acceptodds
Under review as a conference paper at ICLR 2027

Anchor the Past, Calibrate the Present: Maintaining Geometric Consistency for Online Class-Incremental Learning

12 views

Abstract

Realistic open environments demand continual recognition of emerging and previously learned categories, motivating online class-incremental learning (OCIL). As pretrained models (PTMs) adapt to new classes, old-class features can drift from historical classifier directions, while the classifier can become misaligned with the updated feature space. Replaying old examples alone does not maintain geometric consistency across these updates. To bridge this gap, we propose Anchoring in Replay (AIR), which contains two complementary parts. Historical geometry anchoring (HGA) aligns old-class replay features with fixed directions from the previous calibrated classifier, while mixed-geometry calibration (MGC) recalibrates the classifier from memory prototypes in the updated feature space. The calibrated directions support current prediction and anchor the next task. Under OCIL protocol, AIR achieves the best final average accuracy (FAA) among the evaluated replay-based methods on four benchmarks across three replay budgets.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
online continual learning, experience replay, catastrophic forgetting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_LjCYOpenReview ↗
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
  1. The paper is generally clear and easy to follow. Framing continual learning as a mismatch between an evolving representation and the classifier is intuitive.
  2. 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.
  3. 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
  1. 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.

  2. 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.

  3. 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.

  4. 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.

  5. 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
  1. 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
AuthorsOpenReview ↗

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.

CIFAR-100ImageNet-RImageNet-ACUB-200
ClassifierFAACAAFAACAAFAACAAFAACAA
MOS-style84.1184.9371.0172.5158.2065.9176.8679.44
MGC88.9492.5077.1380.0961.7768.2886.0989.89

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):

CIFAR-100ImageNet-RImageNet-ACUB-200
MethodFAACAAFAACAAFAACAAFAACAA
ER-ACE74.2079.2459.4767.9023.2833.7166.7577.57
AIR88.9492.5077.1380.0961.7768.2886.0989.89

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.

CIFAR-100ImageNet-RImageNet-ACUB-200
MethodFAACAAFAACAAFAACAAFAACAA
RanPAC84.8189.4570.5476.3756.1164.8888.1491.77
RanDumb84.0189.0467.6473.7556.2964.6887.5291.43
SimpleCIL73.1280.0260.0465.8347.0756.5784.3588.95
AIR(M=5000)88.9492.5077.1380.0961.7768.2886.0989.89

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.

CIFAR-100ImageNet-RImageNet-ACUB-200
MMethodFAACAAFAACAAFAACAAFAACAA
5000PCR+NCM89.4493.1174.9278.8358.4665.6685.0789.56
5000EARL89.4493.3374.3779.7444.4855.2565.1776.27
5000aL-SAR89.1992.5874.8979.4747.7350.2380.2585.41
5000AIR88.9492.5077.1380.0961.7768.2886.0989.89
AuthorsOpenReview ↗

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-100ImageNet-RImageNet-ACUB-200
VariantFAAHDFAAHDFAAHDFAAHD
Adapter+ER56.160.5150.50627.110.2700.25519.310.1120.19531.490.3650.360
HGA only63.220.6310.61437.430.4230.37326.990.1920.18033.800.4260.388
MGC only87.730.5620.70473.430.4410.59955.210.5250.53682.850.5740.737
AIR88.940.6930.79677.130.5960.70061.770.6570.56286.090.6670.795

The results directly demonstrate AIR's effectiveness in maintaining geometric consistency.

Q1

We apply reservoir sampling after each update: the first MM examples are stored; for the nthn_{th} example (n>Mn > M), we sample jj uniformly from {1,...,n}\{1, ..., n\} and replace buffer entry jj only if j≤Mj ≤ M. 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-100ImageNet-RImageNet-ACUB-200
MethodFAACAAFAACAAFAACAAFAACAA
EASE-R84.2989.5554.0263.8741.4355.0379.5288.29
MOS-R84.4789.5366.0473.8342.5753.1375.3383.08
AIR86.6291.3469.6976.0347.4260.6079.4387.80

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.

