{"id":"9805f69f-904d-4109-a9bd-cbfd34c05098","arxiv_id":"1908.07388","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A cross-modal zero-shot hashing method that jointly learns deep features, category attribute spaces, and hash codes, using a composite similarity to exploit labeled and unlabeled data with different label spaces.","lead":"Researchers propose CZHash, a hashing method for cross-modal image-text retrieval that handles new categories, partially labeled data, and different label sets across modalities. It combines deep feature learning, category attribute embedding, and hash coding into one objective and reports accuracy gains over several baselines on three public datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never defines a hash function H(x) for unseen queries: the objective in Eq. (7) optimizes per-instance auxiliary matrices C(v), and no test-time mapping from a new feature vector to a binary code is specified.","rationale":"The strongest claim is that CZHash is a general cross-modal zero-shot hashing method with significant performance gains. The load-bearing condition is that a trained model can hash unseen-category queries. That condition is not met in the written method: after optimizing Eq. (7), the only variables available to map a new input to a code are the deep feature function φ(·), attribute matrices A(v), and coefficient matrices W(v). But the objective never constrains φ(x) to determine the auxiliary matrix C; C is a free optimization variable of size n×c. Without C, the quantization loss ||C(v)W(v) − B||^2 cannot be evaluated at test time, and B = sign(CW) has no test-time analogue. Section III-A promises hash functions H1 and H2, yet no such functions appear in Section III-C; the optimization updates only the training code matrix B. This is not merely a notation gap: it means the entire experimental section cannot be reproduced from the equations, and the reported MAP scores could only be produced by an unspecified extra inference step. The word2vec semantic-space issue is a softer empirical assumption; even if the attribute space were perfectly calibrated, the method would still lack a way to project a new sample into it. I therefore focus on the missing inference mapping. The reader's rationale already lists 'zero-shot inference step is missing' as one of the reasons for rejection, though the reader's weakest_assumption was stated differently. My read converges on the same verdict: the paper as submitted does not describe a coherent, reproducible method, so the rejection stands. If the authors can supply an explicit inference equation or code, the paper might be salvageable after major revision.","tokens_in":18159,"tokens_out":5354,"duration_ms":52492,"concrete_test":"Write out the inference rule: take one trained model and one held-out unseen-category query x, then compute its hash code using only quantities defined after optimizing Eq. (7), namely φ(x), A(v), W(v), and constants. If no such expression exists, any proposed evaluation must introduce a new variable or network (e.g., solving C = F A^T (A A^T)^{-1}, or a separate hash network). A minimal check is to re-run the Table V protocol using only the equations in Section III; if the procedure cannot be implemented without inventing an extra inference step, the reported MAP values are not attributable to the stated model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires a working zero-shot hashing procedure: train on seen classes, then encode a never-seen query as a binary code. The paper does not supply one. In Eq. (7), the learned entities are deep feature functions φ(x), auxiliary category-space matrices C(1), C(2) ∈ R^{n×c}, attribute matrices A(v), coefficient matrices W(v), and a training-code matrix B. The reconstruction term ||F(v) − C(v)A(v)||^2 treats C(v) as a free per-instance variable; it is not a function of F(v), and no closed-form mapping is derived from it. The quantization term ||C(v)W(v) − B||^2 is likewise only defined on the n training instances. At test time, for a query x from an unseen category, the paper provides no way to compute C(v) or to convert F(v) alone into a code. The hash functions H1 and H2 promised in Section III-A are never defined, and the optimization section (III-C.4) only updates B for the training matrix. The reported experiments in Tables IV–VII therefore rely on an unstated evaluation protocol; without that protocol, the significant-outperformance claim is not executable. The gradient equations (8)–(12) cannot repair this because they only update training variables and contain dimensional inconsistencies, such as S22 appearing in ∂Jloss/∂F(1) and B ∈ R^{n×c} in Eq. (11) versus R^{n×b} elsewhere.