{"id":"a37f135f-98ba-4db3-be94-33158464fa23","arxiv_id":"2507.08404","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Semantic hash centers, derived from a data-dependent classifier-based similarity matrix, improve deep hashing image retrieval over distance-only hash center methods like CSQ and MDS.","lead":"This paper proposes SHC, a deep hashing method that generates hash centers encoding semantic relationships between image classes, learned from a classifier's confusion patterns. On five image retrieval benchmarks it reports consistent improvements over state-of-the-art methods, averaging 7 to 12 percent higher mean average precision.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Optimization in Eq. 29/35 appears internally inconsistent; central hash-center generation step not reproduceable from the paper.","rationale":"The reader identified a real concern about the data-dependent similarity matrix being potentially noisy or uninformative. However, the most load-bearing weakness is more basic: even if the similarity matrix is meaningful, the provided optimization in Stage 2 has equations that are internally inconsistent/dimensionally wrong, and no code is provided. That means the central mechanism generating the semantic hash centers cannot be verified from the manuscript. The concern is concrete and technical (Eq. 29 and 35). The downstream empirical claim is broad and admits several other concerns (no code, no error bars, possible contradiction in Table 3), but the optimization gap is the load-bearing one: if Stage 2 cannot be executed as written, the whole method—not just the similarity matrix—is under-specified. A test that re-derives or runs the center generation would settle it. The reader's weakest_assumption did not identify this specific issue; it is related but more fundamental, so agreement is partial.","tokens_in":22968,"tokens_out":1518,"duration_ms":15566,"concrete_test":"Implement Stage 2 (Algorithm 1) exactly from the printed equations in the paper for CIFAR-100, 16 bits: initialize H by MDS, then iterate with the described M, K, h, lambda, alpha updates. If Eq. 29 is used as written, the update fails or produces dimensions mismatch; correcting it in the natural way (replacing rho h_i by rho H) and similarly correcting Eq. 35, check whether the resulting semantic loss and minimum distance reproduce Table 3 (SHC: d_min=4, S_loss=0.0545). A successful reproduction requires the corrected equations, the exact hyperparameters, and T=20 iterations.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim—that SHC's hash centers are semantically organized and distinct—rests on Stage 2 optimization (Eq. 22–38). Yet the printed updates contain apparent inconsistencies. In Eq. 29, the matrix form for M includes “rho h_i” where the vector h_i is used inside a matrix expression (M = (2/q^2 H H^T + rho I)^{-1}(2/q H S + Λ + rho h_i)), which is dimensionally inconsistent: the product requires a matrix, not a column vector. In Eq. 35, the gradient for h_i mixes terms with internal dimensions and signs: the term “beta_{ij}[2 h_j h_j^T h_i - 2(q-2d-k_{ij} h_j)]” has first part vector (h_j h_j^T h_i is a vector) but second part is also vector (k_{ij} h_j) while q-2d is scalar, and the overall expression seems to have an extra 2\\mu factor before the alpha term and possibly misplaced signs. These are not mere typos in constants; they make it unclear how the discrete centers actually minimize Eq. 22, which is the method's core contribution. Moreover, Eq. 29 is used in Algorithm 1 but cannot be executed as written. The paper promises code release but provides no URL. Therefore the main claim is currently not independently verifiable: a reader cannot re-implement Stage 2 and confirm that SHC produces centers with the values in Table 3 or the downstream gains in Tables 4–6.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SHC, a three-stage deep hashing framework. Stage 1 builds a data-dependent inter-class similarity matrix S from the softmax probabilities of a pre-trained classifier, after masking each sample's top prediction and averaging within classes. Stage 2 generates binary semantic hash centers by optimizing a regularized objective that combines a semantic fidelity term with a minimum-Hamming-distance term, using an augmented Lagrangian scheme with alternating updates. Stage 3 trains a deep hashing network with a central similarity loss and quantization loss, following CSQ. Experiments on CIFAR-100, Stanford Cars, and NABirds (five dataset variants) compare against nine baselines and report average MAP improvements of +7.26%, +7.62%, and +11.71% over the second-best method for MAP@100, MAP@1000, and MAP@ALL, respectively.","tokens_in":23273,"tokens_out":4535,"duration_ms":50809,"significance":"The paper's core idea is attractive and potentially useful: unlike CSQ and MDS, which assign hash centers by data-independent constructions, SHC attempts to make centers reflect real inter-class semantic relatedness while preserving a minimum Hamming separation. The three-stage pipeline is clearly described, and the experimental suite is broad, spanning five dataset variants, three code lengths, nine baselines, ablations, and convergence plots. I found no circularity: the similarity matrix is derived from classifier outputs on training data, the centers are optimized against