{"id":"31ba489d-4a38-4df1-91fd-17021944d76b","arxiv_id":"1908.03464","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Semantic attributes as supervision for feature selection enable models trained on seen classes to select discriminative features for unseen classes, outperforming label-based and unsupervised baselines on SUN, aPY and CIFAR10.","lead":"The authors introduce zero-shot feature selection, where a model trained on labeled classes picks features that also work for entirely new classes, using attribute descriptions rather than labels as supervision. It is worth a look as a first demonstration that semantic attributes can power feature selection that generalizes to unseen concepts.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 5's relaxation drops the cardinality constraint, making the objective unbounded under scaling; reported feature rankings may depend on stopping time rather than the model.","rationale":"The reader's weakest assumption concerned transferability of the linear attribute mapping and noisy attributes. That is a legitimate risk, but the more immediate and verifiable flaw is that Eq. 5 as written has no minimizer. The scaling argument is elementary and does not depend on data or hyperparameters: multiplying s by t and dividing W by t preserves the reconstruction terms while reducing the regularization term, so the objective can be decreased indefinitely. The alternating algorithm therefore cannot converge to a minimizer; any reported s is determined by the unspecified iteration count, step size, and initialization. The empirical comparisons in Figs. 4–5 do not yet support the central claim under a well-posed optimization. The paper does receive credit for a clear problem formulation, a sensible attribute-based idea, and a center-characteristic loss; the flaw is in the relaxation and the convergence argument. Since the fix is straightforward, CONDITIONAL rather than REJECT is appropriate. The proposed test, re-running with the cardinality constraint restored, would settle whether the method's advantage survives a well-posed optimization.","tokens_in":11037,"tokens_out":8877,"duration_ms":102897,"concrete_test":"Analytically verify the scaling direction: for any iterate (s,W) produced by Alg. 1 on SUN with W≠0, define s_t=t s and W_t=W/t for t>1. Since X diag(s_t)W_t = X diag(s)W and \\bar X diag(s_t)W_t = \\bar X diag(s)W, the first two terms of Eq. 5 are unchanged while the γ||W||² term shrinks by t², so J(s_t,W_t) < J(s,W). This proves Eq. 5 has no finite minimizer. The decisive experimental follow-up is to re-run SemFS with s^T1_d=k enforced (or with an ℓ2 bound on s) and check whether the top-20 features and the Figs. 4–5 ACC/NMI curves change materially; if they do, the published rankings are artifacts of the unconstrained relaxation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing problem is not attribute transfer but the optimization itself. In Eq. 4 the search is constrained by s∈{0,1}^d and s^T1_d=k. When the paper relaxes to Eq. 5, it keeps only s≥0 and silently drops s^T1_d=k. Let U = diag(s)W. Then Eq. 5 is ||Ys−XU||²_F + α||Ys−\\bar X U||²_F + γ||diag(s)^{-1}U||²_F. For any fixed U, taking s_i→∞ and W_{i,:}=U_{i,:}/s_i leaves the first two terms unchanged while the γ term tends to 0. Consequently, the objective has no finite minimizer: from any feasible (s,W) with W≠0, scaling s→t s, W→W/t strictly decreases the objective for t>1. Proposition 1 only asserts monotone decrease of the objective value; it does not show convergence of s, and in the stated problem s can diverge while W→0. Thus the ranking of s is an artifact of the unspecified initialization, step size, and iteration count, not a well-defined solution of Eq. 5. Since the paper's headline results are rankings from this procedure, the central empirical claim is not reproducible as stated. This is an internal inconsistency, not a disagreement with consensus; it can be repaired by restoring s^T1_d=k (or an equivalent normalization/sparsity constraint), but the experiments must then be rerun.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the zero-shot feature selection (ZSFS) problem, in which a feature selection model is trained on seen classes and then applied to select features for unseen classes that share no labels with the training classes. The proposed method, SemFS, uses class-attribute descriptions (manually annotated attributes or automatically generated word vectors) as supervision instead of class labels, and adds a center-characteristic loss that encourages the selected features to preserve the central characteristics of seen classes in the attribute space. The binary cardinality-constrained selection problem is relaxed to a non-negative score vector, and optimization alternates between a closed-form update for