{"id":"2f63813d-8ef9-4258-984c-5449c665ef21","arxiv_id":"2501.09753","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A radially symmetric convolution kernel, SRE-Conv, improves rotated-image classification accuracy on all 16 MedMNISTv2 tasks while using fewer parameters than standard CNNs.","lead":"A new convolution kernel, SRE-Conv, ties weights into circular bands so features do not change when the image is rotated. Tested on 16 medical image datasets, it beat rotation-equivariant baselines on rotated inputs while using far fewer parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Rotated-accuracy gains are confounded with kernel size: SRE uses [9,9,5,5] kernels while baselines use ~3x3, so the reported improvements do not isolate rotational symmetry as the cause.","rationale":"The reader's formal weakest assumption is discretization error in exact rotation equivariance; that concern is real but secondary because the empirical rotated-accuracy claim is about relative improvement, not exact invariance. The more load-bearing issue is experimental confound: SRE-CNN is compared to baselines with smaller kernels, so the improved rotated accuracy and parameter counts do not isolate the effect of rotational symmetry. The abstract and Section 3.2 attribute the gains to the SRE kernel's equivariance, but Section 3.1's choice of [9,9,5,5] kernel sizes changes receptive field and compute simultaneously. Table 1 also shows large per-dataset drops (e.g., OCT 76.8 to 52.2 on rotation), so exact invariance is not reflected in the results; the p=0.11 null result is a low-power test, not evidence of invariance. A matched-kernel ablation would settle whether the mechanism matters or whether a larger unconstrained kernel with pooling downsampling would give the same robustness. The paper is otherwise honest about orientation-prior failures and includes code, so the appropriate outcome remains conditional acceptance pending this ablation.","tokens_in":8291,"tokens_out":10183,"duration_ms":119432,"concrete_test":"On 2-3 representative tasks (e.g., PathMNIST, OCTMNIST, OrganAMNIST), train: (A) SRE-CNN with [9,9,5,5] kernels as in the paper; (B) a standard ResNet-18 modified to use the same [9,9,5,5] kernel sizes and the same pooling+1x1 downsampling, with width adjusted to match SRE's 3.9M parameter count; (C) SRE-CNN with [3,3,3,3] kernels. Compare rotated-test accuracy and FLOPs. If (B) matches or beats (A), or if (C) loses SRE's advantage over R18, the paper's claimed equivariance mechanism is not the cause of the reported gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 states that SRE-CNN uses kernel sizes [9,9,5,5] for every stage, whereas the ResNet-18/R3D-18 baselines use mostly 3x3 kernels (7x7 in the first conv). The parameter reduction to 3.9M vs 11.2M is a direct consequence of sharing one weight per radial band (b = floor(k/2)+2), but the actual convolution is a dense 9x9 operation, i.e., about 9x the multiply-adds of a 3x3 conv at the same resolution. Section 2.2's claim of equal computational complexity is true only relative to a standard kernel of the same size, not relative to the baselines actually compared. Thus the central 'efficiency' and 'rotation robustness' claims are not isolated: a standard large-kernel CNN with comparable receptive field and FLOPs might exhibit similar rotated-test accuracy, and a 3x3 SRE-Conv might lose the reported advantage. Without a kernel-size-matched ablation, the experiments do not establish that the radial band-sharing mechanism, rather than increased receptive field, is responsible for the accuracy gains across the 16 tasks.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SRE-Conv, a convolution kernel parameterized by radial distance bands so that all weights within a symmetric band are shared. The authors construct fully convolutional networks using SRE-Conv and evaluate them on 16 MedMNISTv2 tasks (10 2D, 6 3D), reporting improved classification accuracy on rotated test sets for all 16 tasks compared with ResNet-based and equivariant baselines, while using fewer parameters and less GPU memory. The paper also shows qualitative feature-map alignment under rotation and makes the code publicly available.","tokens_in":8518,"tokens_out":5104,"duration_ms":53751,"significance":"If the empirical claims are established, the method would offer a simple and lightweight way to build rotation robustness into CNN backbones for biomedical imaging, and the public code is a useful contribution. The parameter reduction from weight sharing (O(C·k) instead of O(C·k²)) is real, and the authors are careful to avoid geometric data augmentation, making the comparisons