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CZHash, a cross-modal zero-shot hashing method that jointly learns deep feature representations, category attribute spaces, and hashing quantization in a unified objective (Eq. (7)). It introduces a composite similarity measure (Eqs. (2)-(3)) that combines label and feature similarities to handle labeled and unlabeled data, and it allows different label spaces across modalities. The authors report MAP improvements over several cross-modal and zero-shot hashing baselines on Nus-wide, Wiki, and Mirflickr under four scenarios (Tables IV-VII). The central problem, however, is that the paper never specifies how a query from an unseen category is encoded into a binary hash code at test time, and the optimization details in Section III-C.4 are internally inconsistent with the stated objective.","tokens_in":18556,"tokens_out":4285,"duration_ms":44668,"significance":"If the method were fully specified and reproducible, it would address a genuinely useful and under-studied setting: cross-modal zero-shot hashing with semi-supervised data and per-modality label spaces. The problem formulation is well motivated, and the experimental design covers several realistic scenarios. The paper also makes a reasonable attempt to leverage unlabeled data through the composite similarity. However, the significance assessment hinges entirely on whether a test-time hashing procedure exists; since the manuscript does not define one, the claimed contributions are not currently executable or verifiable.","major_comments":[{"comment":"The paper never defines a test-time hash function for unseen queries. Section III-A promises two hashing functions H1 and H2, but they are never specified. In Eq. (7), the learned variables include per-instance matrices C(1) and C(2) for the n training instances, and the optimization in Section III-C.4 updates only these training variables. For a new query x from an unseen category, there is no rule to compute C(v) from the deep feature F(v), and no closed-form mapping from F(v) alone to a binary code is derived. The quantization term ||C(v)W(v) - B||^2 is defined only for the training instances. Thus the zero-shot retrieval procedure promised in the abstract and experiments is not defined, and the results in Tables IV-VII rely on an unstated evaluation protocol.","section":"III-A and III-C.4"},{"comment":"The reported gradients are inconsistent with the objective in Eq. (7). For example, the gradient w.r.t. F(1) in Eq. (8) contains an unexplained 1/2 factor and a term involving S22, but the first term of Eq. (7) for the (1,1) and (1,2) pairs does not produce an S22 contribution in the derivative w.r.t. F(1). The derivative of ||F(1)-C(1)A(1)||^2 w.r.t. C(1) should be -2(F(1)-C(1)A(1))A(1)^T, whereas Eq. (9) has a 1/2 factor and no factor of 2 on the alpha term; the beta term is also missing a factor of 2. Moreover, Eq. (11) states B in {−1,1}^{n×c}, while Eq. (6) and Eq. (7) define B in R^{n×b}, where b is the code length and c is the number of categories; these are different dimensions. These inconsistencies mean that the optimization procedure as written is not a well-defined minimization of Eq. (7).","section":"III-C.4, Eqs. (8)-(11)"},{"comment":"The experimental section does not specify how queries are encoded in the zero-shot scenarios. The text describes splitting categories into seen and unseen and testing on the unseen categories, but it never states whether the learned mapping is applied to the unseen query's feature vector, whether the attribute vector of the unseen category is used, or how the hash code is generated. The MAP definition in Eq. (13) does not clarify the protocol. Without a precise test-time inference rule, the reported performance numbers in Tables IV-VII cannot be reproduced or attributed to the model described in Section III. This is a load-bearing omission because the central claim of the paper is that CZHash can hash unseen categories.","section":"IV-B, Tables IV-VII"},{"comment":"There is an inconsistency between the objective in Eq. (4) and the unified objective in Eq. (7). Eq. (4) contains three terms, with the cross-modal term weighted by 2, while Eq. (7) writes a double sum over v,v'=1,2, which would produce four terms (S11, S12, S21, S22) with equal weighting. If S12 is intended to be symmetric and counted twice, the relationship to Eq. (4) should be stated explicitly. Additionally, the claim in Section III-C.2 that A(1) and A(2) can be set to their union 'without loss of generality' assumes that categories absent from one modality do not distort that modality's learned category space; this assumption is not justified, and it affects the zero-shot transfer mechanism.","section":"III-C.2 and Eq. (4) vs. Eq. (7)"}],"minor_comments":[{"comment":"The table is badly garbled: the check/cross marks do not align with the column headers, and the row for 'PRDH' contains a placeholder '[?]'