a linear mapping W and projected gradient descent on the score vector s. Experiments on SUN, aPY, and CIFAR10 evaluate the selected features by clustering unseen-class data and report ACC and NMI improvements over several unsupervised and supervised feature selection baselines.","tokens_in":11406,"tokens_out":7565,"duration_ms":75506,"significance":"Zero-shot feature selection is a relevant and little-studied problem, and the idea of using attribute embeddings as a transferable supervision signal is natural and potentially impactful. The center-characteristic loss is a sensible adaptation of center loss to feature selection, and the experiments cover both manually annotated and automatically generated attributes. The paper does not provide code, but the pseudocode is clear and the method is simple. However, the technical soundness of the optimization is compromised by a relaxation that yields an objective without a finite minimizer, and the experimental evidence lacks statistical rigor and a full-feature baseline. These issues are load-bearing for the paper's central claims, so the manuscript requires substantial revision before it can be considered for publication.","major_comments":[{"comment":"The relaxation from Eq. (4) to Eq. (5) drops the cardinality constraint s^T 1_d = k as well as the 0/1 constraint, leaving only s ≥ 0. As written, the objective in Eq. (5) has no finite minimizer. Writing U = diag(s)W, the first two terms become ||Y_s - XU||_F^2 + α||Y_s - \\bar{X}U||_F^2, which are independent of s, while the regularization term becomes γ||diag(s)^{-1}U||_F^2. For any feasible (s,W) with positive s_i and U nonzero, the scaling (s,W) → (t s, W/t) strictly decreases the objective for t > 1, so the infimum is approached only as t → ∞. Consequently, the 'solution' s is not well-defined, and the ranking produced by Algorithm 1 is an artifact of the initialization, step size, and stopping criterion. Proposition 1 only shows monotone decrease of the objective value; it does not establish convergence of s, and the cited projected-gradient convergence result does not apply because the minimum does not exist. This undermines the central claim that SemFS selects features by optimizing a well-defined criterion. The authors should restore a normalization or sparsity constraint (e.g., s^T 1_d = k or an equivalent penalty that prevents s from diverging), re-derive the updates, and rerun all experiments.","section":"Section 4.3, Eq. (5)"},{"comment":"The text states that SemFS and SemFS/c 'always outperform the others significantly,' but no error bars, confidence intervals, or significance tests are reported. The clustering metrics are averaged over 20 K-means initializations, yet the variance across those runs is not shown, and for CIFAR10 the variance due to the choice of seen/unseen split is also unaccounted for. Without measures of variability or statistical tests, the claim of 'significantly' better performance is unsupported. Please report standard deviations and, where appropriate, paired significance tests across repeated clustering runs and data splits.","section":"Section 6.2, Figures 4 and 5"},{"comment":"The description of the CIFAR10 experiment is ambiguous. The text says 'we randomly adopt two classes as seen' and then notes there are C(10,2) different seen/unseen splits, but Figures 4(c) and 5(c) appear to show a single curve. It is unclear whether the reported result is from one randomly chosen split or averaged over all 45 splits. Since CIFAR10 is the key demonstration for the extreme case with automatically generated attributes, this ambiguity materially affects the claim. Please clarify the protocol and, ideally, report the average and standard deviation over all 45 splits.","section":"Section 6.1, Figures 4(c) and 5(c)"},{"comment":"No baseline using all 4,096 features is reported. As the paper focuses on feature selection, it is essential to compare the selected features against the full feature set to show that the benefit of SemFS comes from attribute-guided selection rather than from dimensionality reduction alone. A simple K-means clustering result using all 4,096 features for each dataset should be added to Figures 4 and 5, or else the advantage of the method over the full set must be discussed explicitly.","section":"Section 6.2"}],"minor_comments":[{"comment":"In Section 6.5 the text says 'with the fixed regularization parameter λ=1', but Section 6.1.2 reports γ=0.1; please use consistent notation for the regularization parameter.","section":"Section 6.5"},{"comment":"The proof of Proposition 1 relies on an 'appropriate choice of step size' for the projected gradient update, but the algorithm does not specify how the step size is chosen or whether a line search is used. This is a reproducibility issue even apart from the