acceptodds
Under review as a conference paper at ICLR 2027

Memory as the Backbone: A Dual-Store Associative Memory Architecture for Visual Recognition

Abstract

We investigate whether explicit memory can serve as the organizing principle of a general-purpose visual backbone. Inspired by the interaction between working and long-term memory in human cognition, we introduce DuSAM (Dual-Store Associative Memory), a hierarchical memory-native visual backbone coupling image-local working memory (WM) with a fixed-capacity persistent semantic memory (PM). WM organizes local visual evidence into compact, spatially anchored states that are updated and transported across stages, then read back to refine dense features. PM consolidates semantic evidence from training images and annotated regions into prototypes, whose retrieved content enriches the same working states. DuSAM achieves competitive performance relative to representative, similarly sized backbones on ImageNet-1K classification, ADE20K semantic segmentation, and COCO object detection and instance segmentation. Ablations show contributions from both stores, with PM particularly benefiting class groups with fewer annotated training pixels. At , the complete DuSAM-based UPerNet segmentor achieves the throughput of the evaluated DeiT-S/16 counterpart while using less GPU memory. Memory analyses reveal spatially selective WM access and semantic associations between retrieved prototypes and training regions. Class-targeted removal and restoration further demonstrate that category-specific experience can be replenished through state updates without changing network parameters. These results establish explicit memory as a viable organizing principle for visual backbones and provide a foundation for jointly designing contextual computation and addressable semantic memory.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Associative Memory, Visual Backbone

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_mEh3OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents DuSAM, a hierarchical visual backbone that combines image-local working memory with a persistent bank of class-specific prototypes collected from annotated training data. At deeper stages, tokens retrieve semantic prototypes, bind them to transported working states, and read the states back to refine image features. The evaluation covers ImageNet classification, ADE20K semantic segmentation, and COCO detection and instance segmentation, along with memory ablations, high-resolution profiling, and targeted memory removal and restoration.

Critical strengths
  1. Binding retrieved semantic prototypes to transported, spatially anchored working states gives the two memory stores a clear role within the backbone. The retrieved information modifies intermediate features rather than only a task-specific prediction head.

  2. The component ablations are informative. In particular, keeping the semantic-address supervision while disabling persistent memory helps distinguish retrieval gains from an auxiliary classification loss.

  3. The memory interventions make the stored content inspectable. The class-targeted removal and restoration examples show that predictions can change through memory updates without changing the network weights.

Critical weaknesses
  1. I am not convinced that the experiments establish DuSAM as a competitive general-purpose backbone. Tables 1 to 3 compare many older architectures, but omit strong published efficient backbones. In particular, MSVMamba-Micro [1] reports 79.8% ImageNet accuracy with 12M parameters, 45.1 single-scale mIoU with a 42M-parameter UPerNet, and 43.8 box AP / 39.9 mask AP with a 32M-parameter Mask R-CNN. DuSAM reports 77.8% with 20.5M parameters, 42.90 mIoU with 49.9M, and 41.9 / 38.1 AP with 39.5M, respectively. These are published results with potentially different training recipes, not controlled head-to-head measurements. Still, excluding a smaller, already published backbone with stronger results on all three tasks leaves the proposed architecture's advantage unclear. A matched comparison or a careful account of the protocol differences is needed. The omission also extends to more recent backbone families such as MambaVision.

  2. The transfer evidence is narrower than the general-purpose framing suggests. Appendix A.4 states that persistent memory is rebuilt for each downstream task's label vocabulary. Reusing pretrained weights across several supervised recognition tasks is demonstrated, but transfer of the stored semantic experience is not. I would like to see a frozen-feature or memory-free transfer analysis, or a more limited claim.

  3. The benefits of the specific dual-store binding over existing visual memory mechanisms are not isolated. Associative Transformer [2] already uses explicit associative memory for visual processing, while MCIBI [3] retrieves class-level dataset memory using predicted semantic addresses. Table 4 establishes gains over local-only variants, but gives no matched comparison to either memory-based alternative.

  4. The efficiency comparison in Fig. 2 is against global-attention DeiT-S alone, leaving the advantage over modern efficient backbones unresolved. The targeted restoration results in Figs. 6, 10, and 11 are qualitative. Average recovery across classes is not reported.

