{"id":"8f34308a-fdf1-412e-b202-f9643500322a","arxiv_id":"1908.09630","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A VGG-19 network with group Lasso on its first convolutional layer selects 3 to 4 spectral bands per dataset and reaches about 99.9% accuracy on the CMU, HK PolyU, and UWA hyperspectral face datasets.","lead":"Hyperspectral face images contain many narrow wavelength bands. This paper trains a neural network that automatically switches off redundant bands during recognition, using a group Lasso penalty on the first layer, and reports small accuracy gains on three public datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The state-of-the-art claim rests on an uncontrolled cross-paper comparison: Table 4 does not use re-run baselines, gives no split seeds, and Section 9.9's PolyU probe count is internally inconsistent, so the claimed protocol match is unverifiable.","rationale":"After reading the paper, the central claim is the superiority of Deep-SSL over prior band-selection methods in Table 4 and Section 9.9. The method itself is internally coherent: the loss in Eq. (3) is a standard group-Lasso-regularized classification objective, and the grouping of first-layer filter weights by input channel is a sensible way to encourage band sparsity. I do not see a fundamental flaw in the modeling. The place the argument is least secure is the empirical comparison: no code or data are released, split seeds are absent, prior methods are not re-run, and the protocol description contains concrete arithmetic inconsistencies. The PolyU probe-count discrepancy (113 - 48 = 65, not 63) directly undermines the claim that the experimental setup is consistent with prior work, and the near-ceiling accuracies with no error bars amplify the risk that the reported margins are noise. This supports the reader's CONDITIONAL verdict rather than outright acceptance or rejection, so the verdict should remain unchanged.","tokens_in":21320,"tokens_out":6297,"duration_ms":59334,"concrete_test":"Run a controlled re-implementation of PLS* [3], S-CNN+SVM* [51], Deep-Baseline, and Deep-SSL on identical data with the exact Section 9.9 protocol, including a corrected PolyU probe count (113 - 48 = 65 probes), across 10 random gallery/probe split seeds, and report mean plus/minus standard deviation per method. If Deep-SSL's margin over the re-run baselines is not consistently positive on the same splits, or if the margin is within one standard deviation, the state-of-the-art claim in Table 4 is not supported. At minimum, release the exact split files and code so this comparison can be performed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline assertion (Section 9.9) is that Deep-SSL 'outperforms the state-of-the-art methods including PLS* and S-CNN+SVM*.' For this to be true, the accuracies in Table 4 must be measured under the same gallery/probe splits, preprocessing, and training protocol for every method. The paper asserts consistency ('we have been consistent with other compared methods in experimental setup') but does not re-run prior methods and provides no split definitions, seeds, or released code/data. This is not a mere citation detail: the stated PolyU protocol is internally inconsistent. Section 9.9 says the first 24 subjects contain 113 cubes, two cubes per subject are used for gallery (48 cubes), and 'the remaining 63 cubes' are probes; 113 - 48 = 65, not 63. Separately, Section 9.7 says UWA selection produced 3 bands {570, 650, 680} while Table 2 lists four UWA bands {570, 650, 680, 710}, and Section 9.4 reports 300 PolyU cubes while Table 1 and Section 9.9 use 113. The reported accuracies are also near ceiling (99.8-99.9%) with no error bars; a 0.3-0.6 percentage point margin over the all-band baseline corresponds to roughly one or two probe images under the stated probe sizes. If the prior-method numbers come from different splits, or if the margin is within split-to-split noise, the state-of-the-art claim fails even though the group-Lasso mechanism itself is coherent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Deep-SSL, a VGG-19-based CNN for hyperspectral face recognition with automatic spectral-band selection. The first convolutional layer has one input channel per spectral band, and a group-Lasso (structural sparsity learning) penalty Rg(w1) in Eqs. (5)-(6) is added to the softmax-plus-center loss in Eqs. (3)-(4), so that groups of first-layer weights corresponding to redundant bands are driven to zero during joint training; band selection is therefore end-to-end rather than manual or greedy. The method is evaluated on three public datasets (CMU-HSFD, HK PolyU-HSFD, UWA-HSFD) with reported accuracies of 99.9%, 99.8%, and 99.9%, respectively. The