{"id":"a075b15f-2bf5-4387-b9cd-3616dbde8e16","arxiv_id":"1908.03983","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A visual and semantic prototype-guided CNN decomposes generalized zero-shot learning into open set recognition and zero-shot classification, and introduces generalized open set recognition.","lead":"This paper splits generalized zero-shot learning into two stages: reject images that do not belong to known classes, then identify the rejected images using attribute descriptions. It also introduces a new task where unknown images receive attribute-based explanations instead of just an 'unknown' label.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The entropy gate in §III.C.2 is the load-bearing component of the CO-solving claim; it has no theoretical or validation support, so 'effectively solves' is not established.","rationale":"The reader's weakest assumption is correct: the entropy threshold is load-bearing. My stress test confirms this and adds an internal validation issue: Section IV.A.2's G-ZSL-val set contains only seen classes (1/7 of seen training plus seen validation), so δg and δo cannot be calibrated to reject unseen classes. This is not a disagreement with consensus; it is a gap in the argument's support. The empirical results (H=57.0 on CUB, 67.0 on AWA2, 30.9 on SUN, 34.0 on aPY) are competitive but mixed; on SUN the method is below several baselines, which contradicts the strength of the 'effectively solves' claim. The method has positive aspects: end-to-end joint training, a new G-OSR formulation, and reasonable comparisons on some datasets. However, without a quantitative separation analysis or a threshold selection protocol that includes unseen instances, the central claim should remain conditional. I therefore keep the reader's CONDITIONAL verdict, encoded as UNCHANGED.","tokens_in":15160,"tokens_out":6577,"duration_ms":76655,"concrete_test":"On a held-out split containing both seen and unseen instances (e.g., the standard GZSL validation split with both seen val and unseen val), compute the Eq. 1 entropy for every instance, plot the ROC curve for the seen-vs-unseen decision, and report the AUC. Then select δg two ways: (a) maximize H on this mixed validation split; (b) follow the paper's seen-only G-ZSL-val protocol by choosing δg on seen validation only. Compare the resulting H and the seen/unseen error rates on the same held-out unseen split. If the AUC is not high (say below 0.9) or if protocol (b) differs substantially from protocol (a), the entropy gate is unreliable and the CO-solving claim is unsupported. Also report the entropy distributions of seen and unseen instances separately to show the overlap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the central claim that decomposing G-ZSL into OSR and ZSL 'effectively solves' class overfitting, the method must reliably reject unseen instances before semantic classification. The sole rejection mechanism is the scalar entropy threshold applied to the probability distribution of Eq. 1 (Section III.C.2). This is the weakest point of the argument for three reasons. First, Eq. 1 is a softmax over distances to seen-class visual prototypes; it contains no calibrated model of 'unseenness'. An unseen instance that is visually close to one seen prototype (e.g., a wolf near a dog prototype, or a novel bird species near a seen bird species) yields a sharply peaked distribution, low entropy, and is accepted as seen, so the CO problem reappears exactly for the hard cases. The paper reports only t-SNE intuition (Fig. 1), not a quantitative separation of entropy for seen vs. unseen instances. Second, the threshold is a single dataset-specific constant tuned on the G-ZSL-val split introduced in Section IV.A.2, which is constructed entirely from seen training and validation classes; no unseen instance is available in that split, so there is no principled way to select a threshold that rejects unseen data. Third, the threshold grid [0, 0.000002, 0.02] is extremely small relative to the entropy scale log C (e.g., log 150 ≈ 5.01 for CUB), making the domain decision hypersensitive to the exact operating point. Together these issues mean the reported H numbers, while competitive, do not establish the 'effectively solves' claim; they show partial mitigation under favorable thresholds.