGC addresses classifier-feature mismatch caused by continual updates to AIR's shared adapter. These updates also change old-class features; computing prototypes only for new classes while retaining previously computed old-class prototypes, as in the MOS-style strategy, can therefore leave old-class weights misaligned with the current representation. After each stage, MGC re-encodes buffered examples from both old and new classes using the updated model and replaces their classifier weights with the corresponding mean features. Both sets of weights thus reflect the current feature space. Within AIR, with HGA enabled in both variants (M=5000), MGC achieves higher FAA than the MOS-style classifier on all four datasets, with gains of 3.57–9.23%, as shown below. We will clarify this distinction from MOS in the related work.
W2
HGA uses a saved copy of the previous stage's calibrated classifier to supervise old-class replay features through an auxiliary loss (Eq. 10), encouraging alignment with the correct historical class direction relative to the others. This loss updates the adapters. Unlike freezing old-class weights used for prediction, HGA leaves all current classifier weights trainable, allowing adaptation to both old and new classes while retaining historical supervision. ER-ACE (ICLR 2022) provides a related comparison: its incoming-data loss masks classes absent from the current batch. The comparison between ER-ACE and AIR illustrates the difference between freezing old-class weights and our approach (M=5000):
These results also indicate that freezing old-class classifier weights when learning new classes is insufficient to match AIR's performance.
W3–W4
In the rerun, RanPAC achieves 88.14 FAA on CUB-200, exceeding AIR's 86.09 at M=5000, while AIR leads on the other three datasets.
Comparison under the single-update protocol. We have added strong OCL baselines, including PCR+NCM (CVPR 2023), EARL (CVPR 2024), and aL-SAR (ICLR 2025). PCR+NCM and EARL do not have the SGD convergence issue on classifiers. At M=5000, AIR achieves higher FAA than the baselines on ImageNet-R, ImageNet-A, and CUB-200.