acceptodds
Under review as a conference paper at ICLR 2027

Spatially Adaptive Convolutions via Attention-Routed Expert Mixtures

Abstract

Convolutional autoencoders have been widely used for image reconstruction due to the local receptive fields and inductive biases of convolutional operators. However, conventional convolution applies the same learned transformation across spatial locations, limiting the ability of a single operator to adapt to heterogeneous local structures. We investigate whether this limitation can be addressed by learning a spatially varying composition of distinct convolutional transformations. We propose a mixture-of-experts (MoE) based adaptive convolutional autoencoder in which a fixed bank of structurally heterogeneous three-dimensional convolutional experts provides distinct transformations, while a spatial routing mechanism determines their contributions at different locations. In the encoder, localized window representations are processed by an attention module to produce spatially varying expert weights, which are interpolated to form a dense routing field. The resulting formulation retains the parameters of the individual experts while allowing their effective composition to vary spatially. We evaluate the framework on three-dimensional magnetic resonance imaging (MRI) reconstruction using the IXI-T1 dataset, focusing on the behavior of the learned experts and their spatial routing. The individual experts exhibit systematically different responses to image structure and spatial frequency, while routing redistributes their contributions across the feature map and retains a soft, non-uniform mixture rather than hard expert assignment. These observations are consistent with functional specialization emerging among heterogeneous convolutional operators through spatial routing.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Mixture of Experts, Dynamic Convolutions, Attention Mechanisms, Image Reconstruction, Medical Image Analysis

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 2 reviews, above 0% of the others.
Reviewer_FxwEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a spatially adaptive convolution in which the outputs of five heterogeneous 3D convolutional experts are combined with per-location softmax weights, predicted by a window-attention router in the encoder and a small convolutional router in the decoder. The formulation is embedded in an autoencoder trained to reconstruct T1-weighted MRI volumes from IXI. The paper focuses on characterizing expert behavior: routing maps, correlations between expert responses and image-structure measures, responses to sinusoidal inputs, and expert-removal ablations. It concludes that the experts show differentiated behavior consistent with functional specialization under a soft, spatially varying mixture, and reports higher reconstruction fidelity than a MONAI autoencoder and a 3D ViT.

Critical strengths

A good control for routing is included, and the analyses are qualified carefully.

  • The paper compares adaptive routing against uniform aggregation of the same expert bank (Table 5, Appendix A.4), which is the correct control for isolating the effect of routing.
  • It explicitly frames the structural and frequency analyses as correlational rather than causal, and reports that expert roles vary across random seeds (Appendix C.7) rather than presenting a single run as representative.
Critical weaknesses

1. There is no evidence that spatial routing improves the model, and the reconstruction comparisons are not interpretable.

  • Routing adds little over uniform aggregation: The adaptive five-expert model reaches 29.04 dB PSNR and 0.9596 SSIM, against 28.84 dB and 0.9586 for uniform aggregation of the same experts (Table 5). No variance across seeds is reported for these comparisons, so there is no way to establish statistical significance.
  • Simpler configurations match or exceed the full model: A single expert without routing (E2, a standard 1×3×3 convolution) reaches 29.14 dB and 0.9593 SSIM. Removing E0 or E3 from the full model improves PSNR to 29.68 and 29.64 dB.
  • The learned routing is close to uniform: The normalized routing entropy of the full model is 0.957 (Table 2). In the encoder, the colorbars of Fig. 8 indicate that each expert's weight varies only within a narrow range across the slice (e.g., approximately 0.59–0.62 for expert 1). This amounts to an approximately fixed global weighting rather than meaningful spatial adaptation.
  • yet the cost is substantial: Relative to the MONAI autoencoder, inference time is about 13× higher (347.79 vs. 26.35 ms) and peak memory about 7× higher (2.276 vs. 0.313 GB) (Table 4).
  • The baseline comparisons are confounded:
    • The task is identity reconstruction, where achievable fidelity is governed mainly by bottleneck capacity. Latent dimensions and compression ratios are not reported for any model.
    • As described, the decoder upsamples only once, by a factor of two in-plane (Sec. 3.2, 3.2.3). This implies a latent with only four times fewer spatial positions than the input, which may explain the PSNR gap over the MONAI autoencoder and the ViT.
    • All models are trained for only 20 epochs.
  • The most relevant baselines are absent: The paper cites dynamic and conditional convolutions (CondConv, Dynamic Convolution, ODConv, KAConv) and spatial mixture-of-experts (Dryden & Hoefler, 2022), but compares against none of them.