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VSG-CNN, a unified end-to-end framework that decomposes generalized zero-shot learning (G-ZSL) into an open set recognition (OSR) task and a zero-shot learning (ZSL) task, and also introduces a new generalized open set recognition (G-OSR) setting. The network has a shared layer, a visual prototype branch (based on convolutional prototype learning, CPL), and a semantic prototype branch for attribute-based classification. At test time, the entropy of the distance-based softmax over seen-class visual prototypes (Eq. 1) is compared with a threshold to decide whether an instance is seen or unseen; seen instances are classified by nearest visual prototype, unseen instances by nearest semantic prototype. Experiments on CUB, AWA2, SUN, and aPY report harmonic-mean accuracies for G-ZSL compared with 20 non-generative and 8 generative methods, and a G-OSR comparison against CPL only. The paper claims that this decomposition effectively solves the class overfitting (CO) problem and that VSG-CNN is the first method to exploit semantic information of known classes in OSR.","tokens_in":15489,"tokens_out":4008,"duration_ms":38043,"significance":"If the central claim were fully supported, the decomposition of G-ZSL into OSR plus ZSL with joint training would be a notable alternative to generative-data-augmentation approaches, and the introduction of G-OSR with semantic descriptions of unknowns would open a useful new direction. The method is clearly specified, uses standard data splits, and is compared with a broad set of recent G-ZSL methods, which are strengths. However, the load-bearing component—the entropy threshold that separates seen from unseen instances—is not theoretically grounded or quantitatively validated, and the G-OSR evaluation rests on a single baseline. The mixed results on SUN and aPY further weaken the general claim that the CO problem is effectively solved. The paper ships no code and reports no error bars, limiting the reproducibility of the numerical claims.","major_comments":[{"comment":"The entropy gate is the only mechanism that separates seen and unseen instances, yet the paper provides no quantitative evidence that the entropy of the distance-based softmax in Eq. (1) actually separates the two domains. Fig. 1 shows t-SNE of prototypes, not the entropy distributions of test instances, and the G-ZSL-val split described in §IV.A.2 and Fig. 4 contains only seen classes, so the threshold δ_g cannot be validated against any unseen instance. Consequently, the central claim that the decomposition 'effectively solves the CO problem' is not established.","section":"§III.C.2, Eq. (1)"},{"comment":"The threshold grid for δ_g and δ_o is [0, 0.02] with step 0.000002, which is several orders of magnitude smaller than the natural entropy scale of Eq. (1) (e.g., log 150 ≈ 5.01 for CUB). No sensitivity analysis with respect to the threshold is reported, so it is unclear whether the final results are stable or artifacts of a very narrow, dataset-specific operating point.","section":"§IV.A.2"},{"comment":"The G-OSR evaluation compares VSG-CNN with only CPL, and the differences are marginal: 70.5→73.3 on CUB, 84.5→84.3 on AWA2 (a slight decrease), 40.3→42.3 on SUN, and 73.6→74.8 on aPY. With no error bars, significance tests, or additional OSR baselines, the claim that introducing semantic information from known classes improves OSR is not demonstrated.","section":"Table 4"},{"comment":"On SUN, VSG-CNN is substantially worse than the best competitor (H = 30.9 vs 41.0 in Table II and vs 40.6 in Table III), and on aPY it is comparable to or worse than several visual-semantic embedding methods (e.g., H = 34.0 vs 36.6). The paper acknowledges these failures but does not explain why the decomposition does not help there, and the abstract and conclusion nevertheless assert that the CO problem is effectively solved in general.","section":"Tables II and III"}],"minor_comments":[{"comment":"The caption says 'COMPARING CRL' but this appears to be a typo; it should likely read 'COMPARING CPL' or 'COMPARING VSG-CNN'.","section":"Table II caption"},{"comment":"There are typos in the attribute descriptions: 'longnech' and 'nech' should be 'longneck' and 'neck'.","section":"Fig. 5"},{"comment":"The paper states that code will be published later but does not report learning rate, batch size, number of epochs, or other training details; these are needed for reproducibility, especially because the reported gains on some datasets are small.","section":"§IV.A.2"},{"comment":"The 'Baseline (ours)' row is described only as a 'non-co-learning strategy'; the exact training procedure and architecture differences