central assertion (Section 9.9, Table 4) is that Deep-SSL outperforms state-of-the-art methods including PLS* and S-CNN+SVM*, and that band selection improves accuracy over the all-band deep baseline by 0.32-0.59 percentage points (Section 9.8, Table 3).","tokens_in":21639,"tokens_out":16358,"duration_ms":135071,"significance":"If the comparison were controlled, this would be a useful contribution: Eqs. (3)-(6) give a clear group-Lasso formulation for channel-level band selection trained jointly with recognition, avoiding manual and greedy selection, and the evaluation covers three public datasets with a broad slate of comparison methods in Table 4. Credit is due for the clean formulation and the multi-dataset effort. However, the marginal gain of the band-selection mechanism over the all-band deep baseline is only 0.32-0.59 percentage points at near-ceiling accuracy (99.3-99.9%), so the substantive advantage claimed for band selection is small and not yet shown to be statistically stable; the dominant effect is the strength of the deep baseline itself. The headline state-of-the-art claim is not robustly evidenced, and several internal inconsistencies weaken the experimental record. The core idea is plausible and publishable, but the evidence needs substantial strengthening.","major_comments":[{"comment":"The headline assertion that Deep-SSL outperforms PLS* and S-CNN+SVM* is based on accuracies taken from the original papers rather than on re-runs under the protocol used for Deep-SSL. The sentence 'we have been consistent with other compared methods in experimental setup' asserts protocol matching, but no split definitions, random seeds, preprocessing details (e.g., alignment, cropping, image resolution), or per-split results are given for the prior methods, and code is not released. Because the prior methods use entirely different feature pipelines (e.g., 3D-DCT and PLS in [3]) and possibly different gallery/probe compositions, the numbers in Table 4 cannot be verified as comparable; if the prior-method results come from different splits, the stated state-of-the-art advantage fails even if the group-Lasso mechanism itself works. The authors should either re-run all baselines under the single evaluation protocol used for Deep-SSL and report per-split and mean accuracy, or explicitly downgrade the claim to 'competitive with previously published results.'","section":"Section 9.9, Table 4"},{"comment":"The experimental description is internally inconsistent in several places. Section 9.4 states that the HK PolyU dataset contains 300 hyperspectral cubes from 24 subjects, while Table 1 and Section 9.9 state 113 cubes. Section 9.9 says the PolyU gallery uses two cubes per subject (48 cubes) with 'the remaining 63 cubes' as probes, but 113 - 48 = 65. Section 9.7 reports three selected UWA bands {570, 650, 680}, while Table 2 lists four UWA bands {570, 650, 680, 710}. These discrepancies are not merely typographical: the probe-count error changes the interpretation of the claimed 0.36% improvement (a fraction of one probe image), and the band-list conflict makes the selected-band result ambiguous. All of these numbers must be reconciled before the evaluation can be assessed.","section":"Sections 9.4, 9.7, 9.9; Tables 1 and 2"},{"comment":"No error bars or multiple-run statistics are reported, although gallery images are chosen randomly (Section 9.9) and lambda_g is selected by cross-validation (Section 9.3). With probe sets of approximately 99 (CMU), 63 or 65 (PolyU), and 50 (UWA) images, the reported SSL gains over the all-band baseline (0.59%, 0.36%, and 0.32%, respectively) correspond to less than one probe image per dataset, so the improvement attributed to band selection is within plausible split-to-split noise unless demonstrated otherwise. In addition, Table 4 reports Deep-SSL PolyU accuracy as 99.8%, while Table 3 reports 99.88%, which rounds to 99.9%; the discrepancy should be resolved. Report mean and standard deviation over multiple random splits and per-split accuracies for Deep-Baseline and Deep-SSL.","section":"Section 9.8, Tables 3 and 4"},{"comment":"The mapping from the trained weights to the selected-band lists in Table 2 is not specified. Since the network is trained with Adam rather than a proximal-gradient method, the group norms likely only approach zero in practice; the paper does not state the threshold used to declare a band zeroed out, nor whether the network is retrained after the claimed depth reduction C' << C (Section 8.5), and the specific lambda_g values used per dataset are not reported (Fig. 8 gives only a sensitivity range). The initialization of the first-layer filters by duplicating the ImageNet RGB filters (Section 9.2) is an