Reviewer_YeWTOpenReview ↗
Rating
3(+1)
Confidence
2(−2)
Informativeness
1
Presentation
2
Soundness
2(−1)
Read review
Paper summary

This paper proposes Anchoring in Replay (AIR), a replay-based framework for online class-incremental learning (OCIL) with pretrained models. AIR consists of two components: Historical Geometry Anchoring (HGA), which encourages replayed old-class features to align with historical classifier directions, and Mixed-Geometry Calibration (MGC), which replaces classifier weights with class prototypes computed from the replay buffer after each task. Experiments on four benchmarks with different memory budgets demonstrate competitive performance against the evaluated baselines.

Critical strengths

The proposed approach is straightforward to implement and introduces limited architectural overhead. It operates under a single-pass online training protocol with a bounded replay buffer, making it practically relevant to streaming continual learning.

Critical weaknesses
  1. The two proposed components are closely related to widely ideas in continual learning. HGA essentially performs contrastive alignment between current features and historical classifier weights, while MGC directly replaces classifier weights with class-mean prototypes. Geometric preservation, relational knowledge distillation, and classifier alignment have already been extensively investigated, including in LUCIR, TPCIL (ECCV 2020), Relation Knowledge Distillation (AAAI 2021), Model Behavior Preserving (TNNLS), and Geometric Prototype Alignment (ICCV 2025). However, these closely related studies are not adequately discussed or distinguished from the proposed method. Consequently, the technical novelty beyond combining existing techniques remains unclear.

  2. Although the paper targets OCIL, a substantial portion of the main baselines consists of conventional class-incremental learning methods adapted to the online protocol, including DER, MEMO, EASE, MOS, and TagFex. While the paper includes some online continual learning baselines.

  3. The experimental results suggest that most of the improvement originates from the relatively simple prototype-based classifier calibration rather than the proposed historical geometry anchoring. For example, on CIFAR-100 with a memory budget of 1,000, MGC alone improves FAA from 55.88% to 84.93%, whereas adding HGA further improves it to only 85.62%.

Questions for authors

How does HGA fundamentally differ from previous geometric preservation, proxy-based distillation, and classifier alignment approaches, particularly LUCIR and related works?

2 comments
AuthorsOpenReview ↗

Response to Weaknesses and Questions (1/2)

Thank you for your comments. We address your concerns below.

W1 / Q1

LUCIR's feature loss (CVPR 2019) preserves each sample's previous feature direction. TPCIL (ECCV 2020), Relation KD (AAAI 2021), and MBP (TNNLS 2023) preserve graph topology, sample relations, and similarity/label rankings, respectively. GPA (ICCV 2025) initializes the current classifier from prototypes and aligns it with evolving class centroids during training.

HGA instead constrains old-class replay features using class directions from the previous classifier, rather than each sample's previous feature direction: it encourages alignment with the correct historical class direction and separation from the others. The advantage is to target old-class discrimination without penalizing feature adaptations useful for new tasks. And unlike GPA's alignment to evolving class centroids, this historical constraint remains fixed within a stage.

AIR further turns current calibration into historical supervision for the next stage. After training, MGC updates the classifier from mean features of old and new classes in the buffer; HGA uses a fixed copy to supervise old-class discrimination at the next stage. MGC's output therefore defines the targets for subsequent HGA supervision, while HGA constrains the representation used by the next MGC step. This dependence across stages distinguishes AIR from an independent combination of contrastive learning and prototype classification. We will clarify these distinctions in the related work.

HGA loss substitutions, M=5000. With MGC fixed, AIR achieves the highest FAA on all four datasets; on ImageNet-A, it exceeds the GPA and LUCIR loss substitutions by 8.38 and 5.57%, respectively.

CIFAR-100ImageNet-RImageNet-ACUB-200
MethodFAACAAFAACAAFAACAAFAACAA
MGC + LUCIR87.9291.6674.5877.9356.2060.7883.9788.91
MGC + GPA87.8291.6073.7177.1553.3959.9183.4288.44
AIR88.9492.5077.1380.0961.7768.2886.0989.89
AuthorsOpenReview ↗

Response to Weaknesses and Questions (2/2)

W2

We have added PCR+NCM (CVPR 2023), EARL (CVPR 2024), and aL-SAR (ICLR 2025) as additional OCL baselines. Their results will be incorporated into the experimental tables in the revised manuscript. Results for M=5000 are reported below:

CIFAR-100ImageNet-RImageNet-ACUB-200
MMethodFAACAAFAACAAFAACAAFAACAA
5000PCR+NCM89.4493.1174.9278.8358.4665.6685.0789.56
5000EARL89.4493.3374.3779.7444.4855.2565.1776.27
5000aL-SAR89.1992.5874.8979.4747.7350.2380.2585.41
5000AIR88.9492.5077.1380.0961.7768.2886.0989.89

At M=5000, AIR achieves the highest FAA and CAA on ImageNet-R, ImageNet-A, and CUB-200.

