{"id":"c2c26d9e-c8c0-4d73-bce1-f094cb519337","arxiv_id":"2412.14518","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"S5VH uses bidirectional Mamba layers and a hash-center alignment loss to improve self-supervised video hashing accuracy and efficiency.","lead":"This paper builds a self-supervised video hashing model on Mamba, a state-space architecture with linear-time sequence processing, and adds a global hash-center alignment loss. The method reports better retrieval accuracy and faster, more memory-efficient inference than transformer baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hash center ADMM update in Eq. (16) appears to be the sphere projection, not the box projection, so the generated centers are not binary as claimed; this undermines the LCA contribution and method reproducibility.","rationale":"The reader identified the semantic meaningfulness of k-means clusters as the weakest assumption. That concern is real, but the paper's ablation (Table 2, w/o LCA vs. full) provides empirical evidence that LCA helps on these benchmarks, partially addressing the concern. In contrast, the ADMM update error is an internal inconsistency: Eq. (16) does not solve the subproblem it claims to solve. If the released code follows the text, the method produces non-binary centers, breaking the mathematical derivation of Eq. (21) and the central claim of generating binary hash centers. The empirical results might still hold if the authors implemented a correct proximal step, but then the paper's description is flawed and the method is not reproducible from the text alone. This is more load-bearing because it directly targets the correctness of the proposed contribution, not just its interpretation. The concrete test (checking the proximal derivation and the binary nature of Φ from the released code) would settle whether the concern lands. Given the paper's strengths—clear benchmark improvements, efficiency analysis, and ablations—a CONDITIONAL verdict remains appropriate: require the authors to correct the ADMM description and verify (or fix) the implementation before full acceptance.","tokens_in":26920,"tokens_out":7973,"duration_ms":72930,"concrete_test":"Re-derive the proximal operator for the box constraint in Eq. (15): it must be clip(Φ^{k+1} + Υ_b^k/μ_b, -1, 1), not Eq. (16). Then run the released code's hash center generation on UCF101 features (or any dataset) and inspect the converged Φ. Count the fraction of entries that are exactly ±1 (within a small tolerance) and verify that the output is indeed binary. If Eq. (16) is used or the output contains non-binary entries, the described algorithm is not what is implemented, and the LCA loss's mathematical grounding changes. Alternatively, ablate by replacing the generated centers with random binary centers; if the LCA gain persists, the semantic structure is irrelevant, which would also undermine the claimed mechanism.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The method's central novelty is the semantic hash center generation algorithm, which claims to optimize binary centers Φ ∈ {−1,+1}^{Nc×K} under semantic consistency and separation objectives. In the ℓp-box ADMM derivation, Eq. (15) defines the proximal subproblem for the box constraint Sb ≡ [−1,+1]^{Nc×K}. The closed-form solution should be the Euclidean projection onto the box, i.e., clipping each entry to [−1,1]. Instead, Eq. (16) gives Ψ_b^{k+1} = sqrt(NcK) * (Φ^{k+1} + Υ_b^k/μ_b) / ||Φ^{k+1} + Υ_b^k/μ_b||_F, which is the projection onto the Frobenius sphere of radius sqrt(NcK) — the projection for the ℓ2-norm constraint Sp, not the box. If implemented as written, the ADMM updates never enforce the binary/box constraint; the resulting 'hash centers' would be continuous vectors on a sphere, possibly with entries far outside [−1,1]. The subsequent center alignment loss in Eq. (21) simplifies the cosine similarity to φ_c^T b_i / (Kτ) under the assumption that φ_c and b_i are ±1; with non-binary φ_c this simplification is invalid, and the claimed 'well-separated and semantically consistent hash centers' are not produced. Because this algorithm is a key contribution and the paper does not report the actual properties of Φ (e.g., fraction of ±1 entries), the method description is internally inconsistent. This is a concrete correctness risk that cannot be waived by the empirical ablations, which compare black-box variants without verifying the optimization's output.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents S5VH, a self-supervised video hashing model built from bidirectional Mamba layers, with a self-local-global (SLG) learning paradigm that combines masked temporal reconstruction, contrastive learning, and a hash-center alignment loss. The hash centers are derived from k-means clustering of temporally averaged CNN features, followed by an ADMM optimization that aims to map cluster centroids to well-separated binary codes. Experiments on ActivityNet, FCVID, UCF101, and HMDB51 report improved retrieval accuracy over six baselines, cross-dataset transferability, and lower inference cost with linear scaling in sequence length. The paper includes ablations of the bidirectional design, loss terms, and SSM variants, and the code is publicly available.","tokens_in":27280,"tokens_out":6320,"duration_ms":53663,"significance":"If the empirical claims hold, S5VH