[1] Shi et al., Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model, NeurIPS 2024.
[2] Sun et al., Associative Transformer, CVPR 2025.
[3] Jin et al., Mining Contextual Information Beyond Image for Semantic Segmentation, ICCV 2021.

Questions for authors
  1. When transferring from ImageNet to ADE20K or COCO, which PM writer, reader, and semantic-address parameters are retained, and which are reinitialized along with the class-specific prototype bank (Appendix A.4)?

  2. Can the authors clarify the pretrained checkpoints and augmentation protocols behind the Table 2 and 3 baselines? Are results available under a matched setting for a more recent efficient backbone such as MSVMamba-Micro?

6 comments
AuthorsOpenReview ↗

Response to Reviewer mEh3

Dear Reviewer mEh3, thank you sincerely for your valuable review, and for pointing us to specific recent work. We are truly glad that the paper deepened your understanding of the problem and that you saw a clear role for each memory store within the backbone. Your comments helped us sharpen how this new architecture is positioned and presented. We have addressed every point below and updated the revised manuscript accordingly, and we hope this answers your questions.

AuthorsOpenReview ↗

W1/Q2 (Recent efficient backbones, matched setting, and baseline protocol).

Matched comparison. As requested, we added MSVMamba-Micro and MambaVision-T on all three tasks. For ADE20K and COCO, we trained both ourselves under the same protocol as every other baseline, so Tables 2–3 now provide the matched setting you asked about rather than published numbers. (Tables 1–3, pp. 5–7 of the revised manuscript; highlighted in red.)

Baseline protocol (Q2). Table 1 lists published ImageNet-1K results with each model’s augmentation recipe. For Tables 2–3, baselines are initialized from publicly released ImageNet-1K checkpoints and trained by us under the same segmentation and detection settings as DuSAM, including the newly added MambaVision-T and MSVMamba-Micro. (Appendix A.1.2, p. 15 of the revised manuscript; highlighted in red.)

Stronger DuSAM results. With the same augmentation recipe as these baselines, DuSAM improves to 79.1% top-1 (from 77.8%), 43.33/44.50 SS/MS mIoU on ADE20K, and 42.6/38.9 box/mask AP on COCO (DuSAM⋆).

Where DuSAM stands. DuSAM is, to our knowledge, the first general-purpose visual backbone to build its feature hierarchy on explicit memory, with no conventional self-attention or state-space blocks. In this first instantiation, under the matched protocol, it already matches Swin-T in COCO box AP (42.6) with 17% fewer parameters, exceeds DeiT-S, PVT-S, and PoolFormer-S24 by 2.2–2.5 box AP, exceeds PVT-S and PoolFormer-S24 on ADE20K by 0.9 and 1.6 mIoU, and scales to high resolution as efficiently as MSVMamba-Micro (both retain 12.3% of their throughput from 512² to 2048²). On small objects, its box and mask AP are the second highest among all compared models, above ConvNeXt-T, Swin-T, and MSVMamba-Micro (Table 10). MSVMamba and MambaVision build on several generations of refinement of visual state-space models and currently lead in accuracy; DuSAM reaches the range of established backbones in its very first form, with clear room for the same kind of refinement. Beyond accuracy, it offers what these backbones do not: semantic evidence stored explicitly, which can be inspected, diagnosed, and replenished from five annotated regions without changing weights (Sec. 4.4, p. 10).

AuthorsOpenReview ↗

W2/Q1 (Transfer: retained parameters and scope of the general-purpose claim).

Retained and reinitialized parameters (Q1). When transferring from ImageNet-1K to ADE20K or COCO:

  • Retained: all task-independent backbone parameters, including WM construction, cross-stage transport, and readback; the PM reader (content query) and PM-to-WM binding projections; the task-independent layers of the semantic-address heads; and the PM writer’s descriptor construction and EMA update rule.

  • Reinitialized: the class-specific prototype bank, which is rebuilt from downstream annotated regions (ADE20K) or instances (COCO); the semantic-address output layers for the new label vocabulary (150 or 80 classes); and the task-specific segmentation and detection heads.

(Sec. 3.2, pp. 4–5, and Appendix A.1.2, pp. 14-15 of the revised manuscript; highlighted in red.)