non-existence problem in Eq. (5).","section":"Section 5.1, Algorithm 1"},{"comment":"The claim 'to our best knowledge, this is the first study to consider the zero-shot setting in the feature selection problem' is strong and difficult to verify from the provided references; please soften the claim or support it with a more thorough literature review.","section":"Section 2.2"},{"comment":"The axis labels 'Lassoo' and 'Lassos' contain typos and should be corrected to 'Lasso_o' and 'Lasso_s'.","section":"Figure 7"},{"comment":"The transition from Eq. (2) to Eq. (3) replaces Xdiag(s) with Y_s without an explicit derivation; the heuristic of substituting the attribute reconstruction from Eq. (1) should be described more clearly to avoid confusion.","section":"Equations (2) and (3)"},{"comment":"The definition of the class center \\bar{x}_i is ambiguous because y_i is a label vector, not a class index; please clarify, for example by defining a class index c_i and writing \\bar{x}_{c_i}.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an arXiv preprint (v3, 2021) and has not been published elsewhere to my knowledge. The main technical issue is the unbounded relaxation in Eq. (5), which is fixable but requires revisiting the optimization and all reported experiments. In addition, the experimental section needs more rigorous reporting (variance, split protocol, full-feature baseline). Given the significance of the ZSFS problem and the plausibility of the high-level idea, I recommend major revision rather than rejection, provided the authors can resolve the optimization flaw and strengthen the evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing you should know: this paper defines a genuinely new problem, zero-shot feature selection, and the high-level idea — use class attributes instead of labels to supervise feature selection so the chosen features transfer to unseen classes — is sensible and worth taking seriously. The framing is not circular: training uses only seen-class attributes, and evaluation clusters unseen data. I found no sign that the authors tuned α and γ to test performance; the ablation with α=0 is the right way to test the center-characteristic loss.\n\nNow the soft spots, in proportion. The load-bearing one is the optimization, and the stress-test note is correct. Eq. 4 has s ∈ {0,1}^d and sᵀ1 = k. The relaxation to Eq. 5 keeps only s ≥ 0 and silently drops the cardinality constraint. As written, the objective is unbounded below: if you scale s → t·s and W → W/t with t > 1, the first two residual terms are unchanged while γ‖W‖²_F shrinks to 0. So there is no finite minimizer, and Proposition 1 only shows monotone decrease of the objective value, not convergence of s. A monotone decreasing sequence can diverge in s. The reported rankings may therefore depend on initialization, step size, and iteration count rather than on a well-defined solution. This is fixable — restore sᵀ1 = k or add an equivalent normalization or sparsity penalty — but the experiments have to be rerun under the corrected objective. This is not a cosmetic issue; it is the core of the method.\n\nOther concerns are more minor but real. The CIFAR10 split is ambiguous: the text says “randomly adopt two classes as seen,” then mentions C₂₁₀ different splits. Is Fig. 4(c) one random split or an average over all 45? That needs one sentence of clarification. There are also no error bars or significance tests anywhere, despite “significantly” appearing several times; K-means was repeated 20 times, so standard deviations were available. And there is no baseline that uses all 4,096 features, so we cannot tell whether the selected features beat simply keeping everything.\n\nThe citation pattern looks fine, and the paper honestly acknowledges that attributes can be noisy. The problem definition and experimental framework are worth preserving. But the optimization flaw means the central empirical claim is not reproducible as stated.\n\nMy bottom line: this deserves peer review, not desk rejection, because the zero-shot feature selection problem is new and the intended mechanism is plausible. A serious referee should demand the optimization fix, rerun experiments, add error bars, and clarify the CIFAR10 protocol. I would not cite it in its current form, but I would bring it to a reading group to discuss the problem framing and the unbounded-relaxation pitfall.","headline":"Genuinely new problem and a plausible attribute-transfer idea, but the relaxed objective in Eq. 5 is unbounded below, so the reported feature rankings may be artifacts of stopping time until the constraint is restored.","tokens_in":11904,"tokens_out":2620,"would_cite":false,"duration_ms":28598,"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":"Guiding feature selection with class attributes