directly relevant to equivariant learning. However, the central evidence is weakened by a kernel-size confound between the SRE models (9×9 and 5×5 kernels) and the baselines (3×3 kernels), and by the paper's overstated claim of exact equivariance for arbitrary rotation angles on a discrete grid. The lack of any variance information further tempers confidence in the reported gains.","major_comments":[{"comment":"The comparison is confounded by kernel size. SRE-R18 and SRE-R3D-18 use kernel sizes [9,9,5,5] for all stages, while the R18/R3D-18 baselines use 3×3 kernels (7×7 in the first conv). The rotated-accuracy advantage could therefore arise from the larger receptive field rather than from the radial band-sharing mechanism. Please add ablations with kernel-size-matched models: (i) SRE-Conv with 3×3 kernels (or the baseline's kernel sizes) and (ii) standard convolution with [9,9,5,5] kernels. Also report FLOPs and inference time for all models; as it stands, the statement in §2.2 that SRE-Conv has 'the same number of floating operations (FLOPs) as traditional convolution' is true only relative to a standard kernel of the same size, not relative to the 3×3 baselines actually compared.","section":"§3.1 and §3.2, Tables 1–2"},{"comment":"The claim of exact rotational equivariance is not valid for the non-lattice rotations used in the evaluation. On a discrete pixel grid, a kernel constructed by binning Euclidean distances is invariant to rotations that map grid points to grid points (multiples of 90°), but the rotated test sets use 10° increments (2D) and 30° increments about each axis (3D), which require interpolation. The band-binning makes the kernel a radial step function, so the equivariance is approximate for these angles. Please state this qualification explicitly, provide a quantitative characterization of the approximation (e.g., how rotated accuracy varies with rotation angle), and avoid the unqualified statement that the arrangement 'ensures local equivariance'.","section":"§2.1"},{"comment":"No per-run variance is reported. All accuracy numbers appear to come from a single training run per model and dataset, and the paired t-tests are computed across the 10 (or 6) datasets rather than across repeated runs. Thus the reported p-values only show that the mean differences across datasets are unlikely to be zero under a paired test; they do not account for training stochasticity. Please report results from multiple seeds with standard deviations or confidence intervals, or clearly justify the single-run protocol.","section":"§3.2 and Tables 1–2"},{"comment":"The efficiency claim is overstated. While the parameter count is indeed reduced to O(C·k), the actual convolution is a dense operation on a k×k kernel, so a 9×9 SRE-Conv performs roughly 9× the multiply-adds of a 3×3 standard convolution at the same resolution. The paper reports lower GPU memory for SRE-R18, but this does not translate to lower compute. The abstract's phrase 'increasing efficiency' should be qualified as parameter and memory efficiency, and the trade-off with computational cost should be disclosed in the experimental comparison.","section":"§2.2 and Abstract"}],"minor_comments":[{"comment":"The phrase 'increasing efficiency with fewer parameters and reduced memory footprint' leaves 'efficiency' ambiguous; consider using 'parameter efficiency' and explicitly stating that computational cost is comparable to a same-size standard convolution.","section":"Abstract"},{"comment":"The notation involving C is confusing: 'we denote C for Cin and Cout for simplicity' is followed by K ∈ R[C,k,k], which does not represent the usual [Cin,Cout,k,k] shape or a shared kernel unless Cin=Cout. Please clarify the tensor notation.","section":"§2.1"},{"comment":"The construction of the binary index matrix is described as 'assigning 1 to each row-column index that has equal distance values in D'; this is ambiguous because D contains many distinct distance values. Please specify the thresholding or band-assignment rule precisely.","section":"§2.1"},{"comment":"The text states that OrganC and OrganS are skipped because they are repeated with OrganA and Organ3D, but OrganS is not a duplicate of OrganA in MedMNISTv2; please clarify the relationship between the skipped and included datasets.","section":"§3.1"},{"comment":"The description of unrotating feature maps should state the interpolation method used for the 60° increments, since these are not lattice rotations and the qualitative alignment depends on interpolation.","section":"Figure 2"},{"comment":"The table layout for memory usage is inconsistent: in Table 1 the 'Mem.' column appears only in