Scope of the claim. DuSAM is a new memory-native backbone, not a memory module added to an existing CNN, Transformer, or SSM. Our general-purpose claim concerns this architecture and its pretrained representations, which transfer across classification, segmentation, and detection, as with other general-purpose backbones. The learned memory operations are reused across tasks, while persistent semantic content is accumulated for each task’s label vocabulary, as stated in Appendix A.4.

Frozen-feature and memory-free transfer. We added the analysis you asked for on ADE20K. Frozen-feature: with all backbone weights frozen and PM disabled, DuSAM reaches 36.70/37.92 SS/MS mIoU, so its pretrained representations support dense prediction on their own. Memory-free: with fine-tuning and no PM, it reaches 41.75/42.97. Task-adapted PM then adds a further 1.15 points on both metrics (42.90/44.12). The full analysis, including a preliminary test of directly reusing the ImageNet prototype bank, is given in Table 12. Together, these results show that DuSAM learns transferable visual representations as a memory-native backbone, with persistent memory supplying additional task-relevant semantic evidence. (Appendix A.3.6 and Table 12, p. 22 of the revised manuscript; highlighted in red.)

Reviewer_13WHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DuSAM, a hierarchical visual backbone combining image-local working memory (WM) with persistent semantic memory (PM). WM aggregates image features into spatially anchored states, transports these states across stages, and provides context through token-dependent readout. PM stores class-specific prototypes accumulated from annotated training images or regions. Tokens retrieve prototypes using content similarity and predicted category probabilities, and the retrieved information is incorporated into WM before being read back into image features. Experiments on ImageNet-1K, ADE20K, and COCO evaluate classification, semantic segmentation, object detection, and instance segmentation. The paper additionally examines memory configurations, inference-time interventions, and class-targeted memory removal and restoration.

Critical strengths
  1. The paper develops a coherent dual-store architecture that integrates current-image context with persistent semantic experience. Spatially anchored working states, cross-stage transport, and a shared prototype bank provide a concrete implementation of this design. Evaluation on semantic segmentation, object detection, and instance segmentation extends the evidence for memory-centered visual backbones beyond image classification. The competitive results across these tasks constitute a meaningful contribution.

  2. Table 4 provides useful evidence for several design choices, including the contribution of PM, cross-stage WM transport, and PM sharing. On ADE20K, the complete model achieves 42.90 mIoU, compared with 40.45 without both memories and 41.75 without PM. Retaining the semantic-address heads and auxiliary losses while removing PM yields 41.55, helping distinguish the benefit of retrieved content from additional supervision. The capacity analyses, update/retrieval sensitivity study, and inference-time interventions further examine how the model uses its memory states.

Critical weaknesses

Although Table 4 supports the usefulness of the components and several cross-stage design choices, the benefit of the specific PM-to-WM interaction remains insufficiently isolated. Removing PM also removes access to historical semantic information, while separating the PM banks changes cross-stage sharing rather than the fusion pathway. These experiments therefore do not determine whether routing retrieved content through WM is preferable to directly incorporating it into token features while retaining WM for image-local context. Table 5 tests inference-time perturbations of the trained model, rather than this alternative architecture. Since the proposed cooperation between the two stores is central to the architectural contribution, this distinction matters to my assessment. Repeated-run statistics would also help assess the smaller gains from WM transport and PM sharing.

The evidence for reliable class-targeted memory replenishment is limited. The removal analysis in Section 4.4 provides useful quantitative evidence of class-directed effects. However, restoration using five annotated training regions or instances is demonstrated through visual examples in Figure 6 and Appendix A.3.5 (Figures 10–11), without aggregate restoration performance across classes or variability across choices of the five examples. The results therefore establish the possibility of restoring predictions in the illustrated cases, but do not adequately characterize the reliability or extent of this capability. This limits the evidence supporting memory modification without weight updates as a practical benefit of the architecture.

Questions for authors
  1. Could the authors compare the proposed PM-to-WM binding against direct PM-to-token fusion, while retaining WM and keeping the local backbone, memory capacity, auxiliary supervision, and training settings comparable? This would help distinguish the benefit of the proposed interaction pathway from the benefit of accessing PM content. Are the improvements from WM transport and shared PM consistent across independent runs?

  2. How are the five examples used for restoration selected? Please report target-class performance before removal, after removal, and after restoration across evaluated classes, together with variability across sampled example sets and effects on non-target classes. Please also clarify the units and aggregation procedure for the reported non-target IoU change of 0.020.

