{"id":"29d7bcca-8377-4027-84d1-c24eba4d7f58","arxiv_id":"1908.01300","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"SOVNET, a capsule network with group-equivariant convolution predictions and degree-centrality routing, is equivariant to its chosen transformation group, and its capsule-decomposition graph is isomorphic under such transformations.","lead":"This paper introduces SOVNET, a capsule network variant that replaces pairwise part-to-whole prediction networks with group-equivariant convolutions and routes capsules by degree centrality. The authors prove that these layers commute with transformations in the chosen group and report accuracy gains under affine perturbations over several capsule baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Model-level equivariance is asserted for the full SOVNET, but only the routing layer in Algorithm 2 is proved equivariant; the primary capsule construction, modified residual blocks, and output pooling are described without equivariance proofs, so the graph-isomorphism theorem does not yet apply…","rationale":"The reader's weakest assumption identifies exactly the same load-bearing gap: the equivariance proof covers the routing layer but not the surrounding components of the full SOVNET architecture. I agree with this assessment. The paper's headline contribution is that 'this particular instance of our general model is equivariant,' and the only rigorous proof is for Algorithm 2 in isolation. The primary capsule layer, the modified residual blocks, and the final pooling are described in prose and without equivariance proofs, so the model-level claim is under-supported. This does not mean the claim is false; in fact, standard group-convolution layers, pointwise SELU, and residual addition can all be made equivariant, and group max-pooling is a standard invariant operation for classification. But the paper does not show this for its specific architectural choices, and no code is released to check it empirically. The condition for Theorem 2.2 is that the routing procedure is equivariant; the paper only proves that for one layer, not for the full network. Therefore the correct assessment is conditional: the central claim is plausible and partially proven, but the full-network equivariance needs either a formal proof for every component or an empirical equivariance check. This does not change the reader's verdict, which already conditioned acceptance on these missing pieces. I find no reason to move to accept or reject; the paper's core routing-layer proof appears correct in outline, and the main weakness is the unverified extension to the complete architecture.","tokens_in":16819,"tokens_out":4833,"duration_ms":49995,"concrete_test":"Numerically verify full-network equivariance on the described architecture: take the network with random or trained weights, choose a group element g (e.g., a 90-degree rotation in p4), and compare the network output on L_g x with L_g applied to the output on x, using the layer immediately before the final max-pooling. Repeat for the primary capsule layer and for each modified residual block in isolation. Any mismatch beyond numerical tolerance would falsify the model-level equivariance claim. Alternatively, formally specify each component as an equivariant map, prove its equivariance, and compose these proofs to extend Theorem A.1 from Algorithm 2 to the full model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed SOVNET instance is equivariant and therefore preserves the capsule-decomposition graph under transformations (Theorems 2.1 and 2.2). The proof in Appendix A.1 covers only the operations inside Algorithm 2: group-correlation predictions, degree-centrality scoring, weighted summation, and squash. The actual model described in Section 3 includes additional components: a primary capsule layer built from a modified residual block with SELU and group-equivariant convolutions, hidden SOVNET blocks that use a group-equivariant modified residual block as the prediction mechanism, and an output capsule layer followed by max-pooling over group positions. For the model-level equivariance statement to hold, every one of these components must be equivariant with respect to the same group action L_g. The paper does not formally define these blocks or prove their equivariance. In particular, if the residual blocks contain any non-equivariant operation, such as batch normalization or a shortcut that alters the group action, the equivariance chain is broken. The final max-pooling over group positions is not equivariant (it is invariant), so if the output capsules are taken after this pooling, the vertex and edge construction of Theorem 2.2 no longer applies. Thus the paper establishes equivariance for a routing layer, but the central claim requires equivariance of the full network, and that is exactly the part left unproved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SOVNET, a capsule-network framework in which each capsule type is associated with a trainable equivariant prediction function, instantiated with group-equivariant convolutions, and in which routing is performed by a degree-centrality-based aggregation. The paper proves that the routing layer (Algorithm 2) is equivariant (Theorem 2.1) and that any equivariant routing procedure in the family of Algorithm 1 preserves the capsule-decomposition graph up to isomorphism (Theorem 2.2). Experiments on affine-transformed MNIST, FashionMNIST, and CIFAR-10, plus KMNIST and SVHN, compare SOVNET against several capsule baselines and a few convolutional baselines.","tokens_in":17103,"tokens_out":8170,"duration_ms":84262,"significance":"If the full-model equivariance claim were established, this would be a useful contribution: the architecture is more