that matrix, and retrieval is evaluated on held-out queries; the reported MAP is not used to set constants in the method. However, the central hash-center generation step is not reproducible from the printed equations, and one of the paper's headline claims about the generated centers is contradicted by its own table. These issues are load-bearing and must be fixed before the results can be accepted as reported.","major_comments":[{"comment":"Equation (29) is dimensionally inconsistent and cannot be executed as written. The matrix form reads M = (2/q^2 H H^T + ρ I)^{-1}(2/q H S + Λ + ρ h_i), where h_i is a column vector of dimension q but the other two terms in the parentheses are q×C matrices. The scalar update in Eq. (28) suggests the intended matrix form should have ρ H in the last term, not ρ h_i. Because Algorithm 1 calls Eq. (29) directly, this typo blocks re-implementation of the central hash-center generation step.","section":"Section 3.2.3, Eq. (29)"},{"comment":"Equation (35) contains several apparent algebraic errors in the gradient with respect to h_i. The λ term is written as λ_i^T but should be the column vector λ_i; the α-related term appears as −2 μ Σ α_ij h_j, although differentiating ∑ α_ij(q − 2d − h_i^T h_j − k_ij) with respect to h_i gives −Σ α_ij h_j with no factor 2μ; and the β term is written as β_ij[2 h_j h_j^T h_i − 2(q − 2d − k_ij h_j)], where the second part appears to conflate the scalar (q − 2d − k_ij) with the vector h_j. As printed, this is not the gradient of the objective in Eq. (34), so the projected gradient update in Eq. (36) is not well defined. The authors should supply corrected derivations and verify them against their implementation.","section":"Section 3.2.3, Eq. (35)"},{"comment":"The text states that SHC 'achieves the largest d_min and the smallest S_loss across all datasets,' but Table 3 contradicts this on the d_min point. For example, on CIFAR-100 at 64 bits, MDS has d_min = 32 while SHC has d_min = 24; on Stanford Cars-A at 64 bits both have d_min = 23; on NABirds-A and NABirds-B at 64 bits both have d_min = 21. Thus the claim that SHC sets hash centers 'as far apart as possible' is not supported by the reported numbers. The authors should either correct the claim, report the full trade-off between d_min and S_loss, or explain why a lower or equal d_min with a much lower S_loss is the desirable operating point.","section":"Section 4.5.2 and Table 3"},{"comment":"All MAP numbers in Tables 4–6 are reported without standard deviations or error bars, and no code or dataset release URL is provided despite the statement that 'All the curated datasets and codes will be released on Github.' Given the unresolved issues in Eqs. (29) and (35), the reported numerical results cannot currently be independently verified. The authors should report variability across runs and make the implementation available, at least as supplementary material, so that Stage 2 can be reproduced.","section":"Sections 4.3–4.5 and 4.4"}],"minor_comments":[{"comment":"The heading contains the typo 'Gibert-Varshamov bound'; it should be 'Gilbert-Varshamov bound'.","section":"Section 3.2, heading"},{"comment":"The formula for d_min, d_min = min{(q − H^T H ⊙ (1 − I))/2}, is ambiguous because the elementwise product with (1 − I) appears inside the min without clearly indicating a minimum over off-diagonal entries. Please add the missing indexing, e.g., d_min = min_{i≠j} (q − (H^T H)_{ij})/2.","section":"Eq. (44)"},{"comment":"References [11] and [12] cite the same BERT paper with different years and venues; one of them should be removed or both should be unified to a single canonical reference.","section":"References [11] and [12]"},{"comment":"The text says the Precision/Recall curves use 'a large range from 1, 5, 10, ..., to 500' for topK; Precision-Recall curves are not conventionally parameterized by topK, so this sentence needs clarification.","section":"Section 4.2"},{"comment":"The reported improvement ranges, e.g., '(+0.13% ~ +11.53%)' and '(+7.43% ~ +22.64%)', do not match the three metric families listed in the sentence; it would help to label which range corresponds to MAP@100, MAP@1000, and MAP@ALL.","section":"Section 4.5.3"}],"recommendation":"major_revision","confidential_remarks":"The paper's main contribution is potentially sound and the experimental study is extensive, but the central optimization equations are not reproducible as printed and Section 4.5.2 contains a factual contradiction with Table 3. I believe these are fixable within the scope of a revision: the authors can correct Eqs. (29) and (35), provide a derivation or code, and rephrase the d_min claim. If the code is not supplied and the equations remain ambiguous, I would escalate to rejection on reproducibility grounds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick take on arXiv:2507.08404. The core idea—semantic hash centers that respect inter-class similarity while keeping a minimum Hamming separation—is a natural and reasonable extension of CSQ and MDS. The data-dependent similarity matrix built from classifier softmax confusions is the genuine new bit, and the three-stage pipeline is well structured. The experiments are broad: five dataset variants, three code lengths, nine baselines, and the reported MAP gains are consistent and large enough to matter.