Altogether, these make the significance of the contribution unclear.

2. The specialization claim is largely built in by design, and the manuscript contains internal inconsistencies.

  • Differentiation is imposed architecturally: The experts differ by construction. In the decoder, E0 and E1 operate on low-pass and high-pass components respectively, and the others differ in kernel orientation (1×3×3 vs. 3×1×1). Distinct structural and frequency responses are therefore expected from the architecture and are not evidence of emergent functional specialization. Conversely, Table 1 lists encoder experts E1 and E4 both as 1×1×1 convolutions, so the bank is not fully heterogeneous.
  • The routed-correlation analysis is mechanical: The routed contribution is defined as Ck(p)=wk(p)Ak(p)C_k(p) = w_k(p) A_k(p). Changes in its correlation with structural maps follow directly from multiplying by a spatially varying weight. They do not show that the router selects experts based on image structure.
  • Expert roles are not reproducible: The paper reports that an expert's structural association in one run need not hold in another (Appendix C.7). Only aggregate redundancy is stable (0.58–0.63), so specialization at the level of individual experts is not established.
  • The formulation is close to existing operators: As shown in Appendix A (Eq. 12), the operation reduces to a spatially varying kernel Keff(p,δ)=∑kwk(p)Kk(δ)K_{\text{eff}}(p,\delta) = \sum_k w_k(p) K_k(\delta). This is a spatially varying form of dynamic convolution, closely related to the cited KAConv, pixel-adaptive convolution, and spatial mixture-of-experts.
  • The manuscript contains internal inconsistencies:
    • The same model is reported with MSE 0.00147 and SSIM 0.955 in Table 3, but MSE 0.001310 and SSIM 0.9596 in Table 5, at identical PSNR (29.04 dB).
    • Table 1 lists decoder E3 as 3×1×1 and E4 as 1×1×1; Algorithm 2 assigns them the reverse.
    • The encoder is described as having three blocks (Sec. 3) and as two parts of two blocks each (Sec. 3.1).
    • Algorithm 1 subtracts the spatial mean of the routing logits, whereas Eq. 20 subtracts the mean across experts.
    • Figure 13 is captioned as model reconstructions but shows routing maps, and complexity results are attributed to Table 2 rather than Table 4.
    • Template placeholder text remains in the Author Contributions and Acknowledgments sections.
    • Channel widths, latent size, window size, and the number of attention heads are not reported.
Questions for authors
  1. Bottleneck capacity: What are the latent dimensions and compression ratios of the proposed model, the MONAI autoencoder, and the ViT? How do the models compare at matched compression?
  2. Value of routing: Given that a single unrouted expert (E2) matches the full adaptive model, what does spatial routing contribute? Please report mean ± standard deviation over seeds for the adaptive, uniform, and single-expert configurations.
  3. Related operators: How does the method compare with CondConv, Dynamic Convolution, ODConv, and spatial mixture-of-experts at matched parameter count and compute?
  4. Inconsistencies: Please reconcile the reconstruction metrics in Tables 3 and 5, the decoder expert definitions in Table 1 and Algorithm 2, and the logit normalization in Algorithm 1 and Eq. 20.
  5. Downstream relevance: Does the learned expert composition provide any benefit on a task other than identity reconstruction, such as segmentation or super-resolution?
1 comment
AuthorsOpenReview ↗

Clarification of Routing Contributions, Model Comparisons, and Scope

We thank the reviewer for the careful assessment and for recognising the uniform-aggregation control and the qualified treatment of our analyses. We agree that the results warrant a narrower interpretation and address the five questions below.

1. Bottleneck capacity and reconstruction comparisons

We agree that comparisons with MONAI and the 3D ViT are difficult to interpret without reporting latent dimensions and compression ratios. The decoder's 2× in-plane upsampling alone does not establish the overall compression ratio. Further, training all models for 20 epochs does not guarantee convergence. We therefore interpret these results only under the reported training setup, not as capacity-controlled evidence of architectural superiority.