parameter-efficient than pairwise prediction capsule networks, the stepwise equivariance proof for Algorithm 2 is structurally sound and relies only on standard GCNN results, and the graph-based formalization of compositionality is a worthwhile framing. The experimental study is broad and shows consistent gains in transformation robustness over the tested capsule baselines. The significance is tempered, however, by the gap between what is proved (the routing layer) and what is claimed in the abstract (that the particular model instance is equivariant), and by the near-tautological character of Theorem 2.2.","major_comments":[{"comment":"The central claim that the full SOVNET model is equivariant is not established. The proof of Theorem 2.1 in Appendix A.1 covers only the operations inside Algorithm 2: group-correlation predictions, degree-centrality scoring, weighted summation, and squash. Section 3 describes the primary capsule layer, the hidden SOVNET blocks, and the output capsule layer only qualitatively ('We only present an outline'), and no equivariance proof is given for these components. In particular, the output layer takes a maximum over group positions of the two-norm, which is an invariant operation; subgroup pooling in Cohen and Welling is invariant, not equivariant. If the output capsules are formed after this pooling, they do not satisfy the transformation law [L_g f](x)=f(g^{-1}x) that Theorem 2.2 requires. The authors should either define every component and prove its equivariance, or explicitly restrict the equivariance and graph-preservation claims to the internal layers and treat the final pooling as an invariant readout. This is load-bearing because the abstract and conclusion claim that 'this particular instance' is equivariant and hence preserves the compositional representation of an input.","section":"Section 3 / Theorem 2.1 scope"},{"comment":"Theorem 2.2 is close to a tautology. After assuming that the routing procedure is equivariant, the proof's edge-preservation step simply restates this assumption in graph language ('due to the assumed equivariance of the model, ... is routed to ... with routing coefficient ...'), so the theorem adds no content beyond the definition of the graph. Furthermore, the graph definition assigns the same routing coefficient c^{l+1}_{ij}(g2) to every edge entering (i,g2) from any shallower position, so the graph does not represent position-specific part-whole relationships; the isomorphism then holds essentially by construction. The authors should either enrich the graph definition to include position-dependent edge weights or temper the claim that equivariance 'guarantees the preservation of detected compositionalities.' The proof also contains a typo: the transformed edge's first vertex is written as \\tilde{f}^{l+1}_i(h^{-1}\\circ g_1) but should be \\tilde{f}^{l}_i(h^{-1}\\circ g_1).","section":"Appendix A.2 / Theorem 2.2"},{"comment":"The proof of equivariance of DegreeScore requires division by the product of norms of prediction vectors. When a prediction vector is zero, which is a case the paper itself anticipates for unrelated capsule types in Section 1 (such capsules 'will project themselves to a vector of low activation'), the cosine similarity in the DegreeScore procedure is undefined. The appendix restricts to 'non-zero' maps, but Algorithm 2 as stated does not specify how to handle this case. The authors should specify a well-defined procedure for zero prediction vectors (for example, adding an epsilon to the denominator or defining the similarity to be zero when either norm vanishes) and state that this procedure remains equivariant for all inputs on which the network is defined.","section":"Algorithm 2 / Appendix A.1"}],"minor_comments":[{"comment":"The summation index in the step defining f^{l+1}_j(g) starts at i=1 rather than i=0, which is inconsistent with the rest of the algorithm and with Algorithm 2.","section":"Algorithm 1"},{"comment":"Several entries in the CIFAR-10 table lack the percent sign (for example, '73.45' in the Capsnet row), which should be made uniform for readability.","section":"Table 4"},{"comment":"The sentence 'This is an equivariant operation, as it corresponds to the subgroup-pooling of (Cohen and Welling, 2016)' is inaccurate: subgroup pooling yields invariance, not equivariance, and this wording obscures the model-level issue described in the major comments.","section":"Section 3, output capsule layer"},{"comment":"The 'modified residual block' and the primary capsule construction are described only in words, with no formal definition or hyperparameter details; since the code is 'to be released pending publication,' the experimental section as written is not fully reproducible.","section":"Section 3, architecture details"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the model-level equivariance claim is valid and should be communicated clearly: the paper proves equivariance for the routing layer only, while the abstract and conclusion claim it for the full architecture. The graph-isomorphism theorem is largely definitional and would be better framed as a corollary of a more substantive graph definition. I would also ask the editor to encourage the authors to provide the promised code or at least full architectural details, since the experiments are otherwise difficult to reproduce."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read it: the equivariance proof for the routing layer (Algorithm 2) is correct and checkable, but the paper's headline claim — that the full SOVNET model is equivariant — is not actually proven; and Theorem 2.2 is more a formal restatement than a deep result. Despite that, the paper gives the capsule-network subfield two concrete things: it replaces the expensive pairwise prediction networks with group-equivariant convolutions, and it introduces a degree-centrality routing rule that has no iteration hyperparameter. The stepwise proof for Algorithm 2 is sound; it relies on standard GCNN equivariance and shows that the degree-score, softmax, and weighted sum all commute with the group action. That part is solid.