the second block, and it is unclear whether 'Mem.' applies to each model row or only to the listed row. Please format the tables so that every model has a memory value.","section":"Tables 1–2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a compact workshop-style contribution and the core idea is simple and plausible; the public code and the breadth of evaluation are strengths. The kernel-size confound is the main technical obstacle: without a matched-kernel ablation, the headline claim that radial symmetry itself drives the rotated-accuracy gain is not established. A major revision addressing the ablation, the exact-equivariance qualification, and variance reporting would make the paper acceptable. The novelty relative to prior weight-symmetric kernels (e.g., Dudar & Semenov, Fuhl & Kasneci) is incremental, but the systematic MedMNIST evaluation and the code release add value."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clean engineering extension of an existing idea, and the experiments are broader than most in the subfield. The symmetric band-shared kernel is in [11] and [12]; the authors say so, and their real contributions are the arbitrary-size parameterization, the 3D extension, and a 16-task MedMNISTv2 benchmark with public code. Those are useful additions, not a new concept.\n\nWhat they do well: the implementation is careful, the feature-map visualization is a direct demonstration of rotation consistency, and the parameter reduction is real and large (3.9M vs 11.2M in 2D; 2.6M vs 33.2M in 3D). The 3D variant is nontrivial and likely to be practically useful.\n\nThe soft spots are not fatal, but they are genuine. First, the comparison is confounded by kernel size. SRE uses [9,9,5,5] kernels; the ResNet baselines use mostly 3x3. The paper's 'same computational complexity' claim holds only against a standard kernel of the same size, not against the actual baselines. The dense 9x9 conv is roughly 9x the multiply-adds of a 3x3 at that layer, and no baseline with a comparable receptive field is tested. Without a kernel-size-matched control or a 3x3 SRE variant, the rotated-accuracy gains cannot be cleanly attributed to radial symmetry rather than receptive field. Second, the exact-equivalence wording in Sec. 2.1 overstates the case for arbitrary angles. On a discrete grid, 10-degree rotations (or 30-degree in 3D) require interpolation, and the band-binning quantizes distances, so the equivariance is approximate, not exact. That distinction matters to anyone extending the method. Third, per-run variance is absent; the paired t-tests across 10 datasets with one accuracy per dataset are a weak substitute for seed variance.\n\nThese are all fixable with better-controlled experiments and a more precise claim. As is, the paper is a solid practical report, but the abstract's implication that the symmetry mechanism itself causes the gains is not yet established.\n\nI would send this to review. The extension is real, the benchmark is broad, and the flaws are addressable. I would not cite it as evidence for rotation equivariance without the matched-kernel ablation, but I would cite the 3D parameterization. For a reading group, it is a good case study in how kernel size and symmetry interact in experimental design.","headline":"Solid engineering extension of a known kernel idea, with a broad benchmark, but the kernel-size confound and overclaimed exactness leave the mechanism unproven.","tokens_in":9039,"tokens_out":3692,"would_cite":false,"duration_ms":38555,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Radial-band convolution wins all 16 rotated medical-image benchmarks.","keywords":["rotation equivariance","symmetric convolution kernel","radial bands","biomedical image classification","MedMNISTv2","parameter efficiency","3D convolution","equivariant neural networks"],"falsifier":"Take a trained SRE-Conv layer, rotate a fixed input image by an angle that is not a multiple of 90 degrees, run both versions through the layer, rotate the second feature map back, and measure the maximum absolute difference: exact equivariance predicts a difference at machine precision, while interpolation and band discretization would produce a nonzero gap. A second check is to compare the kernel $K$ itself under an arbitrary rotation: if the rotated kernel differs from the original at any non-90-degree angle, the claimed exact equivariance in Section 2.1 is falsified at that angle.","tokens_in":8118,"feed_emoji":"🩻","tokens_out":5673,"duration_ms":54948,"temperature":0.7,"pith_summary":"The paper proposes that a convolutional kernel whose trainable weights are organized by radial distance from the kernel