. This makes the comparison table difficult to interpret.","section":"Table I"},{"comment":"The text introducing Tables VI-VII refers to a method 'DRCMH', but the method described and listed elsewhere is 'RDCMH'. Please correct the typo.","section":"IV-B"},{"comment":"There are several typos and inconsistent notations: 'beacuase' (Section IV-E), 'Mirfilcker' (Fig. 2), 'CZSHash' (Conclusion), 'ZCHash' (Section III-C.2), 'non-liner' (Section III-C.1), and the duplicated 'θ(1)_x and θ(1)_x' in Section III-C.1. These should be corrected.","section":"Throughout"},{"comment":"The axis labels in Figure 2 are garbled and overlapping; the figure should be redrawn with clear tick labels for alpha and beta.","section":"Fig. 2"},{"comment":"The text says Eq. (7) is solved via ADMM, but the subsequent description uses SGD and back-propagation. The connection between these optimization strategies should be clarified, or the ADMM reference should be removed.","section":"III-C.4"}],"recommendation":"reject","confidential_remarks":"The missing test-time hash function is not a local fix: it requires redefining the method's inference procedure and re-running all experiments. The gradient inconsistencies in Section III-C.4 further indicate that the optimization, as written, is not the minimization of Eq. (7). This is beyond the scope of a normal major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper proposes CZHash, a cross-modal zero-shot hashing method that handles different label spaces across modalities and semi-supervised data. That problem setting is genuinely new as far as I know—AgNet does cross-modal ZSH but assumes shared labels and full supervision, so the authors are filling a real gap. The composite similarity (Jaccard on labels, combined with feature similarity) is reasonable, and jointly optimizing deep features, category-space embeddings, and hash codes is a sensible high-level design. The experimental matrix covers four scenarios (fully supervised, zero-shot, semi-supervised ZSH, and different label spaces), which is the right way to test the claim.\n\nThe central method, however, is not actually specified. Section III-A promises hash functions H1 and H2, but they are never defined. Eq. (7) optimizes F(v), C(v), W(v), and B on the n training instances. C(v) is a free per-instance matrix, and there is no formula that takes a new query feature vector and outputs a code. The optimization section only updates training variables. So the paper does not tell us how to encode a never-seen query, which is the whole point of zero-shot hashing. This is not a cosmetic gap; the experiments in Tables V–VII must rely on an unstated evaluation protocol.\n\nThe optimization details are also inconsistent. Eq. (8) includes S22 in the gradient w.r.t. F(1), which does not appear in the objective's derivative; Eq. (9) is missing the expected -2α(F − CA)A^T term and has a wrong sign/coefficient; Eq. (11) writes B ∈ {±1}^{n×c} when b is the code length. A reader cannot reproduce the training procedure from these equations.\n\nThere is also a numerical problem with the headline claim. On Wiki text-to-image, SePH beats CZHash by a wide margin in Table IV (0.535 vs. 0.302 at 16 bits) and in Table V (0.498 vs. 0.253). The paper mentions SePH does well on Wiki in one sentence but does not square that with the abstract's claim of significant superiority. That is a real contradiction, not a rounding error.\n\nThe related work is cited fairly and the comparison methods are standard. But the missing test-time mapping and the internal inconsistencies are load-bearing. As submitted, the manuscript is not reproducible.\n\nWho is this for? Someone working on zero-shot hashing might find the problem formulation useful, and the experimental setup is worth borrowing. But the method needs a major rewrite: define the hash functions, fix the gradients, clarify the evaluation protocol, and reconcile the Wiki results. I would only send it to review if the authors can supply the missing inference step and code; without those, a referee would be reviewing a promise.\n\nRecommendation: desk reject unless the authors provide the inference procedure and code. If they do, it deserves a careful look.","headline":"The problem setting is real and the high-level design is sensible, but the paper never defines a hash function for unseen queries, the gradients don't match the objective, and the reported gains are reversed on at least one benchmark configuration.","tokens_in":18995,"tokens_out":3737,"would_cite":false,"duration_ms":36696,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CZHash claims that a single joint objective—feature learning, category attribute spaces, and hash quantization—lets cross-modal hashing work for unseen categories with partially labeled data and differing label spaces.","keywords":["cross-modal hashing","zero-shot hashing","composite similarity","category attribute space","semi-supervised learning","multi-label data","different label spaces","deep feature learning"],"falsifier":"Hold all training structure fixed but replace the Word2Vec category attribute vectors with random vectors of the same dimension (or with permuted vectors). If MAP for unseen-class queries does not drop to near chance, the zero-shot transfer is not actually mediated by the semantic attribute space the paper describes; if it does drop, the method's promise hinges on the quality of the category-name embedding.","tokens_in":18003,"feed_emoji":"🔎","tokens_out":8189,"duration_ms":70992,"temperature":0.7,"pith_summary":"This