\n\nThe soft spots start with the scope of the claim. The proof covers only the routing layer. The primary capsule construction, the modified residual blocks, and the output pooling are described qualitatively, and no proof shows they preserve the same group action. The output stage actually takes a max over group positions, which is invariant rather than equivariant. So Theorem 2.2, as written, does not apply to the full architecture described in Section 3. This is a genuine gap, though not a fatal one: the authors could either formalize the whole architecture or narrow the theorem to the routing subnetwork.\n\nTheorem 2.2 is close to a tautology. The capsule-decomposition graph is built from the routing weights, so if routing is equivariant, the graph is equivariant by construction. It's a clean formalization of the intuition, but it doesn't add much beyond the definition.\n\nEmpirically, the paper is weaker. No code, no error bars, no ablations. The accuracy tables show SOVNET beating several capsule baselines on affine-transformed MNIST, FashionMNIST, and CIFAR-10, but there are odd points — e.g., on untransformed CIFAR-10, SOVNET drops from 88.3% to 47.6% under a 30° rotation, which the authors explain by noting that p4 only covers multiples of 90°. That's coherent, but it undercuts the broad \"robustness\" framing. Released code and standard deviations would make the empirical claims much more believable.\n\nWho should read it: people actually building capsule networks or working on equivariant representation learning. For them, the degree-centrality routing and the routing-layer equivariance proof are worth engaging with. General readers can skim. It deserves a serious referee — the theoretical part is checkable and the direction is promising. I'd send it to review and ask for code, error bars, and a clear statement of which parts of the architecture the equivariance theorem actually covers.","headline":"Solid routing-layer equivariance proof, but the full-network claim is underproved and the graph theorem is nearly definitional; worth a serious review for the capsule niche.","tokens_in":17640,"tokens_out":5191,"would_cite":true,"duration_ms":48190,"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":"This paper claims that replacing pairwise capsule predictions with per-type equivariant functions and degree-centrality routing makes a capsule network provably equivariant, preserving its part-whole decomposition graph under…","keywords":["capsule networks","equivariance","group-equivariant convolutions","space-of-variation","degree centrality routing","compositionality","transformation robustness","SOVNET"],"falsifier":"Run a trained SOVNET on an input $x$ and on a transformed input $L_gx$ with $g$ in the group used by the convolutions, extract the routing coefficients $c_{ij}(g)$ from both runs, and check whether each edge weight obeys $c_{ij}(h^{-1}\\circ g)$; a mismatch beyond floating-point tolerance would falsify Theorem 2.2 for the full architecture. A second check is layer-level: compare $([L_gF^l]\\otimes \\Psi_j^{l+1})$ with $L_g([F^l \\otimes \\Psi_j^{l+1}])$ for random input capsules and random equivariant filters.","tokens_in":16594,"feed_emoji":"🔁","tokens_out":9189,"duration_ms":84301,"temperature":0.7,"pith_summary":"Most capsule networks learn a separate prediction network for every pair of capsule types, which is expensive and hard to scale, and they lack formal guarantees that their representations transform predictably. This paper claims both problems can be fixed at once: replace pairwise predictions with a trainable equivariant function per capsule-type—a 'space-of-variation'—implemented with group-equivariant convolutions, and route by degree centrality in a graph of predictions. The paper proves that this specific routing layer is equivariant (Theorem 2.1) and that an equivariant routing procedure in its general family preserves the capsule-decomposition graph up to isomorphism under input transformations (Theorem 2.2). If correct, the result gives capsule networks a provable compositional invariance and makes deeper capsule architectures practical, because the number of prediction networks grows with capsule types rather than with all type pairs. Experiments on transformed MNIST, FashionMNIST, CIFAR-10, KMNIST, and SVHN show SOVNET beating several capsule baselines on transformation robustness and matching CNNs in some settings.","feed_headline":"Rotating the input rotates this capsule network's parse tree","feed_subtitle":"Group convolutions and degree-based routing keep the part-whole graph in sync with the input.","key_machinery":"The central object is the space-of-variation (SOV): a trainable, vector-valued function $\\Psi_j^{l+1}$ on a group $G$ that each capsule-type uses to map shallower capsules into predictions for a deeper capsule, replacing the pairwise prediction networks of standard capsule nets. The layer is assembled from four equivariant pieces: the group-equivariant correlation $f_i^l \\star \\Psi_j^{l+1,p}$, the degree-centrality routing weights computed from cosine similarities among predictions, the squash non-linearity, and the weighted sum that produces deeper capsules. The paper defines the capsule-decomposition graph $G(x)$ whose vertices are capsules (type, position, pose, activation) and whose directed, weighted edges are routing coefficients; this graph is the