instead of labels lets a model trained on seen concepts pick features that also separate unseen concepts.","keywords":["feature selection","zero-shot learning","class attributes","semantic supervision","center loss","transfer learning","clustering evaluation"],"falsifier":"On a dataset with a known seen/unseen split, randomly permute the attribute values of the seen classes while keeping everything else fixed; if SemFS's unseen-class clustering advantage over label-based baselines does not drop substantially, then the attribute content is not what drives the result. More directly, construct a dataset whose true attributes are nonlinear functions of the features and show that a nonlinear attribute-reconstruction selector outperforms SemFS on unseen classes, which would indicate the linear transfer assumption is the limiting factor.","tokens_in":10884,"feed_emoji":"🎯","tokens_out":6706,"duration_ms":64420,"temperature":0.7,"pith_summary":"The paper claims that feature selection can be made zero-shot: a model trained on 'seen' classes can rank features that remain discriminative for 'unseen' classes, provided the supervision signal is class attributes rather than class labels. The problem matters because new concepts appear faster than reliable labels can be produced, and traditional selectors trained on seen labels carry no information about unseen classes. To test this, the authors build SemFS, which learns to reconstruct attribute vectors from selected deep features, and add a center-characteristic loss that anchors selected features to their class centers. Evaluations on SUN, aPY, and CIFAR10 show that the resulting feature rankings cluster unseen classes better than rankings from label-based or unsupervised baselines, even when only 10% of seen concepts are labeled and when attributes are automatically generated word vectors.","feed_headline":"Class attributes let feature selection reach unseen concepts","feed_subtitle":"On SUN, aPY, and CIFAR10, attribute-guided rankings cluster unseen classes better than label-based ones.","key_machinery":"The machinery is the joint objective over the feature importance vector $s$ and the linear generating matrix $W$: $$\\|Y_s - X\\operatorname{diag}(s)W\\|$_F^{2}$ + \\$\\alpha$\\|Y_s - \\bar{X}\\operatorname{diag}(s)W\\|$_F^{2}$ + \\gamma\\|W\\|$_F^{2}$,$$ where $Y_s$ holds class attributes, $X$ is the data, and $\\bar{X}$ holds class centers. The first term asks selected features to linearly reconstruct attributes, and the second term (the center-characteristic loss) asks them to reconstruct class centers under the same map, which steers selection toward stable central characteristics rather than idiosyncratic or noisy attribute correspondences. Relaxing the 0/1 selection constraint to $s\\ge 0$ turns $s$ into a feature importance ranking, and alternating updates converge to a solution.","core_discovery":"The central claim is that the reason supervised feature selection fails on unseen concepts is the supervision itself: 0/1 class labels encode only distinctions among seen classes, so nothing transfers. Replacing labels with class-semantic attributes gives the selector a shared semantic layer, and a linear map from selected features to attributes learned on seen classes keeps working for unseen classes. The paper also claims that simply regressing on attributes is not enough because attributes are noisy, so it adds the center-characteristic loss, which requires the same linear map to reconstruct each seen class's center from its selected features. The joint model, SemFS, produces a non-negative feature importance vector through alternating least squares and projected gradient descent, and ranking by that vector yields the selected features. This, the authors argue, is the first formulation of the zero-shot setting for feature selection.","pith_inferences":["The same swap of labels for attributes could be applied to other supervised feature selection objectives, so SemFS is likely an instance of a general recipe rather than a single algorithm.","If attribute noise is the main risk, a natural extension is to model attribute uncertainty explicitly, for example by weighting attributes by reliability or adding a per-attribute variance term.","The paper evaluates transfer only through clustering; a direct test on zero-shot classification with the selected features would be a stronger end-to-end check of the claim.","On CIFAR10 only two seen classes support the transfer, so the extreme low-resource regime is where the assumption of a linear attribute map is most strained and where nonlinear mappings would be worth testing."],"forward_implications":["A feature selection model trained once on seen concepts can be applied directly to unseen concepts, as long as attribute descriptions for those concepts are