is a meaningful step for efficient self-supervised video hashing, being the first Mamba-based model in this area. The linear-complexity inference and the global hash-center alignment signal are both practically motivated, and the paper provides a useful set of ablations. The public code and the ablation coverage are strengths. However, the central learning signal relies on an ADMM derivation that is currently incorrect as written, and the experimental section lacks measures of variance, so the significance can only be confirmed after a carefully revised version.","major_comments":[{"comment":"The update for the box-constraint auxiliary variable Ψ_b is written as the projection onto the Frobenius sphere of radius sqrt(NcK): Ψ_b^{k+1} = sqrt(NcK) * (Φ^{k+1} + Υ_b^k/μ_b) / ||Φ^{k+1} + Υ_b^k/μ_b||_F. This is the projection for the ℓ2-norm constraint S_p, not for the box constraint S_b ≡ [-1,+1]^{Nc×K}. The correct proximal update for S_b is elementwise clipping of Φ^{k+1} + Υ_b^k/μ_b to [-1,1]. As written, the ADMM never enforces the box constraint, so the output of the hash center generation is not guaranteed to be binary. Consequently, the simplification of the cosine similarity in Eq. (21) to φ_c^T b_i / (Kτ), which relies on both φ_c and b_i being ±1 vectors, is invalid unless the centers are actually binarized by some other mechanism. Please correct the update (assigning the sphere projection to the Ψ_p update), verify the actual outputs of the optimization, and report the fraction of ±1 entries or the constraint violation. The method is not reproducible as currently described.","section":"Semantic Hash Center Generation, Eq. (16)"},{"comment":"All retrieval numbers are reported as single runs without standard deviations or statistical significance tests. The cross-dataset gain in Table 1 (e.g., 0.0285 vs 0.0378 on 16-bit codes) and the LCA ablation in Table 2 (e.g., 0.357 vs 0.286 at 16 bits on UCF101) could be affected by run-to-run variation. Since the central claim is that S5VH outperforms state-of-the-art methods, please provide mean ± std over at least three independent runs, or justify why a single run is representative. This is important for evaluating whether the reported improvements are reliable.","section":"Tables 1 and 2; Figure 3"}],"minor_comments":[{"comment":"The metric is called 'geomean' (GmAP) but Eq. (25) computes the root-sum-square (Euclidean norm) of the mAP@N values, not the geometric mean. The geometric mean over the six N values would be the sixth root of the product of the mAP@N values. Please rename the metric or correct the formula, and adjust the captions and text of Figure 1 and Table 2 accordingly.","section":"Eq. (25)"},{"comment":"The values of the loss weights α and β in Eq. (22) are not reported, although they are essential for reproducibility. Please list them, together with the early-stopping criteria and the number of runs.","section":"Implementation Details"},{"comment":"There are several presentation errors: the inference efficiency results are referred to as Figure 1(b) in the Experiments section, but Figure 1 shows inference efficiency in (a); the Model Analyses section refers to 'Figure 1(c)', which does not exist; the footnote defines the abbreviation as 'SSSSSVH' (an extra S); and 'stimulate tensors' should be 'simulate tensors'.","section":"References and figure citations"},{"comment":"The contrastive loss in Eq. (20) is typeset in a way that is difficult to parse. Please rewrite it as the sum of two log-softmax terms, one treating the first view as anchor and the second treating the second view as anchor, to make the objective unambiguous.","section":"Eq. (20)"},{"comment":"The number of semantic centers Nc and the mask ratio ρ are fixed per dataset, but no sensitivity analysis or justification is provided. A brief discussion of how these choices affect the hash center quality and retrieval performance would strengthen the paper.","section":"Hyperparameter sensitivity"}],"recommendation":"major_revision","confidential_remarks":"The ADMM issue in Eq. (16) is the principal technical concern; it appears to be a fixable derivation error rather than a fundamental flaw, and the manuscript's scope can accommodate the correction. The missing variance reporting and hyperparameter details are also important but are standard revision items. The author group has several related publications, but the citation pattern is not excessive for this subfield."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a useful paper with a real novelty—first Mamba-based self-supervised video hashing—and experiments that mostly support the claims. But there is a concrete derivation error in the hash-center generation (Eq. 16 projects onto the sphere, not the box), so the \"binary\" centers are not guaranteed. That needs fixing before I would trust the LCA contribution.\n\nWhat's new: applying bidirectional Mamba to SSVH is a sensible extension of Vim/VMamba, and the hash-center alignment loss is a reasonable global signal. The paper runs standard baselines on four datasets, includes ablations for the main components, and reports an efficiency scaling analysis. The code link is a plus. The cross-dataset transfer result and the convergence speedup from LCA are the strongest empirical points.