unvalidated heuristic that could bias which input channels survive group sparsification. An ablation with random initialization and a precise description of how the band lists are extracted from the trained filters are needed for the band-selection result to be reproducible.","section":"Sections 8.4-8.5, 9.2, 9.7"}],"minor_comments":[{"comment":"The CMU spectral range is stated as 450-1100 nm in the text but 450-1090 nm in Table 1; since 65 bands at 10-nm steps implies 450-1090 nm, the text should be corrected.","section":"Section 9.4 vs. Table 1"},{"comment":"The text assigns Fig. 9 to CMU and Fig. 10 to PolyU, while the figure captions assign Fig. 9 to HK PolyU-HSFD and Fig. 10 to CMU-HSFD; the assignments should be made consistent.","section":"Section 9.8 vs. Figs. 9-10"},{"comment":"An unfinished citation placeholder '[?,?,?,?]' appears in the deep-learning face-recognition sentence and must be completed.","section":"Section 2.2"},{"comment":"References [41] and [86] both cite Simonyan and Zisserman (VGG) and should be merged; the notation L(w) is used for two different losses in Eqs. (1) and (3) and should be disambiguated.","section":"References; Eqs. (1) and (3)"},{"comment":"The notation 'l_12 norm' for the group penalty in Eq. (5) is non-standard; since Eq. (6) is an l1 sum of l2 norms, the standard notation is l2,1 (or group Lasso), and this should be stated explicitly.","section":"Section 8.4, Eq. (5)"},{"comment":"The stray footnote marker in 'Please email us 1' should be removed, and the offer of code and data only by email limits reproducibility; public release of the code and split definitions is recommended.","section":"Section 9.9"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is written as a book chapter (LNCS format), and Sections 3-7 are largely tutorial material on sparsity, compression, regularization, and CNN basics; for a journal-length version these should be condensed substantially, since the actual contribution is in Sections 8-9. The main review risk is the state-of-the-art claim in Section 9.9: if the Table 4 numbers for prior methods are not re-measured under the Deep-SSL protocol, I recommend requiring the claim to be reworded. The internal inconsistencies (300 vs. 113 PolyU cubes; 63 vs. 65 probes; three vs. four UWA bands; 99.88 vs. 99.8) also suggest the experimental section needs a careful audit before the paper can be accepted. I would support acceptance after these issues are resolved; the group-Lasso band-selection idea itself is coherent and plausibly useful."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The core mechanism is straightforward: the authors put a group-Lasso penalty on the first convolutional layer of a VGG-19, with each spectral band as a group, and zero out redundant bands during training. That is a clean, sensible application of known structured-sparsity ideas, and the loss in (3)-(6) is unambiguous. The paper also gives per-band accuracy plots, which help show why the selected bands make sense.\n\nThe soft spot is the evidence for the headline claim. Table 4 compares Deep-SSL against a long list of prior methods, but those numbers appear to be taken from earlier papers rather than re-run under the same splits and preprocessing. The authors assert consistency in setup, but give no split definitions, no seeds, and no code or data at submission. On top of that, the paper has internal arithmetic problems: the PolyU probe count in Section 9.9 is 63, but 113 cubes minus 48 gallery cubes leaves 65; Section 9.4 says 300 PolyU cubes while Table 1 uses 113; and Section 9.7 says UWA selection found three bands, while Table 2 lists four. These are the sort of thing that makes a careful reader stop trusting the reported numbers.\n\nThe gains over the all-band baseline are also tiny: 0.32 to 0.59 percentage points at 99.9% accuracy. Without error bars or multiple random splits, that margin is plausibly one or two probe images, so the \"outperforms all bands\" claim could easily be split noise. The mechanism could still work; the paper just does not demonstrate it convincingly.