Our task is volumetric MRI autoencoding: fully observed IXI T1-weighted volumes serve as both inputs and targets. It does not model undersampled acquisition or k-space data consistency, and we make no claims of clinical utility. We use it as a testbed for studying spatially routed heterogeneous 3D operators on structured volumetric data.

2. Value of spatial routing

We acknowledge that adaptive routing yields a modest gain over uniform aggregation: 0.20 dB PSNR and 0.0010 SSIM. E2-only is also competitive (29.14 dB, 0.9593 SSIM), while removing E0 or E3 improves reconstruction fidelity. Thus, the results do not establish that the full five-expert bank is optimal or that routing is necessary for this task. However, removing E2 substantially degrades the full model (24.87 dB), indicating its importance within the bank. A single expert can achieve competitive fidelity, but cannot exhibit spatial composition among multiple operators. Our focus is on characterising that composition, not claiming consistent reconstruction superiority.

We agree that mean ± standard deviation across seeds for adaptive, uniform, and E2-only configurations would strengthen the comparison. Appendix C.7 examines variability in expert-level routing patterns and structural associations; it does not establish the statistical significance of the reconstruction gain. We therefore make no such claim.

We also acknowledge the computational overhead relative to MONAI (347.79 vs. 26.35 ms inference; 2.276 vs. 0.313 GB peak memory). We describe the model as parameter-efficient, not computationally efficient. Although high normalized entropy (0.957) indicates diffuse expert weighting rather than selective routing, entropy alone does not establish spatial uniformity. The narrow routing ranges nevertheless indicate relatively modest, soft modulation, and we have qualified our description accordingly.

3. Relationship to related operators

We agree that our formulation is closely related to spatially varying dynamic convolution and mixture-of-experts approaches. Appendix A expresses the weighted operation as a spatially varying effective kernel; we do not claim this general formulation as fundamentally new. Our focus is on characterising the spatial contributions of heterogeneous 3D operators and their associations with structural and frequency properties. We have not performed matched-parameter and matched-compute comparisons against CondConv, Dynamic Convolution, ODConv, or spatial MoE, and do not claim superiority over them.

4. Architectural interpretation and inconsistencies

The experts' different operators impose distinct inductive biases by design. Our claim is not that routing creates these properties or discovers fixed semantic roles, but that it modulates their contributions spatially. Since Ck=wkAkC_k=w_kA_k, correlations involving routed contributions can also be affected mechanically by the weights. We therefore treat these as descriptive associations, not evidence that the router independently selects experts based on image structure. Appendix C.7 also shows that individual expert associations vary across seeds; we do not claim reproducible expert-specific roles.

We have corrected the inconsistencies identified by the reviewer: reconstruction metrics in Tables 3 and 5, decoder expert definitions in Table 1 and Algorithm 2, and logit normalization in Algorithm 1 and Eq. 20. We have also corrected the encoder-block description, Figure 13 caption, complexity-table reference, and template placeholders.

5. Downstream relevance

We agree that identity autoencoding cannot establish downstream utility. We have not evaluated segmentation or super-resolution and make no claims of improvement on these tasks. Evaluating transfer to downstream volumetric tasks remains future work.

Overall, we have narrowed the framing to distinguish the analysis of spatially modulated operator contributions from claims of substantial reconstruction gains, universally necessary experts, or downstream utility.

Reviewer_32zfOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a spatially adaptive convolutional autoencoder based on a mixture of five heterogeneous 3D convolutional experts, designed to capture different spatial and frequency characteristics. The contributions of these experts are combined using spatially varying routing weights. Specifically, the encoder employs an attention-based router operating on local window representations, while the decoder uses a lightweight convolutional router.

Critical strengths

None

Critical weaknesses

1. Limited novelty and insufficient differentiation from closely related prior work.

The core formulation of the proposed method, which uses spatially varying, input-dependent weights to combine a bank of convolutional operators, is closely related to existing adaptive convolution methods.

In particular, Revisiting Spatial Invariance with Low-Rank Local Connectivity (Elsayed et al., ICML 2020) already formulates spatially varying convolutional filters as weighted combinations of shared basis filters and further introduces input-dependent combining weights (Section 3.2.2). This is mathematically very close to the central formulation of the proposed method. However, this highly relevant work is not discussed or compared against.