from VSG-CNN should be specified for the comparison to be meaningful.","section":"Table II, Baseline"},{"comment":"The definition of G-OSR in item 4) is informal; a formal formulation of the task and its evaluation protocol would help clarify the claimed novelty.","section":"§III.A"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the proposed decomposition is interesting, but the entropy-threshold mechanism is not adequately validated, and the G-OSR evaluation is too thin to support the novelty claim. The authors should be asked to provide quantitative separation evidence for the entropy criterion, a sensitivity analysis of δ_g/δ_o, and a stronger G-OSR comparison, as well as to discuss the SUN and aPY failures more candidly. The self-citation [6] is not an issue by itself, but the 'first method' claim for G-OSR should be checked against the broader open-set-recognition-with-side-information literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the joint end-to-end training of the OSR and ZSL branches is the one genuinely new thing here, and the baseline-vs-VSG-CNN ablation supports it. But the abstract's 'effectively solves the class overfitting problem' is not supported by the reported numbers, and the entropy-based seen/unseen gate is the least grounded part of the method. The paper deserves a serious referee, but it needs revision.\n\nWhat's actually new: The OSR+ZSL decomposition is a known domain-separating idea (CMT, COSMO are cited). The twist is training the two branches jointly in one CNN with a shared frozen backbone, visual prototypes for the OSR branch and predefined semantic prototypes for the ZSL branch. The ablation against a separately trained baseline shows consistent gains on all four datasets, which is real evidence that the co-learning helps. The G-OSR problem—using semantic side-information from known classes to describe rejected unknowns—is a reasonable new task, though the evaluation is thin.\n\nWhere it wobbles:\n\n1. The 'effectively solves' claim. On CUB and AWA2 the harmonic mean is strong (57.0 vs 54.3 for 3ME; 67.0 vs 64.2 for BAAE). But on SUN the method is below the best generative methods (30.9 vs 40.6), and on aPY it is below VSE (34.0 vs 36.6). That does not support 'effectively solves'; it says 'partially mitigates on some datasets'. The paper does acknowledge the SUN failure, but the abstract and intro go further than the data.\n\n2. The entropy gate. The entire rejection of unseen instances rests on a threshold applied to the entropy of the distribution over visual prototypes (Eq. 1). The paper gives t-SNE intuition but no quantitative separation analysis. The G-ZSL-val split used for threshold selection contains only seen classes, so choosing δg on it doesn't directly tell you how the threshold will behave on unseen data. Some might worry that the threshold grid [0, 2e-6, 0.02] is too small relative to log C; I think that part is speculative—for a sharp softmax with γ=1, entropies can easily be well below 0.02—but the validation-set criticism is fair.\n\n3. G-OSR comparison. Only against CPL, which is a single baseline. No comparison to other OSR methods, no ablation of the semantic component in the OSR branch, no error bars. For a new task, that's minimal evidence.\n\nBottom line: The paper is clearly written and the core idea is sensible. The ablation is a good piece of evidence. But the central claim is overreaching, the rejection mechanism is under-analyzed, and the G-OSR evaluation needs more substance. I'd send it to peer review—there's enough here to be worth referee time—but I'd expect revisions to temper the claims and add at least entropy-separation analysis and a broader G-OSR comparison.","headline":"A clearly written domain-separating G-ZSL paper with a real joint-training ablation, but the 'effectively solves' claim overreaches and the entropy gate is under-analyzed.","tokens_in":16039,"tokens_out":5175,"would_cite":false,"duration_ms":48561,"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":"Generalized zero-shot learning is decomposed into a two-stage pipeline—reject what is not a known class, then identify the rejected image by its attributes—and the authors show this split solves the class overfitting problem that plagues…","keywords":["generalized zero-shot learning","open set recognition","convolutional prototype learning","visual-semantic embedding","class overfitting","entropy-based rejection","attribute-based classification","domain separation"],"falsifier":"On a held-out