paper proposes CZHash, a cross-modal hashing scheme that handles three complications at once: data in two or more modalities, categories that never appear during training, and label sets that differ between modalities while most instances are unlabeled. The central claim is that one unified objective—deep feature learning guided by a composite label-and-feature similarity, category attribute space learning, and hash-code quantization—lets the model transfer knowledge from seen to unseen categories. A reader should care because real retrieval systems face exactly this mix: partially labeled multi-modal data with new concepts arriving over time, while existing methods assume away at least one of these conditions. The paper reports that CZHash outperforms seven representative hashing methods on three benchmark image-text datasets across four experimental scenarios.","feed_headline":"Cross-modal hashing reaches unseen categories without retraining","feed_subtitle":"CZHash blends labeled and unlabeled image-text data with different label sets, beating seven baselines in zero-shot retrieval.","key_machinery":"The load-bearing objects are the composite similarity matrices and the unified objective. Within a modality, composite similarity supplements Euclidean feature similarity with Jaccard label similarity when both instances are labeled; across modalities, feature similarity is estimated by the average of paired within-modality similarities, and label overlap is added when both sides are labeled. This construction lets unlabeled data and partially overlapping label spaces steer deep feature learning. The unified objective ties this similarity-preserving feature learning to category-space learning, where each modality's deep features are reconstructed by a category representation matrix times a shared attribute matrix built from the union label space, and to hash-quantization learning, where category-space codes are linearly mapped to a shared binary code matrix; a sign-based update gives the closed-form code assignment at each alternating step.","core_discovery":"The discovery, stated on the paper's own terms, is that zero-shot cross-modal hashing remains feasible when labeled data are scarce and the modalities' label spaces only partially overlap, provided label information is folded into a composite similarity rather than used as a hard semantic embedding. The paper builds a composite similarity that uses Jaccard overlap of labels when labels exist and Euclidean feature similarity otherwise, with inter-modality similarity approximated by the average of paired within-modality similarities. It then minimizes a unified objective that simultaneously preserves this similarity in deep feature spaces, projects features into a category attribute space built from class-level Word2Vec attribute vectors, and quantizes the projected codes into shared binary codes. The optimization alternates between feature networks, category-space coefficients, hashing coefficients, and the binary code matrix. On NUS-WIDE, Wiki, and MIRFlickr, the paper reports consistently higher mean average precision than the compared deep and non-deep methods, including the cross-modal zero-shot baseline AgNet, with the largest gains in the semi-supervised and different-label-space scenarios.","pith_inferences":["The paper fixes both modalities' attribute matrices to their union; an untested alternative is to weight each modality's contribution to the union, since a label present in only one modality could otherwise distort the other modality's learned category space.","Because the attribute space is built from category-name embeddings rather than hand-engineered attributes, the method is portable to new vocabularies; this also implies the text-embedding model's quality sets a ceiling on zero-shot transfer, so swapping the embedding model should measurably change MAP.","The composite similarity is computed on raw features before deep learning; replacing it with a self-supervised learned similarity on unlabeled data is a natural extension that could further reduce the need for labels.","The sign-based code update connects CZHash to binary embedding methods; a testable extension is whether the learned codes also serve as features for zero-shot classification, not only retrieval."],"forward_implications":["A deployed retrieval system could serve newly appeared categories without retraining the hash functions, provided it can obtain the new category's attribute vector (its name's embedding).","Partially labeled multi-modal collections with mismatched label vocabularies become usable for hash-code learning: unlabeled instances contribute through feature similarity, and labels from either modality contribute even when the label appears in only one modality.","Jointly optimizing feature learning, category-space learning, and quantization avoids the feature/quantization incompatibility the paper attributes to AgNet, and the reported MAP gains reflect this design.","The closed-form binary-code update at each iteration keeps the alternating