center can make a CNN rotation- and reflection-equivariant at no extra computational cost. It reports that replacing standard kernels in a ResNet-style fully convolutional network with these symmetric SRE-Conv kernels yields the best classification accuracy on rotated test sets for all 16 MedMNISTv2 datasets, in both 2D and 3D, while using roughly a third of the parameters in 2D and under a tenth in 3D. The motivation is that biomedical images often have no canonical orientation, so a model that is stable under rotation could reduce reliance on data augmentation and improve generalization. If true, the result offers a parameter-efficient drop-in convolution layer for biomedical image classifiers.","feed_headline":"Radial-band convolution wins all 16 rotated medical benchmarks","feed_subtitle":"A symmetric distance-band kernel makes CNNs rotation-invariant and cuts parameter count dramatically.","key_machinery":"The central object is the SRE-Conv kernel, defined as $K = M_I \\Theta$, where $\\Theta \\in \\mathbb{R}^{[C,b]}$ contains the trainable weights and $M_I \\in \\mathbb{R}^{[b,k,k]}$ is a fixed binary index matrix that assigns each kernel position to one of $b = \\lfloor k/2 \\rfloor + 2$ radial bands by Euclidean distance from the center, with the four corners zeroed to create a circular support. This construction reduces trainable parameters from $O(C k^2)$ to $O(C k)$, allows arbitrary kernel sizes, and preserves the same floating-point operations as standard convolution because the full kernel is precomputed at inference. The equivariance claim rests on the radial symmetry of the bands combined with convolution's translation equivariance.","core_discovery":"The paper's central claim is that a convolution kernel parameterized by $b$ non-overlapping circular bands, defined as equal-width ranges of Euclidean distance from the kernel center, produces feature maps that stay consistent when the input is rotated or reflected. Because ordinary convolution is translation-equivariant, sliding this radially symmetric kernel over the image extends the local symmetry to global rotation and reflection equivariance, and global adaptive pooling makes the final classification invariant. Across all 16 MedMNISTv2 tasks, the resulting SRE-CNN improves rotated-test accuracy over both conventional ResNet baselines and five rotation-equivariant competitors in 2D, and over two 3D baselines, with no statistically significant drop on original test sets overall. The paper acknowledges that on datasets with a strong anatomical orientation prior, such as OrganA in 2D and Organ in 3D, the symmetric constraint can lower original-test accuracy.","pith_inferences":["Because the equivariance is exact only for rotations that map grid points to grid points, such as multiples of 90 degrees, the strong gains on 10-degree and 30-degree rotated inputs are partly a discrete approximation; a direct pixel-level invariance test on non-grid rotations would clarify how much comes from exact symmetry versus learned robustness.","The band-binning parameterization is effectively a radial basis with hard boundaries, so replacing hard bands with smooth radial basis functions could preserve the parameter savings while reducing interpolation artifacts at arbitrary angles.","A hybrid that combines SRE-Conv with a small orientation-estimating branch could recover the lost performance on orientation-prior datasets like OrganA while keeping rotation robustness on orientation-free tasks.","The same kernel construction could be applied beyond classification, such as to segmentation or detection backbones, where local rotation equivariance may help when individual objects rotate within a larger scene."],"forward_implications":["Rotation and reflection data augmentation may become unnecessary for biomedical classifiers built on SRE-Conv, since rotated and reflected test performance matches original performance.","Large kernels become affordable: because parameters scale linearly with kernel size rather than quadratically in 2D or cubically in 3D, SRE-Conv can use 9x9 or larger kernels to enlarge receptive fields without blowing up model size.","The same construction extends to 3D, where parameter savings are even larger, making it practical for volumetric medical images with arbitrary orientation.","SRE-Conv can replace standard convolution layers in existing CNN backbones, so the accuracy and efficiency trade-off is available without architectural redesign.","On datasets with a strong anatomical orientation prior, the symmetric constraint can hurt original-test accuracy, so pure SRE-CNNs are best suited to orientation-agnostic tasks or need