validation set, compute the entropy values for seen and unseen images under the learned visual prototypes and check whether a single threshold separates them. If the entropy distributions overlap substantially (e.g., more than a few percent of unseen images fall below the best seen/unseen threshold), then no threshold can prevent unseen images from being absorbed into seen classes, and the claimed solution to class overfitting fails on that data.","tokens_in":14979,"feed_emoji":"🐾","tokens_out":6875,"duration_ms":65877,"temperature":0.7,"pith_summary":"This paper argues that generalized zero-shot learning (G-ZSL) — recognizing both classes seen in training and classes never seen, using shared attribute descriptions — is best solved by splitting it into two tasks rather than mixing seen and unseen classes in one classifier. A first branch, built on convolutional prototype learning, recognizes seen classes from visual features and rejects anything that does not look like a seen class; the rejected images are then handed to a second branch that matches them to semantic attribute prototypes of unseen classes. The authors claim this decomposition removes the class overfitting problem, where unseen images get forced into seen classes, and they support it with a network called VSG-CNN that trains both branches jointly. On four standard benchmarks, VSG-CNN reaches the best harmonic-mean accuracy on CUB, AWA2, and aPY, and is competitive on SUN.","feed_headline":"Decomposing zero-shot learning fixes seen-class overfitting","feed_subtitle":"A two-branch prototype network rejects unknown images first, then identifies them by attributes; it tops three of four benchmarks.","key_machinery":"The central object is VSG-CNN, a convolutional network with two branches built on a shared frozen feature extractor. A visual-prototype branch (a convolutional prototype network) learns a set of per-class visual prototypes and assigns probabilities by distance-based cross-entropy; a semantic branch maps the same shared features to predefined class attribute vectors. The load-bearing mechanism at test time is an entropy gate: the probability distribution of the test image over seen-class visual prototypes is computed, and its entropy decides the domain—low entropy means 'seen', high entropy means 'unseen'. The threshold is set on a validation split. This gate is what separates the two tasks and is claimed to eliminate the class overfitting problem.","core_discovery":"The paper's central claim is that the class overfitting problem in G-ZSL can be effectively solved by decomposing the task into open set recognition (OSR) followed by zero-shot learning (ZSL). Seen classes are recognized directly in visual space via learned visual prototypes; unseen classes first fail a visual-prototype acceptance test and are then classified in semantic space against predefined attribute prototypes. This separation, implemented in the jointly trained VSG-CNN, means unseen images are never forced into seen classes, and seen images retain the richer discriminability of visual features. The same architecture is repurposed for a new task the paper calls generalized open set recognition (G-OSR), where rejected unknowns are not merely discarded but given a rough attribute description derived from known-class semantic information. On CUB, AWA2, SUN, and aPY, the method reports state-of-the-art or competitive harmonic-mean accuracy, with notably smaller gaps between seen and unseen accuracy than visual-semantic embedding baselines.","pith_inferences":["The same two-branch decomposition should transfer to few-shot learning, where base classes are handled by the visual branch and novel classes by the semantic branch; the entropy gate would need retraining for the new class set.","Because the semantic branch is a simple regression to attribute vectors, swapping it for a more powerful alignment (e.g., a semantic autoencoder) could push harmonic mean higher; the paper explicitly allows this substitution.","A natural evaluation for G-OSR would measure how useful the output attribute vectors are to a human identifying the unknown class, rather than only rejection accuracy.","On datasets where seen and unseen classes are visually very similar, the entropy distributions may overlap; testing on such a dataset would show whether the claimed CO solution degrades gracefully or collapses."],"forward_implications":["Generative models for synthesizing unseen-class instances become unnecessary, sidestepping the