Furthermore, Spatial Mixture-of-Experts (Dryden and Hoefler, NeurIPS 2022) has already explored fine-grained spatial routing over expert transformations. Although this work is cited, the paper does not sufficiently establish what fundamental advantage or new capability is introduced by the proposed formulation.

The heterogeneous 3D expert bank and attention-based routing mechanism constitute specific architectural choices, but their distinct contributions over existing approaches have not been convincingly demonstrated. Consequently, the novelty and significance of the proposed method remain unclear.

2. Insufficient experimental evidence supporting the necessity and effectiveness of the proposed architecture.

The ablation results in Appendix Table 5 raise substantial concerns about whether the proposed five-expert architecture and adaptive routing mechanism are necessary. First, adaptive routing improves PSNR by only 0.20 dB over uniform expert aggregation (29.04 vs. 28.84 dB), with an SSIM improvement of only 0.0010.

More importantly, several simpler expert configurations perform comparably or better. The single-expert E2 model achieves 29.14 dB, already slightly exceeding the complete five-expert model. Removing E0 or E3 further improves PSNR to 29.68 dB and 29.64 dB, respectively. These results suggest that the proposed heterogeneous expert composition may introduce unnecessary complexity, and the benefit of combining all five experts is not established.

3. Fundamental confusion about the research problem, lack of scientific rigor, and unacceptably poor presentation.

My most serious concern is that the paper lacks a coherent scientific motivation and demonstrates a fundamental confusion about the problem it claims to address.

The abstract explicitly states that the framework is evaluated on "three-dimensional magnetic resonance imaging (MRI) reconstruction." However, what the authors actually perform is a straightforward autoencoding task, where a fully reconstructed 3D MRI volume is provided as input and the network is trained to reproduce that same volume. This is fundamentally different from MRI reconstruction in its conventional sense, which involves recovering images from k-space measurements under an acquisition model. The experiments involve no k-space measurements, undersampling, forward operator, or data consistency. Calling this task "3D MRI reconstruction" is highly misleading. More importantly, the authors provide no convincing explanation of what meaningful scientific or clinical problem this autoencoding task is supposed to solve. Why is reconstructing an already reconstructed MRI volume an important research problem in this context? What does this evaluation establish about the proposed method's relevance to medical imaging? This fundamental mismatch raises serious doubts about the authors' understanding of the application domain and substantially undermines the motivation and significance of the entire work.

Beyond this conceptual problem, the manuscript itself is remarkably incoherent and poorly presented. The paper reads like a loosely assembled collection of existing architectural components, heuristic design choices, and post hoc analyses, without a clear scientific hypothesis or a convincing logical progression from motivation to method and validation. It is frequently unclear why particular components are introduced, what specific problem they are intended to solve, or how the presented experiments support the central claims. The figures are also of extremely poor quality, with small or difficult-to-read elements and insufficient explanations, making several results unnecessarily difficult to interpret. Combined with the inconsistencies between the mathematical descriptions, architecture specifications, and pseudocode, the overall presentation falls substantially below the standard expected of an ICLR submission.

The paper gives a strong impression of an insufficiently scrutinized LLM-driven "auto-research" exercise rather than a carefully conceived and rigorously validated scientific investigation. Notably, the authors' AI Use Statement explicitly acknowledges the use of generative AI across nearly every stage of the research process, including hypothesis formulation, mathematical development, experimental design, implementation, interpretation, and manuscript preparation. While AI assistance is not inherently problematic, and the disclosure alone does not establish autonomous research generation, the pervasive conceptual confusion, weak scientific justification, and numerous presentation and implementation inconsistencies raise serious concerns about the adequacy of independent scientific scrutiny.

Overall, I find it difficult to identify a well-defined, meaningful research question that this paper convincingly addresses. In its current form, the work appears insufficiently motivated, inadequately validated, and far below the level of scientific rigor and clarity expected at ICLR.

Questions for authors

1. Clarification of novelty relative to prior work.

Could the authors clarify the essential methodological differences between the proposed formulation and input-dependent Low-Rank Local Connectivity (Elsayed et al., ICML 2020), particularly its formulation of spatially varying effective kernels through learned combinations of shared basis filters? What advantages are provided specifically by the heterogeneous expert design and attention-based routing? A direct comparison with these closely related methods, or appropriately adapted implementations, would materially strengthen the novelty claim.