  3. Figure 2 establishes a high-resolution efficiency advantage over the evaluated global-attention DeiT-S/16 baseline, and Table 2 reports FLOPs and training memory for additional backbones. To clarify the practical efficiency trade-off, how does high-resolution throughput compare with a hierarchical backbone under the same profiling setup, and what training-time overhead is introduced by the separate writer path?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 13WH

Dear Reviewer 13WH, thank you sincerely for such a thoughtful review. It means a lot to us that you found DuSAM a coherent dual-store architecture and its results across tasks a meaningful contribution. Your questions were insightful, and the experiments they inspired further confirm the value of the dual-store design. We have run every experiment you asked for, answered each question below, and updated the revised manuscript accordingly, and we hope this resolves your concerns.

AuthorsOpenReview ↗

Q1/W1(PM-to-WM binding vs. direct PM-to-token fusion; consistency across runs):

PM-to-WM binding. To isolate the benefit of PM-to-WM binding, we compared it against direct PM-to-token fusion while retaining WM and keeping the local backbone, memory capacity, auxiliary supervision, and training settings unchanged (identical parameter count, 19.43M). PM-to-WM binding achieves 42.90 SS mIoU on ADE20K versus 42.03 for direct fusion (+0.87), with the largest gain in the Low-frequency group (34.81 vs. 31.59, +3.22). Integrating persistent semantic evidence into WM before readback is therefore more effective than adding it directly to token features. (Table 5 and the “Mechanism comparison” paragraph in Section 4.4, p. 8 of the revised manuscript; highlighted in green.)

Repeated-run stability. We repeated the Standard model and the three variants with an independent seed. Across the two runs, Standard achieves 42.90/42.74 SS mIoU, compared with 42.48/42.27 for w/o WM transport, 42.39/42.28 for Separate PM (2+2), and 42.37/42.46 for Separate PM (4+4). Standard outperforms all three variants in both runs (mean gains of 0.44, 0.48, and 0.40), so the benefits of WM transport and a shared PM are consistent across independent runs, with small run-to-run differences (at most 0.21). The role of transport is further supported by Table 6 (p. 9 of the revised manuscript): disabling it in the trained model lowers mIoU by 4.71.

AuthorsOpenReview ↗

Q2/W2(Reliability of class-targeted memory restoration):

We have added a quantitative removal-and-restoration evaluation across all 150 ADE20K classes on the full validation set.

How the five examples are selected. Five annotated training regions of the target class are sampled uniformly at random, with three independent draws per class, and written back with the training-time writer while network parameters stay fixed.

Performance before removal, after removal, and after restoration. Across all 150 ADE20K classes, mean target-class IoU goes from 42.90 (Standard) to 42.27 (Remove) and back to 42.76 (Restore-5), recovering 78% of the lost IoU. Reliance on PM varies across classes. Removal leaves the network parameters intact, so some classes change little; this is expected, because PM is designed to complement learned parameters and image-local WM, and matters most for classes that these capture less well. For classes whose IoU drops by at least 1 pp, where recovery can be measured reliably, mean IoU goes from 43.00 to 39.91 and back to 42.24, with a mean per-class recovery of 77% (67–92% across the four frequency groups); a single region already recovers 60%, and selecting the five largest regions raises recovery to 93%. Among these classes, drops are largest in the two rarer frequency groups (mean drops of 3.1 and 4.7 pp, versus 2.2 and 2.0 pp), consistent with the PM benefits in Table 4.

Variability across sampled example sets. Restoration is stable across different choices of the five examples: the mean within-class SD of restored IoU over the three draws is only 0.41 pp.

Effects on non-target classes; units and aggregation of 0.020. The reported 0.020 is in IoU percentage points. For each target-class removal, we average the absolute IoU changes of the other 149 classes relative to Standard, then average over all target classes with equal weight. After restoration, the same quantity is only 0.11 pp, so both removal and restoration are highly selective.

Practical benefit. We also added class-level memory diagnostics, which reveal both the correct predictions that individual PM entries support and the specific confusions they reinforce, such as water predicted as sea. Together, these results establish a key benefit of DuSAM’s explicit memory: stored semantic evidence can be inspected, selectively modified, and replenished from a few annotated regions without changing network parameters.