\n\nSoft spots:\n1. Eq. (16) is wrong for the box constraint Sb. The update is the projection onto the Frobenius sphere of radius sqrt(NcK), which is the update for the ℓ2 sphere constraint, not the box. If implemented as written, the centers are not binary. The loss simplification in Eq. (21) assumes φ ∈ {±1}, so the derivation is internally inconsistent. This is a major flaw in a claimed contribution, not a cosmetic typo. A corrected derivation and a check on the actual fraction of ±1 entries in Φ are required.\n2. Eq. (25) labels a root-sum-square as 'GmAP'; the geometric mean would be the 6th root of the product. Minor but confusing.\n3. No error bars or repeated runs anywhere. For a paper whose conclusions are comparative, that's a standard requirement.\n\nThe k-means pseudo-label concern the reader raised is real but secondary; even if the clusters are imperfect, the loss still acts as a soft prototype alignment, which can help. The binary-center issue is more damaging.\n\nThe stress-test note is correct and lands. The reader's confidence was a bit generous; the paper needs revision.\n\nWho this is for: people working on self-supervised video hashing or efficient video backbones. It deserves a serious referee: the topic is timely, the experiments are reasonably thorough, and the flaw is fixable. I would not desk-reject. But I'd make the authors (a) correct the ADMM update, (b) report the binary rate of the generated centers, (c) add error bars, and (d) rename the metric.\n\nRecommendation: send to peer review with major revision.","headline":"First Mamba-based SSVH with solid experiments, but the hash-center ADMM update in Eq. (16) is the sphere projection, not the box projection, so the binary-center claim is not justified; needs a fix before the LCA contribution can be trusted.","tokens_in":27814,"tokens_out":2498,"would_cite":false,"duration_ms":20744,"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":"S5VH shows Mamba backbones can beat Transformers in self-supervised video hashing.","keywords":["Mamba","state space models","self-supervised video hashing","hash center alignment","video retrieval","Hamming space","bidirectional Mamba","contrastive learning"],"falsifier":"Run S5VH on a video dataset whose category structure is not captured by the pretrained CNN features, and compare it with and without the center alignment loss; if the GmAP gap vanishes or reverses, the semantic meaningfulness of the k-means pseudo-labels is what carries the improvement. A cleaner control is to replace the real k-means centroids with random binary centers: if the retrieval gain survives, the global signal is doing the work, not the semantics.","tokens_in":1877,"feed_emoji":"🎬","tokens_out":3527,"duration_ms":69432,"temperature":0.7,"pith_summary":"The paper sets out to show that Mamba, a selective state-space model with linear-time sequence processing, can serve as the backbone for self-supervised video hashing, replacing the quadratic-cost Transformers that currently dominate the task. It introduces S5VH, whose encoder and decoder are built from bidirectional Mamba layers, and pairs it with a self-local-global learning scheme: masked-frame reconstruction gives a self signal, contrastive learning gives an inter-sample local signal, and a hash center alignment loss adds a corpus-level global signal. The global signal is obtained by clustering temporally averaged CNN features with k-means and then solving a constrained optimization that converts the cluster centroids into semantically consistent, well-separated binary hash centers. If the paper is right, state-space models offer both better retrieval accuracy and a scalable inference regime for video hashing, and hash center alignment is a reusable way to inject global semantics into self-supervised hash learning.","feed_headline":"Mamba backbone beats Transformers in self-supervised video hashing","feed_subtitle":"Bidirectional state-space layers cut inference cost while a hash-center loss speeds training by about 65 percent.","key_machinery":"The machinery has two parts. First, the bidirectional Mamba layer: each layer sums a forward and a backward Mamba block, where each block is a gated selective scan (linear projection, layer norm, 1D convolution, SiLU, and a selective state-space module with input-dependent $\\Delta$, $B$, and $C$); this carries temporal modeling at linear complexity in frame number. Second, semantic hash center generation: k-means on temporally averaged video features yields $N_c$ centroids, and an $\\ell^p$-box ADMM solver minimizes $\\|\\Phi\\Phi^\\top - K W\\|_F^2 + \\tfrac{1}{2}\\sum_{i,j}\\phi_i^\\top\\phi_j$ over binary $\\Phi$, where $W$ is the cosine similarity matrix of centroids; the resulting hash centers are both consistent with feature-space semantics and mutually separated. The center alignment loss $\\mathcal{L}_{CA} = -\\log \\frac{\\exp(\\phi_{c_i}^\\top b_i/(K\\tau))}{\\sum_{c}\\exp(\\phi_c^\\top b_i/(K\\tau))}$ then pulls each video's hash code toward its assigned center during training.","core_discovery":"The central claim is that S5VH is the first Mamba-based self-supervised video hashing model and that it outperforms the state of the art on ActivityNet, FCVID, UCF101, and HMDB51 at 16-, 32-, and 64-bit code lengths, transfers better across datasets, and scales better at inference. The reported gains include a GmAP of 0.0378 in cross-dataset transfer from UCF101 