2. Necessity of the full expert bank and attention-based routing.

How do the authors explain the observation that a single E2 expert achieves comparable or slightly better reconstruction performance than the complete five-expert model, while removing E0 or E3 further improves performance? Additionally, is the modest improvement over uniform routing consistent across multiple random seeds? An ablation comparing attention-based routing with a simpler convolutional router under matched settings would help establish whether the additional routing complexity is justified.

3. Evaluation setting and baseline fairness.

The experiments reconstruct the same fully observed MRI volume provided as input. Could the authors clarify the bottleneck dimensions and compression constraints of the autoencoder, and explain how this evaluation supports the intended MRI reconstruction application? Furthermore, are the baseline architectures and training configurations sufficiently matched to isolate the benefits of the proposed convolutional formulation?

4. Implementation and reporting inconsistencies.

Several details require clarification. Algorithm 2 describes uniform aggregation of decoder experts, whereas Section 3.2.2 specifies adaptive routing in the first decoder block. There are also differences in the expert definitions between Table 1 and Algorithm 2, as well as in the encoder routing-logit normalization described in Appendix A.2 and Algorithm 1. Additionally, Tables 3 and 5 report different MSE and SSIM values for the proposed full model despite reporting identical PSNR. Could the authors clarify which implementations and experimental settings produced the reported results?

1 comment
AuthorsOpenReview ↗

Clarification of Novelty, Experimental Scope, and Architectural Design

We thank the reviewer for the detailed feedback. We address the principal concerns below.

1. Novelty relative to prior work

We acknowledge that spatially varying combinations of convolutional operators have been explored, particularly by Elsayed et al. (ICML 2020), whose LRLC formulation combines shared basis filters using input-dependent spatial weights. We do not claim to introduce this general formulation. Instead, we investigate the spatial composition of structurally heterogeneous 3D convolutional operators and characterise their routed contributions in volumetric data.

Although the experts possess different inductive biases by construction, we compare their structural associations before and after routing. Expert 0 changes from negative to positive correlations across all four structural measures, while Expert 4 becomes more negatively associated with them. This suggests that routing modifies how the operators contribute in relation to image structure. We therefore describe this as routing-dependent functional differentiation, not the discovery of fixed semantic roles.

The encoder uses attention over local windows, while the decoder employs a lightweight convolutional router. We acknowledge that we do not provide a matched ablation comparing these routing mechanisms and therefore do not claim that attention is superior to alternative routers.

2. Necessity of the expert bank and adaptive routing

We agree that E2 alone achieves comparable reconstruction performance to the full model, with marginally higher PSNR and nearly identical MSE and SSIM. We therefore do not claim that the five-expert configuration is optimal for reconstruction. Our objective is to investigate spatial composition, which an individual expert cannot capture.

Removing E2 from the adaptive bank reduces PSNR from 29.04 to 24.87 dB, indicating its importance within the composition. However, we acknowledge that some reduced configurations perform better than the full bank.

Adaptive routing improves over uniform aggregation by only 0.20 dB PSNR and 0.0010 SSIM. We agree this is a modest reconstruction improvement. Appendix C.7 examines routing variability across three seeds, but does not establish the consistency of this performance gain across seeds. We will moderate our claims accordingly.

3. Evaluation setting and baseline fairness

We agree that our experiment is volumetric MRI autoencoding: fully reconstructed IXI-T1 volumes serve as both inputs and targets. It does not involve k-space measurements, undersampling, an acquisition model, or data consistency. Calling this conventional MRI reconstruction is inaccurate, and we will correct the terminology throughout.

We use MRI autoencoding as a testbed to study how heterogeneous 3D operators respond to image structure and frequency characteristics, and how routing modulates their contributions. This evaluation does not establish clinical utility; downstream segmentation and undersampled k-space reconstruction remain future directions.

We also acknowledge that latent dimensions and compression ratios were not reported, preventing a capacity-matched baseline comparison. Overall compression depends on encoder downsampling and latent channel dimensionality, not just decoder upsampling. Furthermore, equal training duration of 20 epochs does not guarantee comparable convergence. We will clarify these limitations and moderate our reconstruction comparisons.