\n\nWho is this for? Researchers working on hyperspectral face recognition or band selection. It deserves a serious referee because the method is coherent and the datasets are public in principle, but the current write-up needs real changes: re-run or clearly identify the baseline numbers, provide error bars and split info, fix the inconsistencies, and release code/data. If those are addressed, it could become a solid niche contribution. As it stands, the stated SOTA advantage is unverified.","headline":"A coherent group-Lasso band-selection method whose reported state-of-the-art advantage is not verifiable from the paper as written.","tokens_in":22197,"tokens_out":2840,"would_cite":false,"duration_ms":26562,"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":"A CNN regularized with group Lasso automatically selects a small set of spectral bands for face recognition, and on three public hyperspectral datasets this selection beats both all-band training and prior band-selection methods.","keywords":["hyperspectral face recognition","band selection","group lasso","structured sparsity","convolutional neural network","center loss","spectral band selection","deep learning"],"falsifier":"Re-run PLS* and S-CNN+SVM* under the exact gallery/probe splits, preprocessing, and random seeds used for Deep-SSL on all three datasets. If either baseline reaches or exceeds Deep-SSL's accuracy under identical conditions, the paper's central state-of-the-art claim fails; if Deep-SSL still wins, the claim stands even under scrutiny.","tokens_in":21081,"feed_emoji":"🎯","tokens_out":5948,"duration_ms":54535,"temperature":0.7,"pith_summary":"This paper claims that hyperspectral face recognition is best done not with all available spectral bands but with a small automatically chosen subset, and that a convolutional network can learn that subset while it learns to recognize faces. The proposed method, Deep-SSL, feeds images from every band into a CNN and adds a group-Lasso penalty on the first-layer filter weights, which drives the weights of redundant spectral bands to zero during training. On three public hyperspectral face datasets (CMU, HK PolyU, and UWA), the resulting network selects three or four bands, mostly in the green, red, and near-infrared range, and reaches accuracy slightly above what the same network achieves with all bands. If correct, the work replaces manual or greedy band selection with a selection that is learned end-to-end as part of recognition, and it points to which wavelengths actually carry identity information.","feed_headline":"Sparse band selection beats full-spectrum face recognition","feed_subtitle":"A group-Lasso penalty makes a CNN drop redundant wavelengths and top prior methods on three hyperspectral face datasets.","key_machinery":"The load-bearing mechanism is group-Lasso regularization applied to the weights of the first convolutional layer. Each group consists of all first-layer weights that share the same input channel, and since each input channel is one spectral band, zeroing a group removes that band from the network. The total loss is the face-recognition loss (softmax plus center loss) plus a scaled $\\ell_{1,2}$ norm of the groups; the $\\ell_1$ part encourages entire groups to vanish while the $\\ell_2$ part keeps weights within a surviving group intact. During training the optimizer jointly minimizes recognition error and this structured sparsity penalty, so the network itself decides which wavelengths are redundant. The center loss pulls features of the same subject together, supplying the discriminative signal that the band-selection penalty is balanced against.","core_discovery":"The central claim is that Deep-SSL, a CNN whose first convolutional layer is regularized by a group-Lasso term, achieves the highest face recognition accuracy reported on the CMU, HK PolyU, and UWA hyperspectral face datasets, outperforming prior methods including PLS-based regression and a CNN with AdaBoost band selection. The same network without the sparsity term already scores in the high 99% range, and adding the group-Lasso penalty raises accuracy by roughly 0.3 to 0.6 percentage points while shrinking the input from 33 or 65 bands to 3 or 4 bands. The selected bands cluster in the green, red, and infrared portions of the spectrum, and the blue bands are zeroed out first. The paper presents this as evidence that jointly optimizing band selection and face recognition, rather than selecting bands greedily or manually, yields both higher accuracy and a smaller, cheaper spectral acquisition requirement.","pith_inferences":["Beyond the paper, the same group-Lasso channel-selection trick could be applied to hyperspectral remote sensing or medical imaging, where the selected wavelengths would be task-specific rather than tied to faces.","A testable extension is to verify whether the selected band set is stable across illumination conditions, skin types, and sensor hardware; the paper's three datasets are all indoor, and outdoor or cross-session data could shift the optimal wavelengths.","The accuracy gains over using all bands are small, so the practical value may lie less in raw accuracy and more in the ability to build cheaper multi-spectral cameras that capture only the selected bands."],"forward_implications":["If Deep-SSL's results hold, hyperspectral face recognition systems can drop