{"id":"592bea85-ae4a-43e1-ba62-b670fa22e9fb","arxiv_id":"1908.05006","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DEMUD-VIS detects novel images and visualizes their novel content by inverting CNN feature residuals back into images.","lead":"This paper presents DEMUD-VIS, a method that automatically finds unusual images in large collections and shows users a picture of what makes each one unusual. It combines an existing anomaly detector with a neural network that turns abstract image features back into a visual image.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Visual explanations are generated from fc6 while discovery uses fc8; the paper does not show that fc6 residual images isolate the novelty signal that drove selection.","rationale":"The reader's conditional verdict identifies the same load-bearing assumption: the visual explanation must faithfully reflect the semantic novelty that caused DEMUD to select the image. My stress-test agrees with that assessment and finds no additional concern that would change the verdict. The class discovery results are solid and the method is clearly described, but the explanation component is the paper's primary contribution, and it rests on an inversion step that the paper itself reports is unreliable for the best-performing fc8 features. The workaround of generating explanations from fc6 while performing discovery with fc8 introduces a feature-space mismatch that is never validated. This is an internal consistency problem, not a disagreement with external consensus, so it is directly relevant to correctness risk. A concrete inversion-fidelity test, such as checking additive consistency and content selectivity of residual visualizations, would settle whether the concern lands. Because the method could be accepted if that validation is supplied, I keep the reader's CONDITIONAL recommendation unchanged.","tokens_in":17151,"tokens_out":3581,"duration_ms":41035,"concrete_test":"Run a controlled inversion-fidelity experiment on the ImageNet-Yellow data. For each of the first 20 DEMUD-VIS selections using fc6 features, compute the reconstruction and residual from Eqs. 2-3, invert both with the UC network, and also invert the original image's fc6 feature vector. Test two quantitative criteria: (1) additive consistency: the pixel-space distance between the original image and UC(reconstruction) + UC(residual) should be no larger than the distance to UC(original feature vector) alone; (2) content selectivity: using segmentation or bounding-box annotations of the known novel regions (e.g., the dark daisy center in Figure 10c), measure whether the residual image's pixel energy or saliency concentrates on those regions significantly more than a control residual of equal norm from a non-selected image.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DEMUD-VIS's central claim is that the generated visual explanation shows why an image was selected as novel. This requires the up-convolutional inversion of DEMUD's residual vector to be a faithful, content-selective rendering of the information that produced the reconstruction error. The paper's own Section 3.3 reports that UC DeePSiM fails on fc6 and fc8 residuals, that fc6 is fixed only by a global mean-shift transformation (Eq. 4), and that fc8 visualization remains unsolved. Section 4.4.1 then chooses to visualize fc6 features, while the discovery experiments in Section 4.3 show that fc8 gives the best class discovery. Thus the explanation stage and the discovery stage operate in different feature spaces, and no check is provided that the fc6 residual being visualized corresponds to the fc8 novelty that caused the image to be selected. The qualitative examples in Sections 4.4.2 and 4.4.3, and the user study in Section 4.5, do not establish this correspondence: the user study measures only whether explanations change user votes, not whether the visualized residual isolates the algorithm's actual novelty signal. If the inversion is not content-selective, the residual image could highlight arbitrary differences, leaving the claimed explanation of novel discovery unsupported. This is an acknowledged limitation of the method, not merely a disagreement with prior expectations, so it is directly load-bearing for the paper's primary contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DEMUD-VIS, a pipeline for visual explanations of novel image content. It combines CaffeNet CNN features, the DEMUD reconstruction-error novelty detector, and an up-convolutional network (UC DeePSiM) to invert DEMUD's reconstruction and residual feature vectors into pixel-space images. Experiments cover balanced and imbalanced ImageNet subsets, ImageNet-Yellow, Mars-Curiosity, and STONEFLY9, measuring class discovery nAUC and comparing DEMUD with SVD and random baselines. A user study with 280 complete responses tests whether adding explanations changes human novelty judgments. The paper reports that DEMUD with fc8 features yields the best discovery, while fc6 features yield the most interpretable explanations, and that explanations increase detection of within-class novelty in some cases.","tokens_in":17441,"tokens_out":7146,"duration_ms":64850,"significance":"The problem of explaining unsupervised novelty/class discovery in image data is important and under-served. DEMUD-VIS combines three existing components in a novel way and is, to my knowledge, the first to render DEMUD residuals as images. The paper ships data, features, and scripts, and it includes a user study rather than only anecdotal examples. The