(Sec. 4.4 and Figs. 5–6, p. 10; Table 11 and Appendix A.3.5, pp. 20–23 of the revised manuscript; highlighted in green.)

Reviewer_FuGxOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary
  1. The paper proposes DuSAM, a 4-stage visual backbone that organizes computation around two explicit memory stores borrowed from cognitive psychology.
  2. Working Memory aggregates token features via Sinkhorn-balanced soft assignments, is transported across stages, and is read back into dense features through a gated residual.
  3. Persistent Semantic Memory is a per-class prototype bank shared across Stages 3–4, written with training-region descriptors via EMA, retrieved by content similarity plus a learned semantic-address prior, and bound back into Working Memory.
  4. A class-targeted removal/restoration experiment shows it can be edited without weight updates.
Critical strengths
  1. This paper maintains a disciplined division of labor (the WM is rebuilt per image, while the PM persists and is frozen at inference). This constitutes more than a rebranding of attention.
  2. The linear-scaling argument is correct and empirically supported. Replacing pairwise global attention with K-anchor read/write yields O(NKC), and the efficiency result (Fig. 2: 2.36× throughput and 59.5% memory reduction vs. DeiT-S/16) is convincing practical claim.
  3. The memory-editing experiment is the most novel result. Class-targeted PM removal combined with restoration from five annotated regions (Fig. 6) demonstrates a capability that parameter-only backbones simply do not possess.
Critical weaknesses
  1. DuSAM's 77.8% top-1 sits below DeiT-S (79.8), Swin-T (81.2), ConvNeXt-T (82.1), and PoolFormer-S24 (80.3) at comparable or smaller parameter counts. The paper claims it is because of "without Mixup/CutMix", but I am doubtful why DuSAM does not adopt these strategies. The comparison table mixes augmentation recipes across rows, so the reader cannot tell whether the 2-point gap to DeiT-S is due to architecture or the training recipe.
  2. DuSAM is similar to LookupViT [1], and the authors should clarify how the two differ.
  3. The related work cites Associative Transformer and Vision Hopfield Memory Networks and claims that DuSAM's contribution lies in "how transient image context and accumulated semantic experience should cooperate across feature scales." However, the paper never provides an empirical comparison against either method.
  4. The most recent model we found in experience, ViM, was published in 2024, and the authors should cite more recent papers from 2025 and 2026.
  5. The 2.36× efficiency claim is baseline-fragile. DeiT-S/16 is a global-attention transformer and is the expected loser at high resolution. Beating it on throughput is not surprising and does not establish that the memory architecture is efficient relative to, e.g., Swin (windowed attention), ViM (linear SSM), or ConvNeXt (local CNN). The paper should include at least one linear-complexity baseline in the high-resolution comparison.
Questions for authors

None

7 comments
AuthorsOpenReview ↗

Response to Reviewer FuGx

Dear Reviewer FuGx, thank you sincerely for such a careful and constructive review. We are truly encouraged that you recognized the distinct roles of DuSAM’s two memories and saw memory editing as a capability unique to this new kind of backbone. Your five points were precise and actionable; each led to a new experiment and made the paper considerably stronger. We have addressed all of them below and updated the revised manuscript accordingly, and we hope this resolves your concerns.

AuthorsOpenReview ↗

W1 (ImageNet accuracy and training recipe):

Many thanks for this suggestion. With limited time and compute before submission, we trained DuSAM with a lighter recipe. We have now retrained it with the DeiT-style recipe used by most baselines (Mixup, CutMix, random erasing, and stochastic depth), with all other settings unchanged (details in Appendix A.1.2, p. 15 of the revised manuscript; highlighted in blue).

Top-1 accuracy improves from 77.8% to 79.1% with the same 20.5M parameters, within 0.7 points of DeiT-S (79.8%, 22.1M) and without any self-attention. This controlled comparison attributes most of the original 2-point gap to the training recipe rather than the architecture. We also note that Swin-T and ConvNeXt-T have 28.3M and 28.6M parameters, about 40% more than DuSAM. The stronger pretraining carries over to downstream tasks (Tables 1, 2, 3, and 10 of the revised manuscript; highlighted in blue.) :

ImageNet top-1ADE20K SS / MSCOCO box / mask AP
DuSAM (submitted)77.842.90 / 44.1241.9 / 38.1
DuSAM⋆79.143.33 / 44.5042.6 / 38.9

These results show that a backbone built around explicit, addressable memory, without self-attention or SSM blocks, achieves strong general visual recognition. It also offers what parameter-only backbones do not: as you noted, memory can be edited without changing weights. We now go further and quantify this across all 150 ADE20K classes: writing back just five annotated regions restores 78% of the lost IoU on average, and 93% for the classes that rely most on PM when the five largest regions are used, while other classes change by at most 0.11 pp (Sec. 4.4, "Class-targeted memory updates", p. 10; highlighted in green).