\n\nBut the math at the center of the method does not hold up as printed. Eq. 29 mixes a vector h_i into a matrix update, and Eq. 35 has multiple dimension and sign problems—the alpha term gets a spurious 2-mu factor, and the beta term subtracts a vector from a scalar. These are not cosmetic. Algorithm 1 depends on these updates, so a reader cannot reimplement Stage 2 from the paper. Also Section 4.5.2 claims SHC achieves the largest d_min, but Table 3 shows MDS beating or tying SHC at 64 bits (e.g., CIFAR-100: MDS 32 vs SHC 24). No code or data link is given despite the promise, and there are no error bars or significance tests. The ablation and BERT comparison are useful, and the convergence curves are relevant.\n\nThe circularity worry is unfounded: nothing in the method fits MAP; the similarity matrix is computed on training images and retrieval is on held-out queries. The classification accuracies are modest (NABirds ~50%), but the within-class averaging idea is plausible, though not deeply analyzed.\n\nNet: the conceptual contribution is sound and the experimental trend is convincing, but the central optimization and the contradiction need fixing. If code were released and the equations corrected, this would be a solid practical paper. I'd send it to peer review rather than desk-reject, but with a strong request for code release and corrected derivations.","headline":"A sensible, potentially useful extension of hash-center methods, but the core optimization is broken as printed and the central quality claim is overstated.","tokens_in":23816,"tokens_out":4055,"would_cite":false,"duration_ms":44754,"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":"Semantic hash centers that place related classes close in Hamming space improve deep-hashing image retrieval by several points of mean average precision.","keywords":["learning to hash","semantic hash centers","image retrieval","deep hashing","Hamming distance","similarity matrix","quantization","classification softmax"],"falsifier":"Replace the Stage 1 similarity matrix S with a random symmetric matrix that has the same diagonal and the same minimal distance d, rerun the full SHC pipeline, and compare MAP: if retrieval accuracy does not fall toward the level of the minimum-distance-only baseline, the semantic constraint is not what produces the reported gains.","tokens_in":22732,"feed_emoji":"🔍","tokens_out":9890,"duration_ms":113522,"temperature":0.7,"pith_summary":"This paper tries to establish that the fixed binary codes assigned to classes in deep hashing should be arranged by meaning, not just by mathematical convenience. Its hypothesis is that semantically related classes deserve hash centers with small Hamming distance, while unrelated classes should sit far apart. To test that, the paper builds a three-stage pipeline: it derives a class-similarity matrix from a pre-trained image classifier's softmax responses, solves a constrained optimization that produces binary centers respecting both those similarities and a minimum Hamming separation, then trains a hashing network to push images toward their class centers. If the hypothesis is right, large-scale image retrieval should become noticeably more accurate, and the paper reports average gains of about +7.26%, +7.62%, and +11.71% in mean average precision over strong existing deep hashing methods.","feed_headline":"Semantic hash centers lift image retrieval gains by up to 11.7%","feed_subtitle":"SHC places related classes closer in Hamming space and lifts mean average precision across five standard benchmark setups","key_machinery":"The load-bearing object is the semantic hash center: a binary vector $h_c \\in \\{-1,+1\\}^q$ assigned to each class $c$, generated as a solution of $$\\min_H \\|S - \\tfrac{1}{q} H^\\top H\\|$_F^{2}$ + \\mu\\sum_{i \\ne j} h_i^\\top h_j \\quad \\text{s.t. } h_i^\\top h_j \\le q - 2d,\\; h_i \\in \\{-1,+1\\}^q,$$ where $S$ is the Stage 1 data-dependent similarity matrix and $d$ is the minimal Hamming distance obtained from the Gilbert-Varshamov bound. The identity that makes the optimization tractable is the Hamming-Euclid relation $D(h_i, h_j) = \\tfrac{1}{2}(q - h_i^\\top h_j)$, which converts separation in Hamming space into an inner-product inequality. The paper solves the NP-hard binary problem by introducing proxy variables and using an Augmented Lagrangian scheme, alternating updates for the proxy, slack, center, and multiplier variables.","core_discovery":"On its own terms, the paper's central claim is that the internal distance geometry of hash centers is a reusable semantic resource. Prior point-wise methods generate centers from combinatorial constructions and then treat any large separation as good; SHC instead computes, for each class pair, a data-dependent similarity score from the classifier's confusions, and then solves for centers in which normalized inner products approximate those scores while no two centers come closer than the Gilbert-Varshamov minimum distance $d$. Given those centers, a standard deep hashing network trained with a center-similarity and quantization loss produces binary codes that inherit the semantic layout. Across five dataset configurations and code lengths 16, 32, and 64, this pipeline reports the best mean average precision among the methods compared, with the interaction between class count and code length shaping how large