optimization tractable on large collections with minibatch SGD.","The formulation is written for two modalities but the authors state it extends to m≥3 modalities, so multi-view zero-shot retrieval is a direct corollary of the setup."],"supporting_citations":[{"why":"Defines the cross-modal zero-shot hashing task and serves as the primary zero-shot baseline CZHash is compared against.","marker":"[17]"},{"why":"Deep cross-modal hashing baseline that preserves semantic similarity between modalities; CZHash extends its setting with unseen classes.","marker":"[24]"},{"why":"Semi-supervised deep cross-modal hashing baseline that also exploits labeled and unlabeled data, against which CZHash compares in the semi-supervised scenarios.","marker":"[11]"},{"why":"Zero-shot hashing method transferring supervised knowledge through label embeddings; CZHash's category-space learning is contrasted with this approach.","marker":"[15]"},{"why":"Supplies the Word2Vec category-name embeddings used as attribute vectors, the semantic bridge the zero-shot transfer depends on.","marker":"[35]"},{"why":"Provides the ADMM-style alternating optimization that CZHash uses to solve the unified objective.","marker":"[34]"},{"why":"Defines the CNN-F image network whose first six layers CZHash adopts for the image modality.","marker":"[33]"},{"why":"Probability-based semantics-preserving hashing baseline that achieves strong text-to-image results on Wiki and is included in the comparison.","marker":"[23]"}],"fun_headline_variants":["Cross-modal zero-shot hashing works with few labels","Mixed-label zero-shot hashing beats baselines across modalities","Unseen categories retrieved via composite similarity hashing","Cross-modal hashing that learns from partial label overlap","Zero-shot cross-modal hashing handles scarce labeled data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The zero-shot transfer rests on Word2Vec vectors of category names placing seen and unseen categories close enough that the learned mapping to hash codes generalizes; if those vectors misplace unseen classes, the hash codes for those classes are effectively arbitrary.","fun_headline_variants_meta":{"raw":{"variants":["Cross-modal zero-shot hashing works with few labels","Mixed-label zero-shot hashing beats baselines across modalities","Unseen categories retrieved via composite similarity hashing","Cross-modal hashing that learns from partial label overlap","Zero-shot cross-modal hashing handles scarce labeled data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1391,"prompt_tokens":995,"completion_tokens":396,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":320}},"tokens_in":611,"tokens_out":396,"duration_ms":4365,"temperature":1.0,"reasoning_tokens":320,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:39:46.958320+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold all training structure fixed but replace the Word2Vec category attribute vectors with random vectors of the same dimension (or with permuted vectors). If MAP for unseen-class queries does not drop to near chance, the zero-shot transfer is not actually mediated by the semantic attribute space the paper describes; if it does drop, the method's promise hinges on the quality of the category-name embedding.","supporting_citations":[{"cited_title":"Attribute-guided network for cross-modal zero-shot hashing,","cited_arxiv_id":null,"evidence_quote":"Defines the cross-modal zero-shot hashing task and serves as the primary zero-shot baseline CZHash is compared against."},{"cited_title":"Deep cross-modal hashing,","cited_arxiv_id":null,"evidence_quote":"Deep cross-modal hashing baseline that preserves semantic similarity between modalities; CZHash extends its setting with unseen classes."},{"cited_title":"Ranking-based Deep Cross-modal Hashing","cited_arxiv_id":"1905.04450","evidence_quote":"Semi-supervised deep cross-modal hashing baseline that also exploits labeled and unlabeled data, against which CZHash compares in the semi-supervised scenarios."},{"cited_title":"Zero-shot hashing via transferring supervised knowledge,","cited_arxiv_id":null,"evidence_quote":"Zero-shot hashing method transferring supervised knowledge through label embeddings; CZHash's category-space learning is contrasted with this approach."},{"cited_title":"Distributed representations of words and phrases and their compositionality,","cited_arxiv_id":null,"evidence_quote":"Supplies the Word2Vec category-name embeddings used as attribute vectors, the semantic bridge the zero-shot transfer depends on."},{"cited_title":"Return of the devil in the details: Delving deep into convolutional nets,","cited_arxiv_id":null,"evidence_quote":"Defines the CNN-F image network whose first six layers CZHash adopts for the image modality."},{"cited_title":"Cross-view retrieval via probability- based semantics-preserving hashing,","cited_arxiv_id":null,"evidence_quote":"Probability-based semantics-preserving hashing baseline that achieves strong text-to-image results on Wiki and is included in the comparison."}],"review_version":1}