additional orientation-aware components."],"supporting_citations":[{"why":"Provides the MedMNISTv2 benchmark, the 16 classification tasks and evaluation protocol used for all experiments.","marker":"[15]"},{"why":"Supplies the ResNet-18 backbone that SRE-R18 modifies and the conventional CNN baseline.","marker":"[16]"},{"why":"Group-equivariant CNNs serve as a rotation-equivariant baseline to compare against.","marker":"[7]"},{"why":"Oriented response networks provide a baseline that learns orientation information during training.","marker":"[8]"},{"why":"Harmonic networks provide a circular-harmonics rotation-equivariant baseline.","marker":"[5]"},{"why":"E(2)-steerable CNNs provide a steerable-filter baseline for 2D comparison.","marker":"[6]"},{"why":"RIC-CNN provides a rotation-invariant coordinate-system baseline.","marker":"[9]"},{"why":"Marr-Hildreth edge detection is cited as the inspiration for using a centrally symmetric kernel.","marker":"[13]"},{"why":"Prior symmetric-kernel CNNs establish the weight-symmetric approach this work extends.","marker":"[11]"},{"why":"Provides the 3D ResNet baselines R3D-18 and R2plus1D-18 used for 3D comparison.","marker":"[17]"}],"fun_headline_variants":["SRE-Conv kernel wins all 16 rotated medical tasks","Rotation-equivariant CNN beats baselines with fewer params","Radial-band convolution boosts rotated image accuracy","Symmetric kernel slims models, perfects rotation invariance","SRE-Conv: 16/16 rotated datasets, smaller footprint"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that grouping kernel positions into radial distance bands makes the convolution exactly rotation equivariant; on a discrete pixel grid this is exact only for rotations that map grid points onto grid points, and the test rotations (10 degrees in 2D, 30 degrees about each axis in 3D) require interpolation, so the claimed mechanism is approximate at those angles.","fun_headline_variants_meta":{"raw":{"variants":["SRE-Conv kernel wins all 16 rotated medical tasks","Rotation-equivariant CNN beats baselines with fewer params","Radial-band convolution boosts rotated image accuracy","Symmetric kernel slims models, perfects rotation invariance","SRE-Conv: 16/16 rotated datasets, smaller footprint"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1269,"prompt_tokens":935,"completion_tokens":334,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":253}},"tokens_in":551,"tokens_out":334,"duration_ms":4404,"temperature":1.0,"reasoning_tokens":253,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:41:06.090430+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained SRE-Conv layer, rotate a fixed input image by an angle that is not a multiple of 90 degrees, run both versions through the layer, rotate the second feature map back, and measure the maximum absolute difference: exact equivariance predicts a difference at machine precision, while interpolation and band discretization would produce a nonzero gap. A second check is to compare the kernel $K$ itself under an arbitrary rotation: if the rotated kernel differs from the original at any non-90-degree angle, the claimed exact equivariance in Section 2.1 is falsified at that angle.","supporting_citations":[{"cited_title":"Use of sym- metric kernels for convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the MedMNISTv2 benchmark, the 16 classification tasks and evaluation protocol used for all experiments."},{"cited_title":"Rotated ring, radial and depth wise separable radial convolutions,","cited_arxiv_id":null,"evidence_quote":"Supplies the ResNet-18 backbone that SRE-R18 modifies and the conventional CNN baseline."},{"cited_title":"Gradient-aligned convolution neural network,","cited_arxiv_id":null,"evidence_quote":"Oriented response networks provide a baseline that learns orientation information during training."},{"cited_title":"Revisiting data augmenta- tion for rotational invariance in convolutional neural net- works,","cited_arxiv_id":null,"evidence_quote":"E(2)-steerable CNNs provide a steerable-filter baseline for 2D comparison."},{"cited_title":"Harmonic net- works: Deep translation and rotation equivariance,","cited_arxiv_id":null,"evidence_quote":"RIC-CNN provides a rotation-invariant coordinate-system baseline."},{"cited_title":"RIC-CNN: Rotation- Invariant coordinate convolutional neural network,","cited_arxiv_id":null,"evidence_quote":"Marr-Hildreth edge detection is cited as the inspiration for using a centrally symmetric kernel."},{"cited_title":"Theory of edge detection,","cited_arxiv_id":null,"evidence_quote":"Provides the 3D ResNet baselines R3D-18 and R2plus1D-18 used for 3D comparison."}],"review_version":1}