explicit discussion of the trade-off between explanation interpretability and detection performance is commendable. However, two load-bearing issues (the degenerate SVD baseline and the fc6/fc8 feature-space mismatch) and the mixed user-study evidence prevent acceptance in the current form.","major_comments":[{"comment":"The SVD baseline is degenerate as configured. With k set to min(d,n), the SVD model spans the entire data set, so the reconstruction error R(x) in Eq. (5) is identically zero for every item. The paper nonetheless reports SVD nAUC values in Tables 1-3 that are neither zero nor random (e.g., Table 1, balanced ImageNet: SVD CNN-fc8 nAUC70=60.58), which is internally inconsistent. Please specify the actual k used, the tie-breaking rule, or correct the description; without a meaningful SVD baseline, the claim in Section 4.3 that DEMUD outperforms SVD is unsupported.","section":"4.2.1, Eq. (5)"},{"comment":"The explanation stage and the discovery stage use different feature layers. Section 4.3 identifies DEMUD with fc8 features as the best discovery method, but Section 4.4.1 states that all visualizations are generated from fc6 features because fc6 explanations are more interpretable. No evidence is provided that an fc6 residual corresponds to the fc8 novelty that caused an image to be selected. The user study in Section 4.5 therefore evaluates a different detector than the one that achieved the headline discovery results. Please either add analyses linking fc6 residuals to fc8 selections (e.g., comparing explanations for the same selected images across layers, or measuring how well fc6 residuals predict fc8 novelty) or explicitly delimit the claim of 'explaining novel image discovery' to the fc6-based DEMUD detector.","section":"4.4.1 and 4.3"},{"comment":"The user study provides only weak and partially contradictory evidence for explanation utility. Of 20 images, only 3 (Figure 17) show significantly more 'Yes (new type of object)' votes with explanations, while 3 other images (Figure 18) show significantly fewer such votes with explanations. The paper's explanation for the latter ('users were confused') is speculative, and the significant increase in 'Not sure' responses with explanations contradicts Hypothesis 1. These results should be reported as evidence that explanations can change judgments in both directions, rather than as a clean demonstration that explanations improve users' understanding of novelty.","section":"4.5, Figures 17-18"}],"minor_comments":[{"comment":"The input image size is stated as 227x227 in Section 3.1 and 277x277 in Section 4.1; please reconcile.","section":"3.1 vs 4.1"},{"comment":"The method name 'UC DeePSim' should be 'UC DeePSiM' for consistency.","section":"3.3"},{"comment":"The SIFT comparison reports the best kSIFT among tested values, which is a form of selection on the test data; this should be acknowledged as favorable to SIFT.","section":"4.2.2"},{"comment":"The paper does not report effect sizes or confidence intervals for the significant differences; given the large number of tests (20 images), a multiple-comparison correction would be helpful.","section":"4.5"},{"comment":"The annotation in panel (a) is referenced in the text but is not fully defined; please label the axes or panels consistently.","section":"Figure 10"}],"recommendation":"major_revision","confidential_remarks":"The SVD issue is likely a simple experimental setup error (perhaps k was intended to be a fraction of n), but it must be fixed. The fc6/fc8 mismatch is more conceptual; if the authors can show that fc6 and fc8 selections are largely identical on these data sets, that would mitigate the concern. The user study should be repositioned as a preliminary study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, DEMUD-VIS is a genuine first: nobody had coupled DEMUD's residual explanations with CNN feature inversion to produce human-viewable explanations of novel image content. Second, the paper's own evaluation is more mixed than the abstract suggests, and there are two real problems that need fixing before this should appear in a journal.\n\nThe good stuff. The pipeline is clearly described, and the adaptations—pre-ReLU inversion and the mean-shift transform for fc6 residuals—are sensible engineering. The discovery experiments are consistent across four datasets: DEMUD with CaffeNet features reliably beats random selection, and with fc8 it is near-oracle on balanced ImageNet and strong on the Mars and stonefly data. The authors ship data, features, and scripts. They are also commendably honest about failure modes: fc8 visualizations are unsolved, and the user study did not confirm the first hypothesis.\n\nThe soft spots. The SVD baseline is configured with k equal to the maximum possible number of components. With a full-rank SVD, reconstruction error is zero for every item, so the baseline cannot rank anything. All SVD comparisons are vacuous and should be removed or redone with meaningful k. This does not affect the DEMUD results, but it is a clear methodological error.\n\nThe bigger issue is the layer mismatch. Best discovery uses fc8; best explanation uses fc6. The paper shows that fc6 visualizations are interpretable and fc8 are not, then generates all explanations from fc6. But it never demonstrates that the fc6 residual being visualized corresponds to the fc8 novelty that caused the selection. The user study uses fc6 explanations and therefore does not validate explanations for the best-performing discovery configuration. The central claim should be narrowed: DEMUD-VIS explains fc6-based novelty, not fc8-based novelty. This is a limitation the authors partly acknowledge, but they do not confront how much it weakens the abstract's claim.