4. Implementation and reporting inconsistencies

We thank the reviewer for identifying inconsistencies in the mathematical descriptions and pseudocode. We have corrected the expert definitions in Table 1 and Algorithm 2, aligned the encoder routing-logit normalization in Appendix A.2 with Algorithm 1, and clarified that only the first decoder block uses adaptive routing while the remaining two use uniform aggregation. We have also corrected the inconsistent MSE and SSIM values in Tables 3 and 5, along with related figure captions and architectural descriptions.

Finally, we have clarified the central research question: whether spatial routing modulates the contributions of heterogeneous 3D convolutional operators and how these contributions relate to structural and frequency properties. We have narrowed our claims to reflect what the experiments establish, without asserting fixed expert roles or universal superiority over simpler alternatives.

We thank the reviewer for helping us improve the manuscript's clarity, technical consistency, and scientific scope.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Clarifying the Contribution and Evidence for Spatially Adaptive Expert Routing

We thank the reviewer for the constructive assessment and for recognising the concrete design for volumetric data and the analyses beyond aggregate reconstruction scores. We agree that the contribution and the evidence for the full architecture require more precise qualification.

1. Contribution beyond existing spatial MoE and dynamic convolution

We agree that attention-weighted expert mixtures and spatially varying convolution are established ideas. As shown in Appendix A, our formulation can be expressed as a spatially varying effective kernel; we do not claim this general operation as fundamentally new. Our contribution is narrower: an empirical investigation of how a fixed bank of heterogeneous 3D convolutional operators is spatially combined, and how its routed contributions relate to structural and frequency properties of volumetric data. We have not performed matched-resource comparisons against closely related spatial MoE or dynamic convolution methods, and therefore do not claim superiority over them.

2. Does spatial routing consistently help?

We acknowledge that the measured gain over uniform aggregation is modest: 0.20 dB PSNR and 0.0010 SSIM. E2-only also achieves competitive reconstruction (29.14 dB PSNR, 0.9593 SSIM), while removing E0 or E3 improves the full model's reconstruction scores. Thus, our results do not establish that the full expert bank is optimal or that spatial routing consistently improves reconstruction.

The existing seed analysis in Appendix C.7 examines variability in expert-level routing patterns and structural associations; it does not establish the seed-wise consistency or statistical significance of the adaptive-versus-uniform reconstruction gain. We have not conducted the additional comparisons against learned spatially constant weights or a convolutional router using the same expert bank. These would help isolate the contribution of spatial adaptation and encoder attention, and we acknowledge their absence.

3. Learned behaviour versus architectural inductive biases

We agree that the experts' directional and frequency responses are partly prescribed by their construction. The observed differences cannot, by themselves, establish emergent specialization. Our analysis instead examines how routing modulates the contributions of these predefined operators. Since the routed contribution is Ck=wkAkC_k=w_kA_k, its correlation with structural measures can also be affected mechanically by the routing weights; we interpret these correlations descriptively, not as proof that the router independently discovers image structure.

We have not evaluated untrained experts or a homogeneous expert bank. Consequently, we cannot quantify how much of the observed behaviour is attributable to learning rather than architectural design. We have narrowed the specialization claims accordingly. Appendix C.7 also indicates that individual expert associations vary across seeds, so we do not claim fixed, reproducible semantic roles for individual experts.

4. Bottleneck capacity and training conditions

We agree that comparisons with MONAI and the ViT require explicit latent dimensions and compression ratios to establish capacity comparability. The decoder's 2× in-plane upsampling alone does not determine the overall encoder-to-latent compression ratio. Further, equal training epochs do not guarantee convergence, particularly when inference and computational costs differ. We therefore interpret the reported reconstruction scores under the stated training setup, not as evidence of superiority under matched bottleneck or convergence conditions.

We also acknowledge the runtime overhead relative to MONAI (347.79 vs. 26.35 ms inference; 2.276 vs. 0.313 GB peak memory). The model is parameter-efficient, but not computationally efficient.

5. Reconstruction metric discrepancy

We agree that the differing MSE and SSIM values for the full model in Tables 3 and 5 were a reporting inconsistency. We have reconciled the values in the revised manuscript. This correction is important given the small adaptive-versus-uniform reconstruction difference.

Overall, we have narrowed the claims to the analysis of spatially modulated contributions from heterogeneous 3D operators. We do not claim that the full expert bank is optimal, that routing provides a substantial reconstruction gain, or that the observed expert behaviour is independent of architectural inductive biases.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.