to HMDB51, versus 0.0335 for the strongest baseline, and a fitted inference-time scaling law that is linear in frame count where the Transformer and MLP-Mixer baselines are quadratic. The paper also claims that the proposed center alignment loss, which aligns each video's hash code to the binary hash center of its k-means pseudo-class, contributes roughly six percent GmAP improvement and about 65 percent training-time acceleration, making the self-local-global paradigm the reason S5VH converges faster and better.","pith_inferences":["Hash center alignment is a plug-in training signal: it could be dropped into non-Mamba self-supervised hashers whose codes are compared by cosine or Hamming similarity, and the ablation pattern suggests it would help any such model.","The k-means pseudo-labels and ADMM hash centers are fixed before training, so a natural variant the paper does not test is re-estimating them periodically as features drift during training.","The linear-complexity inference result suggests state-space backbones are worth trying in other long-sequence retrieval tasks, such as near-duplicate video search or dense video captioning retrieval.","The gain from center alignment likely depends on how well the cluster count $N_c$ matches the true semantic structure; the paper fixes $N_c$ by dataset scale, so sensitivity to $N_c$ is a testable next step."],"forward_implications":["S5VH outperforms six baseline methods in mAP@N and precision-recall across ActivityNet, FCVID, UCF101, and HMDB51 at 16-, 32-, and 64-bit code lengths.","In cross-dataset transfer, training on UCF101 and testing on HMDB51 yields a GmAP of 0.0378 versus 0.0335 for the strongest baseline.","Inference time scales linearly with video length, with a fitted cost of roughly $1.5\\times10^{-3}L$ milliseconds, whereas the Transformer and MLP-Mixer baselines fit quadratic scaling laws.","The center alignment loss contributes about six percent GmAP and roughly 65 percent faster convergence, so the global signal is a substantial efficiency lever during training.","Ablations show bidirectional Mamba beats forward-only and backward-only variants, and beats LSTM, RetNet, and RWKV when substituted into the same framework."],"supporting_citations":[{"why":"Supplies the Mamba selective state-space layer that S5VH converts into bidirectional temporal encoder and decoder blocks.","marker":"Gu and Dao 2023"},{"why":"Provides the contrastive masked autoencoder baseline ConMH and the masked-frame reconstruction loss that S5VH extends.","marker":"Wang et al. 2023a"},{"why":"Provides the $\\ell^p$-box ADMM optimization scheme that S5VH uses to solve the binary hash center generation problem.","marker":"Wu and Ghanem 2018"},{"why":"Establishes the bidirectional Transformer self-supervised video hashing baseline BTH that S5VH compares against.","marker":"Li et al. 2021"},{"why":"Supplies the MCMSH baseline and the multi-granularity contextualization approach used in the efficiency comparison.","marker":"Hao et al. 2022"},{"why":"Defines the original self-supervised video hashing formulation and the dataset split protocols adopted in the experiments.","marker":"Song et al. 2018a"},{"why":"Supplies the ResNet-50 backbone used to extract frame features for the ActivityNet experiments.","marker":"He et al. 2016"}],"fun_headline_variants":["Mamba beats Transformers in video hashing, 65% faster","State-space model speeds video hashing, boosts accuracy","S5VH: Mamba-based hashing with linear complexity wins","Video hashing leap: Mamba outperforms, cuts compute","Faster, better video hashing with Mamba's selective scan"],"cache_read_input_tokens":29824,"weakest_assumption_plain":"The load-bearing premise is that k-means clusters of temporally averaged, pretrained CNN features correspond to semantic categories; if the clusters are noisy or semantically meaningless, the center alignment loss will pull hash codes toward arbitrary centroids and the reported gains from LCA would disappear.","fun_headline_variants_meta":{"raw":{"variants":["Mamba beats Transformers in video hashing, 65% faster","State-space model speeds video hashing, boosts accuracy","S5VH: Mamba-based hashing with linear complexity wins","Video hashing leap: Mamba outperforms, cuts compute","Faster, better video hashing with Mamba's selective scan"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000264,"raw_usage":{"total_tokens":1608,"prompt_tokens":951,"completion_tokens":657,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":568}},"tokens_in":567,"tokens_out":657,"duration_ms":5498,"temperature":1.0,"reasoning_tokens":568,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:09:44.030218+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run S5VH on a video dataset whose category structure is not captured by the pretrained CNN features, and compare it with and without the center alignment loss; if the GmAP gap vanishes or reverses, the semantic meaningfulness of the k-means pseudo-labels is what carries the improvement. A cleaner control is to replace the real k-means centroids with random binary centers: if the retrieval gain survives, the global signal is doing the work, not the semantics.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MCMSH baseline and the multi-granularity contextualization approach used in the efficiency comparison."}],"review_version":1}