available.","Unlabeled unseen data can be clustered on the selected features, since the paper's evaluation shows K-means on SemFS-selected features yielding higher ACC and NMI than baselines.","Supervised selectors that use class labels may be actively misleading for unseen concepts; attribute-based supervision avoids this label bias.","Automatically generated text-derived attributes, not only manually curated ones, are sufficient to transfer feature selection knowledge."],"supporting_citations":[{"why":"Supplies the attribute-based zero-shot learning mechanism the method builds on.","marker":"[13]"},{"why":"Introduces center loss, the basis of the center-characteristic loss for learning discriminative features.","marker":"[34]"},{"why":"Provides the word-vector representations used as automatically generated attributes for CIFAR10.","marker":"[16]"},{"why":"Documents that attributes are noisy, motivating the center-characteristic loss.","marker":"[17]"},{"why":"Supplies the 4096-dimensional deep features used for all datasets.","marker":"[14]"},{"why":"Provides the SUN scene attribute benchmark and its seen/unseen split.","marker":"[27]"},{"why":"Unsupervised feature selection baseline (MCFS) that the method must beat.","marker":"[4]"},{"why":"Classical supervised baseline, used to show attributes outperform labels.","marker":"[33]"}],"fun_headline_variants":["Semantic attributes unlock zero-shot feature selection","From seen to unseen: attribute-guided feature selection","Zero-shot feature selection with class attributes","Transfer knowledge to unseen classes for feature selection","Attributes make feature selection work for unseen classes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a linear map from selected features to attributes learned on seen classes remains valid for unseen classes, and that the attributes themselves carry discriminative information rather than noise.","fun_headline_variants_meta":{"raw":{"variants":["Semantic attributes unlock zero-shot feature selection","From seen to unseen: attribute-guided feature selection","Zero-shot feature selection with class attributes","Transfer knowledge to unseen classes for feature selection","Attributes make feature selection work for unseen classes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000583,"raw_usage":{"total_tokens":2696,"prompt_tokens":853,"completion_tokens":1843,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":1777}},"tokens_in":469,"tokens_out":1843,"duration_ms":12691,"temperature":1.0,"reasoning_tokens":1777,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:12:10.253889+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a dataset with a known seen/unseen split, randomly permute the attribute values of the seen classes while keeping everything else fixed; if SemFS's unseen-class clustering advantage over label-based baselines does not drop substantially, then the attribute content is not what drives the result. More directly, construct a dataset whose true attributes are nonlinear functions of the features and show that a nonlinear attribute-reconstruction selector outperforms SemFS on unseen classes, which would indicate the linear transfer assumption is the limiting factor.","supporting_citations":[{"cited_title":"Describing objects by their attributes","cited_arxiv_id":null,"evidence_quote":"Supplies the attribute-based zero-shot learning mechanism the method builds on."},{"cited_title":"A discriminative feature learning approach for deep face recognition","cited_arxiv_id":null,"evidence_quote":"Introduces center loss, the basis of the center-characteristic loss for learning discriminative features."},{"cited_title":"H., Socher, R., Manning, C","cited_arxiv_id":null,"evidence_quote":"Provides the word-vector representations used as automatically generated attributes for CIFAR10."},{"cited_title":"Zero-shot recognition with unreliable attributes","cited_arxiv_id":null,"evidence_quote":"Documents that attributes are noisy, motivating the center-characteristic loss."},{"cited_title":"Semi-supervised active learn- ing with cross-class sample transfer","cited_arxiv_id":null,"evidence_quote":"Supplies the 4096-dimensional deep features used for all datasets."},{"cited_title":"Sun attribute database: Discovering, an- notating, and recognizing scene attributes","cited_arxiv_id":null,"evidence_quote":"Provides the SUN scene attribute benchmark and its seen/unseen split."},{"cited_title":"Unsupervised feature selection for multi- cluster data","cited_arxiv_id":null,"evidence_quote":"Unsupervised feature selection baseline (MCFS) that the method must beat."},{"cited_title":"Regression shrinkage and selection via the lasso","cited_arxiv_id":null,"evidence_quote":"Classical supervised baseline, used to show attributes outperform labels."}],"review_version":1}