W3

We agree that HGA provides a modest FAA gain on CIFAR-100 at M=1,000 (0.69%). However, Table 3 and Appendix Table 5 show that, with MGC fixed, HGA improves FAA on ImageNet-A by 6.87/6.56/7.48% at budgets of 1000/5000/10000, respectively. At M=1000, it also improves CAA on ImageNet-A by 8.69% and FAA on CUB-200 by 3.67%.

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-100ImageNet-RImageNet-ACUB-200
VariantFAAHDFAAHDFAAHDFAAHD
Adapter+ER56.160.5150.50627.110.2700.25519.310.1120.19531.490.3650.360
HGA only63.220.6310.61437.430.4230.37326.990.1920.18033.800.4260.388
MGC only87.730.5620.70473.430.4410.59955.210.5250.53682.850.5740.737
AIR88.940.6930.79677.130.5960.70061.770.6570.56286.090.6670.795

The results support the respective geometric roles of HGA and MGC.

Reviewer_az69OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies online class-incremental learning with a frozen ViT, bottleneck adapters, a cosine classifier and replay. It argues that adapter updates make old-class features drift from their earlier class directions, while the classifier falls out of step with the new feature space. AIR handles this in two ways. HGA pulls old-class replay features toward the previous task's frozen classifier directions with a softmax over cosine similarities. MGC, at the end of each task, replaces each classifier weight with that class's mean feature computed from the replay buffer. Experiments cover CIFAR-100, ImageNet-R, ImageNet-A and CUB-200 with three buffer sizes and three seeds, plus ablations, confidence intervals, task-wise results, runtime and a ViT-Tiny comparison.

Critical strengths
  1. The empirical study is thorough: four datasets, three budgets, three seeds with confidence intervals, full task-wise tables, runtime and sensitivity analysis. The gains on ImageNet-A and ImageNet-R are clear. This matters because the results are easy to check and the reported numbers look reliable.

  2. The paper is easy to follow, the geometric framing is intuitive, and Figs. 1 and 2 explain the method well. The method itself is simple and cheap, so it would be easy to adopt as a strong replay baseline.

Critical weaknesses
  1. MGC is the class-mean (NCM/prototype) classifier from exemplars used in iCaRL, SCR and SimpleCIL. HGA is a proxy-anchoring loss close to PCR, with the proxies frozen from the previous task. Table 15 supports this: replacing HGA with PCR while keeping MGC lands within about 1.7 FAA points of AIR and gives higher CAA on CIFAR-100 and ImageNet-R. Table 3 shows that most of the gain comes from MGC alone. The "geometric inconsistency" story is essentially semantic drift plus classifier bias, and that line of work is not discussed, e.g. SDC (Yu et al., CVPR 2020), LDC (Gomez-Villa et al., ECCV 2024) and SLCA (Zhang et al., ICCV 2023). This feels fundamental. The contribution reads as a combination of known components, and the paper needs clearer positioning and evidence that HGA does something these methods don't.

  2. The strongest simple baselines are missing or misrepresented. RanPAC and RanDumb in Table 2 (about 9.5% FAA with FM around 94) are copied from PROL, where they come from an adaptation that clearly broke them; PROL's own joint variants reach around 80% on CIFAR-100. The RanDumb paper reports 86.8 on CIFAR-100 and 88.5 on CUB in a harder online split with no exemplars, and RanDumb is also missing from the references. SimpleCIL needs no training and no memory, yet reports 86.73 on CUB-200, above AIR's 86.09 at M=5000. The larger budgets are also big relative to the data. ImageNet-A has about 6K training images, so M=10000 holds the whole stream, and Table 11 shows the M=5000 and M=10000 runs are identical for the first 8 tasks. EASE-R and MOS-R are dropped at M=1000 instead of being run with class-balanced sampling. Against MOS-R at M=5000, the FAA gain is only 0.01 on CIFAR-100 and 0.13 on CUB. This is fixable: rerun RanPAC, RanDumb and SimpleCIL properly, use budgets that make sense relative to dataset size, and keep EASE-R and MOS-R at M=1000.