\n\nThe user study is also partial: for only 3 of 20 images did explanations significantly increase \"Yes\" responses, and explanations increased \"Not sure\" responses overall. That is not strong evidence of utility, though the within-class novelty examples are suggestive.\n\nWho it's for: people working on explainable anomaly detection or on novelty discovery in scientific image collections. It deserves a serious referee. I would accept conditionally, after the SVD baseline is fixed or removed, the layer mismatch is addressed up front, and the claims are scaled back to what the evidence supports.","headline":"A genuine first step for visual novelty explanations, but the SVD baseline is broken and the explanation/discovery layer mismatch weakens the central claim.","tokens_in":17964,"tokens_out":3408,"would_cite":true,"duration_ms":34925,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DEMUD-VIS generates visual explanations of novel image content by inverting reconstruction residuals back into image space, and a user study shows these explanations help people notice within-class novelty.","keywords":["novelty detection","visual explanations","image discovery","convolutional neural networks","feature inversion","reconstruction error","DEMUD","interpretable machine learning"],"falsifier":"Take a labeled image pair where DEMUD-VIS selects a within-class novelty, mask or inpaint the region highlighted in the residual, and check whether the image's reconstruction error drops to near baseline; if removing the residual-highlighted content does not reduce the novelty score, the visualization is not pointing at the semantically novel content.","tokens_in":16948,"feed_emoji":"🖼️","tokens_out":5431,"duration_ms":51440,"temperature":0.7,"pith_summary":"The paper introduces DEMUD-VIS, which it claims is the first method to give visual explanations of novel image content. The pipeline uses a convolutional network to encode images, DEMUD's reconstruction-error ranking to pick images whose content is not yet modeled, and an up-convolutional network to turn the reconstruction and residual feature vectors back into images. The residual image shows what is new in the selected image, while the reconstruction shows what the model already knew. On ImageNet, Mars rover, and stonefly image sets, DEMUD with CNN features, especially from layer fc8, outperformed random and SVD baselines at discovering new classes. A user study found that the explanations shifted judgments toward detecting within-class novelty, although they did not reduce \"not sure\" responses.","feed_headline":"See why an image is new, not just that it is","feed_subtitle":"A CNN novelty detector now draws the residual content that makes each image stand out.","key_machinery":"The central mechanism is the residual vector $r = x' - \\hat{x}'$ produced by DEMUD's incremental SVD model, combined with an up-convolutional network (UC DeePSiM) trained to invert CNN feature vectors back into images. The residual is what carries the novelty: it is the part of the image's feature vector that the current model cannot reconstruct. The paper modifies UC DeePSiM to operate on pre-ReLU feature vectors so that the signs in the residual are preserved, and applies a mean-shift transformation to fc6 residuals to make them visualizable. This inversion step is the load-bearing bridge between feature-space novelty and human-comprehensible explanation.","core_discovery":"The paper's central claim is that novelty detection in large image collections becomes useful for discovery only when the algorithm also says what is new, and that for images this explanation can be rendered visually. DEMUD-VIS does this by treating DEMUD's SVD reconstruction and residual in CNN feature space as images to be generated: the reconstruction visualizes content the model already knew, and the residual visualizes the content that fell outside the model. The paper reports near-oracle class discovery on balanced ImageNet, where the fc8 representation discovered a new class on each of the first 13 selections, and best-in-experiment discovery scores on imbalanced and yellow ImageNet subsets, Mars-Curiosity images, and the STONEFLY9 insect data set. In a user study, users given the visual explanations were more likely than users without them to identify within-class novelty for three specific images at p≤0.05, while the overall rate of \"yes, new type of object\" responses did not change significantly. The paper also documents a trade-off: the best class discovery comes from fc8 features, but the most interpretable visualizations come from fc6 features, so the reported explanations are generated from fc6.","pith_inferences":["If residual inversion could be made reliable at fc8, the method would combine its best discovery performance with its best explanations; the paper's own mean-shift fix works for fc6 but not fully for fc8, so that is the immediate extension.","The approach should transfer to any pretrained CNN whose features can be inverted, so newer architectures with better-trained decoders could improve explanation fidelity without changing the discovery algorithm.","The user-study pattern suggests that explanation utility should be measured after user training rather than on first exposure; a testable design would randomize order or provide a tutorial and then measure novelty-judgment