AuthorsOpenReview ↗

W2 (Difference from LookupViT):

Method differences. DuSAM differs from LookupViT in three ways. (1) Backbone: LookupViT retains Transformer blocks and processes its compressed tokens with self-attention; DuSAM uses no self-attention. (2) Aggregation and readback: LookupViT initializes compressed tokens by downsampling, gathers information through cross-attention, and reuses the same attention weights for readback; DuSAM assigns features to spatially anchored memory states by content similarity and anchor distance, balanced by Sinkhorn normalization, and recomputes content-based read weights from the updated states. (3) Scope of memory: both token sets in LookupViT come from the current image; DuSAM additionally binds long-term, class-specific prototypes accumulated across training images (PM) into its working states. (Section 2, pp. 2–3 of the revised manuscript; highlighted in blue.)

Experimental comparison. We further compared the two aggregation designs within DuSAM, replacing our spatially anchored Sinkhorn writer with pooled-query cross-attention inspired by LookupViT, while keeping PM, WM capacity, readback, and all training settings unchanged. Our writer achieves 42.78 SS mIoU on ADE20K versus 41.98 (+0.80), with gains of 2.35 and 2.25 points on the two rarer class groups (Lowest and Low). (Table 5 and the “Mechanism comparison” paragraph in Section 4.4, p. 8 of the revised manuscript; highlighted in green.)

Comments on the paper

1 comment

We sincerely thank all three reviewers for the time and care they devoted to our paper. Your comments were specific and constructive, and we believe that working through them has made this paper substantially more solid.

What this work is: DuSAM is not an incremental variant of an existing design. It is, to our knowledge, the first general-purpose visual backbone whose feature hierarchy is organized around explicit memory, with no self-attention and no state-space blocks. It shows that a memory-native architecture can serve image classification, semantic segmentation, and object detection. We are grateful that the reviews recognize this: the design “constitutes more than a rebranding of attention” and offers “a capability that parameter-only backbones simply do not possess” (FuGx), forms “a coherent dual-store architecture” whose results “constitute a meaningful contribution” (13WH), and gives the two memory stores “a clear role within the backbone” (mEh3). All three reviewers found the paper clear and easy to follow, with no substantive soundness concerns from Reviewer FuGx and a substantially deeper understanding gained by Reviewer mEh3.

Stronger results: Following the reviewers’ suggestions, we retrained DuSAM with a standard augmentation recipe. Performance improves on all three tasks: ImageNet-1K top-1 from 77.8% to 79.1%, ADE20K from 42.90 / 44.12 to 43.33 / 44.50 SS/MS mIoU, and COCO from 41.9/38.1 to 42.6/38.9 box/mask AP, with the same 20.5M parameters.

What we added:

  1. Controlled comparisons with the memory mechanisms of MCIBI, AiT, and LookupViT inside DuSAM (Reviewers 13WH, mEh3, FuGx).

  2. Quantitative memory removal and restoration across all 150 ADE20K classes, with class-level memory diagnostics (Reviewers 13WH, mEh3).

  3. High-resolution efficiency against 10 baselines, with kernel-level profiling and the training cost of the memory writer (Reviewers FuGx, 13WH, mEh3).

  4. Recent baselines (MambaVision-T, MSVMamba-Micro) on all three tasks under a matched setting (Reviewers mEh3, FuGx).

  5. A frozen-feature and memory-free transfer analysis (Reviewer mEh3).

  6. Expanded related work, further implementation details, and repeated-run results (Reviewers FuGx, mEh3, 13WH).

Every point raised by the reviewers has been addressed, and all corresponding changes are already in the revised manuscript, highlighted in color. Detailed responses follow under each review. We hope the revision resolves your concerns, and we would be glad to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.