the majority of their captured bands and still match or slightly exceed all-band accuracy, which simplifies sensor design and reduces data volume.","The selected bands from all three datasets fall in the green, red, and near-infrared range and exclude blue, suggesting those wavelengths carry the most identity information for faces under indoor illumination.","Because the sparsity penalty acts on the first layer of any CNN, the same training scheme can be transplanted to other deep backbones and other hyperspectral tasks without changing the network's task-specific layers.","The comparison table places Deep-SSL above PLS-based regression and CNN-plus-AdaBoost band selection on all three datasets, making it the current best reported method for hyperspectral face recognition when all methods use their optimal bands."],"supporting_citations":[{"why":"Provides the CMU hyperspectral face dataset used in the main evaluation.","marker":"[98]"},{"why":"Provides the HK PolyU hyperspectral face dataset and the manual band-selection baseline that the method is compared against.","marker":"[18]"},{"why":"Provides the UWA hyperspectral face dataset used for evaluation.","marker":"[99]"},{"why":"Is the PLS-based hyperspectral face recognition method (PLS*) whose accuracy Deep-SSL claims to beat.","marker":"[3]"},{"why":"Is the prior CNN-based band selection method (S-CNN+SVM*) that Deep-SSL claims to outperform.","marker":"[51]"},{"why":"Is the group-Lasso regression formulation that supplies the structured sparsity penalty used for band selection.","marker":"[54]"},{"why":"Is the center-loss formulation used as the recognition loss that the sparsity term is balanced against.","marker":"[95]"},{"why":"Is the deep CNN architecture used as the backbone for the proposed method.","marker":"[86]"}],"fun_headline_variants":["CNN prunes to 4 wavelengths, beats full-spectrum face recognition","Group-Lasso bands: 65 down to 4, accuracy up on hyperspectral faces","Deep-SSL auto-selects 3-4 bands, outperforms prior face methods","Sparse band CNN tops hyperspectral face recognition with fewer wavelengths","Joint band selection and face matching wins with 4 wavelengths"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that Deep-SSL beats prior methods rests on the assumption that the accuracies reported for PLS*, S-CNN+SVM*, and other baselines were measured under exactly the same gallery/probe splits, preprocessing, and evaluation protocol as Deep-SSL; the paper states that the setup was consistent but does not re-run the baselines or publish the splits and seeds.","fun_headline_variants_meta":{"raw":{"variants":["CNN prunes to 4 wavelengths, beats full-spectrum face recognition","Group-Lasso bands: 65 down to 4, accuracy up on hyperspectral faces","Deep-SSL auto-selects 3-4 bands, outperforms prior face methods","Sparse band CNN tops hyperspectral face recognition with fewer wavelengths","Joint band selection and face matching wins with 4 wavelengths"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1370,"prompt_tokens":907,"completion_tokens":463,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":365}},"tokens_in":523,"tokens_out":463,"duration_ms":5348,"temperature":1.0,"reasoning_tokens":365,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:05:01.099886+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run PLS* and S-CNN+SVM* under the exact gallery/probe splits, preprocessing, and random seeds used for Deep-SSL on all three datasets. If either baseline reaches or exceeds Deep-SSL's accuracy under identical conditions, the paper's central state-of-the-art claim fails; if Deep-SSL still wins, the claim stands even under scrutiny.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CMU hyperspectral face dataset used in the main evaluation."},{"cited_title":"Hyperspectral face recognition using 3D-DCT and partial least squares.,","cited_arxiv_id":null,"evidence_quote":"Provides the UWA hyperspectral face dataset used for evaluation."},{"cited_title":"Hyperspectral cnn for image classifi- cation & band selection, with application to face recognition,","cited_arxiv_id":null,"evidence_quote":"Is the prior CNN-based band selection method (S-CNN+SVM*) that Deep-SSL claims to outperform."},{"cited_title":"Model selection and estimation in regression with grouped variables,","cited_arxiv_id":null,"evidence_quote":"Is the group-Lasso regression formulation that supplies the structured sparsity penalty used for band selection."},{"cited_title":"A discriminative feature learning approach for deep face recognition,","cited_arxiv_id":null,"evidence_quote":"Is the center-loss formulation used as the recognition loss that the sparsity term is balanced against."}],"review_version":1}