blurriness, mode collapse, and instability that limit VAE and GAN approaches.","Seen classes are classified in visual space, preserving subclass discriminability that semantic embeddings blur, as the paper's t-SNE visualization illustrates.","Open set recognition gains an output beyond rejection: rejected unknowns receive a rough attribute description assembled from known-class semantic knowledge.","Joint end-to-end training of the two branches beats training them separately, as the Baseline comparison shows, so the two tasks complement each other.","The smaller seen-unseen accuracy gaps on most datasets indicate the class overfitting problem is mitigated, not just accuracy improved."],"supporting_citations":[{"why":"Supplies the convolutional prototype learning method (distance-based cross-entropy plus prototype loss) that forms the visual-prototype branch of VSG-CNN.","marker":"[5]"},{"why":"Provides the domain-separating baseline and the G-ZSL-val validation protocol used to select the entropy thresholds.","marker":"[44]"},{"why":"Supplies the standard data splits for CUB, AWA2, SUN, and aPY used in all experiments.","marker":"[45]"},{"why":"Defines the harmonic mean evaluation metric on which all G-ZSL comparisons are based.","marker":"[22]"},{"why":"Frames the generalized zero-shot learning problem and the seen/unseen class gap that motivates the class overfitting concern.","marker":"[2]"}],"fun_headline_variants":["Split G-ZSL into open set and zero-shot tasks","Prototype net rejects unknowns, then describes them","Visual-semantic prototypes end seen-class overfitting","Two-step prototype model fixes G-ZSL bias","Open set plus semantic cues beat G-ZSL baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the entropy of a test image's distribution over seen-class visual prototypes cleanly separates seen from unseen instances; if an unseen image happens to be visually close to a seen prototype and produces a low-entropy score, it will be misclassified as seen and the class overfitting problem reappears.","fun_headline_variants_meta":{"raw":{"variants":["Split G-ZSL into open set and zero-shot tasks","Prototype net rejects unknowns, then describes them","Visual-semantic prototypes end seen-class overfitting","Two-step prototype model fixes G-ZSL bias","Open set plus semantic cues beat G-ZSL baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00066,"raw_usage":{"total_tokens":3077,"prompt_tokens":1061,"completion_tokens":2016,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":677,"completion_tokens_details":{"reasoning_tokens":1941}},"tokens_in":677,"tokens_out":2016,"duration_ms":15689,"temperature":1.0,"reasoning_tokens":1941,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:54:58.247764+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out validation set, compute the entropy values for seen and unseen images under the learned visual prototypes and check whether a single threshold separates them. If the entropy distributions overlap substantially (e.g., more than a few percent of unseen images fall below the best seen/unseen threshold), then no threshold can prevent unseen images from being absorbed into seen classes, and the claimed solution to class overfitting fails on that data.","supporting_citations":[{"cited_title":"Robust classification with convolutional prototype learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the convolutional prototype learning method (distance-based cross-entropy plus prototype loss) that forms the visual-prototype branch of VSG-CNN."},{"cited_title":"Adaptive confidence smoothing for gen- eralized zero-shot learning,","cited_arxiv_id":null,"evidence_quote":"Provides the domain-separating baseline and the G-ZSL-val validation protocol used to select the entropy thresholds."},{"cited_title":"Zero-shot learning- a comprehensive evaluation of the good, the bad and the ugly,","cited_arxiv_id":null,"evidence_quote":"Supplies the standard data splits for CUB, AWA2, SUN, and aPY used in all experiments."},{"cited_title":"Zero-shot learning-the good, the bad and the ugly,","cited_arxiv_id":null,"evidence_quote":"Defines the harmonic mean evaluation metric on which all G-ZSL comparisons are based."},{"cited_title":"An empirical study and analysis of generalized zero-shot learning for object recognition in the wild,","cited_arxiv_id":null,"evidence_quote":"Frames the generalized zero-shot learning problem and the seen/unseen class gap that motivates the class overfitting concern."}],"review_version":1}