formal object whose isomorphism under transformations is the paper's compositional guarantee.","core_discovery":"The central claim is that a capsule network can be built so that its routing layer is equivariant to a group $G$ of transformations, and that this equivariance is enough to preserve the compositional parse it computes. Theorem 2.1 states that for the layer defined in Algorithm 2, with group-equivariant correlation as the prediction mechanism and degree-centrality as the routing rule, $([L_gF^l]\\otimes \\Psi_j^{l+1}) = L_g([F^l \\otimes \\Psi_j^{l+1}])$: transforming the input first and then routing is the same as routing first and then transforming. Theorem 2.2 states that any network whose routing belongs to the summation family of Algorithm 1 and is equivariant has isomorphic capsule-decomposition graphs for $x$ and $L_gx$, where the isomorphism sends a capsule at group position $g$ to the same type at $h^{-1}\\circ g$. The paper reads this as a formal guarantee that detected part-whole relationships, with their routing strengths, are preserved under transformations.","pith_inferences":["The theorem's scope is layer-level: the full-network guarantee is only as strong as the equivariance of the primary-capsule construction, residual blocks, and output pooling, none of which receives a proof in the paper.","The p4 group used in experiments is equivariant only to rotations in multiples of 90°, so SOVNET's improved accuracy on 30° and 60° test rotations is an empirical extrapolation, not something the proof promises.","A direct empirical test of Theorem 2.2—extract routing coefficients for $x$ and $L_gx$ and check the graph isomorphism—is not reported; it would be a cheap and decisive validation.","The same SOV design could in principle be combined with learned or continuous group grids (for example, finer roto-translation grids), and the routing rule would remain equivariant; the open question is the computational cost of the grid."],"forward_implications":["A layer needs one prediction network per capsule-type rather than one per pair of types, so the parameter count scales with the number of types, not the product of type counts.","Routing has no iteration-count hyperparameter, unlike dynamic routing and EM routing, so there is one less schedule to tune when building deeper models.","Because the layer is equivariant, training on geometrically transformed data should teach the same capsule decomposition at transformed positions, which is what the improved train-time augmentation results show.","The capsule-decomposition graph provides a concrete certificate: any equivariant routing rule from the Algorithm 1 family automatically yields isomorphic graphs under the group, so future routing rules can be checked by checking equivariance."],"supporting_citations":[{"why":"Supplies the group-equivariant convolution and correlation operator $f \\star \\Psi$, including the equivariance result that Theorem 2.1 steps through.","marker":"((Cohen and Welling, 2016))"},{"why":"Defines dynamic routing and the squash non-linearity used in Algorithm 2, and provides the margin and reconstruction losses used in training.","marker":"((Sabour et al., 2017))"},{"why":"Defines EM routing and convolutional capsule layers, the capsule baseline and comparison point for SOVNET's prediction-phase design.","marker":"((Hinton et al., 2018))"},{"why":"DeepCaps is the deep-capsule baseline SOVNET is measured against, and supplies the margin-loss schedule adopted for the second half of training.","marker":"((Rajasegaran et al., 2019))"},{"why":"Group-equivariant capsule networks formalize capsules as group elements; SOVNET's discussion of the geometric-only versus convolutional-equivariance tradeoff is framed against this work.","marker":"((Lenssen et al., 2018))"},{"why":"Spectral routing is cited as an example of a routing family outside the summation-based Algorithm 1, marking the scope of Theorem 2.2.","marker":"((Bahadori, 2018))"}],"fun_headline_variants":["Equivariant routing: capsule parse trees that rotate with the input","Keep capsule graphs in sync: group-equivariant routing","Capsule nets with provable equivariance for transformations","Transformations don't break this capsule network's parse"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The full-network guarantee depends on the primary capsule layer, the residual blocks, and the output pooling also being equivariant, but the paper proves equivariance only for the routing step in Algorithm 2.","fun_headline_variants_meta":{"raw":{"variants":["Equivariant routing: capsule parse trees that rotate with the input","Keep capsule graphs in sync: group-equivariant routing","Capsule nets with provable equivariance for transformations","Transformations don't break this capsule network's parse"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000818,"raw_usage":{"total_tokens":3613,"prompt_tokens":1008,"completion_tokens":2605,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":2538}},"tokens_in":624,"tokens_out":2605,"duration_ms":18606,"temperature":1.0,"reasoning_tokens":2538,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:17:51.493193+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a trained SOVNET on an input $x$ and on a transformed input $L_gx$ with $g$ in the group used by the convolutions, extract the routing coefficients $c_{ij}(g)$ from both runs, and check whether each edge weight obeys $c_{ij}(h^{-1}\\circ g)$; a mismatch beyond floating-point tolerance would falsify Theorem 2.2 for the full architecture. A second check is layer-level: compare $([L_gF^l]\\otimes \\Psi_j^{l+1})$ with $L_g([F^l \\otimes \\Psi_j^{l+1}])$ for random input capsules and random equivariant filters.","supporting_citations":[],"review_version":1}