Questions for authors
  1. The learning rate is tuned per method and "the best setting" is reported. Was it selected on test accuracy or on held-out data?

  2. Do the Table 1 baselines (iCaRL, DER, MEMO, TagFex, SCR, ER-ACE) fine-tune the full backbone or use the same adapters as AIR? If the former, part of the gap may come from full fine-tuning versus PEFT rather than from AIR.

  3. How does AIR compare with SimpleCIL (streaming class means on the frozen backbone) and with correctly run RanPAC and RanDumb under the same protocol? This would change my assessment the most.

  4. Is there any direct evidence for the geometric story, such as the cosine between old-class features and their historical directions over tasks, with and without HGA and MGC?

2 comments
AuthorsOpenReview ↗

Response to Weaknesses and Questions (1/2)

Thank you for your constructive feedback. We address each concern below.

W1

Class-mean classification (iCaRL (CVPR 2017), SCR (CVPRW 2021), and SimpleCIL (IJCV 2025)), drift compensation (SDC (CVPR 2020) and LDC (ECCV 2024)), and classifier alignment (SLCA (ICCV 2023)) can correct predictions, but these operations alone do not turn the corrected directions into fixed constraints on subsequent learning. The contrastive objectives of SCR and PCR (CVPR 2023) also depend on sample features or proxies that evolve during training.

AIR's novelty lies in turning current calibration into historical supervision for the next stage: MGC updates old- and new-class directions, and HGA uses a fixed copy to supervise old-class discrimination at the next stage while allowing the current classifier to adapt. MGC's output thus defines the targets for subsequent HGA supervision, while HGA constrains the representation used by the next MGC step. This coupling across stages distinguishes AIR from an independent combination of contrastive learning and prototype classification. We will expand the related work to discuss these methods.

W2 / Q3

After rerunning the three rehearsal-free baselines, AIR (M=5000) achieves the highest FAA on the first three datasets, whereas RanPAC performs better on CUB-200. We will also add the RanDumb citation.

CIFAR-100ImageNet-RImageNet-ACUB-200
MethodFAACAAFAACAAFAACAAFAACAA
RanPAC84.8189.4570.5476.3756.1164.8888.1491.77
RanDumb84.0189.0467.6473.7556.2964.6887.5291.43
SimpleCIL73.1280.0260.0465.8347.0756.5784.3588.95
AIR(M=5000)88.9492.5077.1380.0961.7768.2886.0989.89

EASE-R and MOS-R, M=1,000. 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. AIR achieves the highest FAA and CAA on the first three datasets but trails EASE-R with a little margin on CUB-200.

CIFAR-100ImageNet-RImageNet-ACUB-200
MethodFAACAAFAACAAFAACAAFAACAA
EASE-R84.2989.5554.0263.8741.4355.0379.5288.29
MOS-R84.4789.5366.0473.8342.5753.1375.3383.08
AIR86.6291.3469.6976.0347.4260.6079.4387.80

Fixed 10% memory budget. Retaining only 10% of the training samples and using class-balanced sampling, AIR achieves the highest FAA on all four datasets and highest CAA on three.

CIFAR-100ImageNet-RImageNet-ACUB-200
MethodFAACAAFAACAAFAACAAFAACAA
EASE-R85.4490.0458.2466.6537.2652.1579.4788.28
MOS-R85.8690.4267.8275.0939.7252.1775.1682.71
AIR89.1992.5475.1779.3140.4257.4379.5287.88
AuthorsOpenReview ↗

Response to Weaknesses and Questions (2/2)

Q1

For each dataset, we select the learning rate from the candidate values based on the final FAA.

Q2

iCaRL, DER, SCR, ER-ACE, MEMO, and TagFex use the ViT-B backbone with adapters.

Q4

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-100ImageNet-RImageNet-ACUB-200
VariantFAAHDFAAHDFAAHDFAAHD
Adapter+ER56.160.5150.50627.110.2700.25519.310.1120.19531.490.3650.360
HGA only63.220.6310.61437.430.4230.37326.990.1920.18033.800.4260.388
MGC only87.730.5620.70473.430.4410.59955.210.5250.53682.850.5740.737
AIR88.940.6930.79677.130.5960.70061.770.6570.56286.090.6670.795

The results directly support the respective geometric roles of HGA and MGC.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.