the improvement is.","pith_inferences":["Beyond the paper, the same three-stage construction should transfer to any retrieval domain with a pre-trained classifier and discrete codes, such as audio, text, or graph retrieval.","Because the similarity matrix inherits the Stage 1 classifier's confusions, the hash centers will encode whatever drives those confusions; on datasets where confusions track background, pose, or lighting rather than category semantics, the centers will encode that structure too.","The observed pattern of gains suggests a testable prediction: at a fixed code length, the advantage of semantic centers should grow with the number of classes, since a crowded Hamming space is exactly where layout-by-meaning matters most.","A natural next experiment is to compare the learned class graph against an explicit human taxonomy; divergences on fine-grained classes would point to the similarity estimate, not the center optimization, as the bottleneck."],"forward_implications":["On the five dataset configurations in the paper, SHC beats all nine compared deep hashing methods, with average MAP gains of +7.26% at MAP@100, +7.62% at MAP@1000, and +11.71% at MAP@ALL over the strongest baselines.","Ablations removing either the semantic constraint or the minimum-distance constraint lower MAP in most settings, so both terms in the center-generation objective contribute to the result.","Swapping the data-dependent similarity matrix for a label-text embedding matrix lowers retrieval performance, supporting the claim that adapting the similarity estimate to the data distribution is a real component of the gain.","The size of the improvement depends on how crowded the Hamming space is: gains are clearest at 32-bit codes for the 196-class Stanford Cars setting and grow with code length for the 555-class NABirds setting."],"supporting_citations":[{"why":"Introduces the hash-center family of point-wise methods and supplies the central-similarity plus quantization loss used to train the Stage 3 hashing network; it is also the main comparison point the paper improves on.","marker":"[42]"},{"why":"Defines minimal-distance-separated hash centers using the Gilbert-Varshamov bound, provides the initialization for H in Algorithm 1, and serves as the ablation baseline without the semantic constraint.","marker":"[37]"},{"why":"Provides the proof of the Gilbert-Varshamov bound that the paper uses to compute the minimum allowable Hamming distance d from the class count and code length.","marker":"[34]"},{"why":"Supplies the augmented Lagrangian framework that lets the paper decompose the binary-constrained center generation into alternating sub-problems.","marker":"[1]"},{"why":"Provides the pre-trained language-model embeddings of class labels used as the label-text alternative similarity matrix in the ablation study.","marker":"[12]"}],"fun_headline_variants":["Semantic hash centers boost image retrieval MAP up to 11.7%","Hash centers that respect class semantics improve retrieval","Data-driven semantic hash centers lift MAP by up to 11.7%","Semantic hash centers: up to +11.7% MAP on image retrieval"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result rests on whether the Stage 1 similarity matrix really captures which classes are semantically related; if the classifier's averaged softmax similarities are mostly noise, the semantic constraint has no true signal to encode.","fun_headline_variants_meta":{"raw":{"variants":["Semantic hash centers boost image retrieval MAP up to 11.7%","Hash centers that respect class semantics improve retrieval","Data-driven semantic hash centers lift MAP by up to 11.7%","Semantic hash centers: up to +11.7% MAP on image retrieval"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000365,"raw_usage":{"total_tokens":1994,"prompt_tokens":1002,"completion_tokens":992,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":915}},"tokens_in":618,"tokens_out":992,"duration_ms":8380,"temperature":1.0,"reasoning_tokens":915,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:20:05.492156+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the Stage 1 similarity matrix S with a random symmetric matrix that has the same diagonal and the same minimal distance d, rerun the full SHC pipeline, and compare MAP: if retrieval accuracy does not fall toward the level of the minimum-distance-only baseline, the semantic constraint is not what produces the reported gains.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the hash-center family of point-wise methods and supplies the central-similarity plus quantization loss used to train the Stage 3 hashing network; it is also the main comparison point the paper improves on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines minimal-distance-separated hash centers using the Gilbert-Varshamov bound, provides the initialization for H in Algorithm 1, and serves as the ablation baseline without the semantic constraint."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the proof of the Gilbert-Varshamov bound that the paper uses to compute the minimum allowable Hamming distance d from the class count and code length."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the augmented Lagrangian framework that lets the paper decompose the binary-constrained center generation into alternating sub-problems."}],"review_version":1}