accuracy against held-out labels.","The residual-as-explanation idea generalizes beyond images: any reconstruction-error detector operating on invertible features could render explanations in the input domain, for example spectra or volumetric medical data, as long as a decoder exists."],"forward_implications":["On large image collections, reviewing only the top DEMUD-VIS selections concentrates human attention on genuinely new classes; on balanced ImageNet the first 13 fc8 selections each introduced a new class.","The reconstruction-plus-residual explanation pair can expose within-class changes such as dust deposition on a Mars rover instrument or coat-color variation in foxhounds, not just entirely new categories.","Mislabeled images can be caught automatically: applied to tiger cub images, the ninth selection was a leopard cub, and the residual highlighted the green leaf and leopard spots that explained the labeling error.","Users need some practice with the explanations: in the user study, early selections showed no benefit or even confusion, while later selections showed increased detection of within-class novelty, suggesting that training or reference material would help.","Because fc8 gives the best discovery but fc6 gives the best visual interpretability, the current DEMUD-VIS explanations use fc6 features, implying a direct trade-off between discovery quality and explanation quality."],"supporting_citations":[{"why":"Supplies the DEMUD novelty detector: incremental SVD with reconstruction error and residual-vector explanations that DEMUD-VIS builds on.","marker":"Wagstaff et al. (2013)"},{"why":"Supplies the up-convolutional network that inverts CNN feature vectors into images, the core of the visualization step.","marker":"Dosovitskiy and Brox (2016b)"},{"why":"Supplies DeePSiM, the feature-space plus adversarial loss variant that the paper adapts for pre-ReLU residual and reconstruction visualization.","marker":"Dosovitskiy and Brox (2016a)"},{"why":"Defines the CaffeNet architecture whose fully connected layers fc6, fc7, and fc8 provide the semantic feature representations.","marker":"Krizhevsky et al. (2012)"},{"why":"Supplies the Caffe implementation and CaffeNet model used to extract the CNN features in the experiments.","marker":"Jia et al. (2014)"},{"why":"Supplies the ImageNet benchmark data from which the ImageNet-Random and ImageNet-Yellow subsets are drawn.","marker":"Russakovsky et al. (2015)"},{"why":"Supplies the STONEFLY9 freshwater-stream insect image set and its motivation for filtering non-target images.","marker":"Lytle et al. (2010)"},{"why":"Supplies the labeled Mars-Curiosity rover image data set used for the scientific discovery experiments.","marker":"Stanboli and Wagstaff (2017)"}],"fun_headline_variants":["Drawing the unexpected in images","Show, don't tell: novel image discovery","See the new part of any image instantly","Residual visualization: understanding novelty","Explaining what's new in an image visually"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach stands on the assumption that the generated residual image faithfully shows the content that made the image novel, and since the authors show the inversion can fail for some network layers, the whole utility rests on that fragile step.","fun_headline_variants_meta":{"raw":{"variants":["Drawing the unexpected in images","Show, don't tell: novel image discovery","See the new part of any image instantly","Residual visualization: understanding novelty","Explaining what's new in an image visually"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000243,"raw_usage":{"total_tokens":1536,"prompt_tokens":963,"completion_tokens":573,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":510}},"tokens_in":579,"tokens_out":573,"duration_ms":6334,"temperature":1.0,"reasoning_tokens":510,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:25:58.737403+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a labeled image pair where DEMUD-VIS selects a within-class novelty, mask or inpaint the region highlighted in the residual, and check whether the image's reconstruction error drops to near baseline; if removing the residual-highlighted content does not reduce the novelty score, the visualization is not pointing at the semantically novel content.","supporting_citations":[{"cited_title":"In: Proceedings of the Twenty-Seventh Conference on Artificial Intelligence, pp 905--911","cited_arxiv_id":null,"evidence_quote":"Supplies the DEMUD novelty detector: incremental SVD with reconstruction error and residual-vector explanations that DEMUD-VIS builds on."},{"cited_title":"In: Advances in Neural Information Processing Systems 25, pp 1097--1105","cited_arxiv_id":null,"evidence_quote":"Defines the CaffeNet architecture whose fully connected layers fc6, fc7, and fc8 provide the semantic feature representations."},{"cited_title":"International Journal of Computer Vision 115(3):211--252","cited_arxiv_id":null,"evidence_quote":"Supplies the ImageNet benchmark data from which the ImageNet-Random and ImageNet-Yellow subsets are drawn."},{"cited_title":"Journal of the North American Benthological Society 29(3):867--874","cited_arxiv_id":null,"evidence_quote":"Supplies the STONEFLY9 freshwater-stream insect image set and its motivation for filtering non-target images."},{"cited_title":"Data set on Zenodo, doi:10.5281/zenodo.1049137","cited_arxiv_id":null,"evidence_quote":"Supplies the labeled Mars-Curiosity rover image data set used for the scientific discovery experiments."}],"review_version":1}