{"id":"46cc5b71-d24a-4041-a680-0468219e1248","arxiv_id":"1908.04389","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"NeuroMask optimizes a spatial mask so a pre-trained classifier's output stays stable, then presents the mask as an explanation of which image regions drove the prediction.","lead":"NeuroMask learns a mask over an input image that hides unimportant pixels while keeping the classifier's output unchanged, producing a heat-map style explanation for a deep network's decision. The paper compares the resulting masks visually with six existing explanation methods on CIFAR-10 and ImageNet images.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on the unvalidated assumption that mask-based output preservation reveals model evidence; the paper provides no quantitative localization test, no ground-truth comparison, and no user study, and its own future-work section admits the missing evaluation.","rationale":"The reader's weakest assumption and my concern are the same: the paper defines explanation by construction, never validating that the optimized mask actually corresponds to the model's decision evidence. This is load-bearing because the central claim is explicitly about localization accuracy and interpretability, and the only support offered is qualitative visual inspection by the authors. The paper's own statement in Section IV-B that user studies are future work is a direct admission that the interpretability claim is not yet tested. I also note the Section III versus Algorithm 1 inconsistency in Lsparse, which makes the described method non-reproducible and obscures which objective actually produced the displayed masks. However, neither the lack of validation nor the inconsistency disproves the method; they show that the paper's central claim is premature. A synthetic known-decision-rule experiment would settle the core question by providing ground truth for what the network should rely on, avoiding the ambiguity of natural-image object annotations. If NeuroMask passes such a test, the conditional acceptance would be justified; if it fails, the central claim would need to be substantially weakened. Since the reader already assigned CONDITIONAL with the same underlying concern, my read does not change the verdict, so I recommend UNCHANGED with the concrete test as a natural condition for acceptance.","tokens_in":8071,"tokens_out":5370,"duration_ms":60558,"concrete_test":"Train a small CNN on synthetic images where each label is determined solely by a shape rendered at a random position on a uniform background. Apply NeuroMask to 100 held-out test images, threshold the resulting mask, and compute the mean IoU between the thresholded mask and the ground-truth shape bounding box. Also run Fong and Vedaldi's method [15] as a baseline on the same task. If NeuroMask's mean IoU is not clearly above chance and not comparable to [15], then the assumption that output-preserving masks recover the model's decision evidence is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"NeuroMask's central claim — that its masks 'successfully localize the parts of the input image which are most relevant to the DNN decision' and are 'both accurate and interpretable' (Abstract; Section IV-B) — rests on the premise, introduced in Section III, that minimizing the combination of Lpred, Lsparse, and Lsmooth yields masks that reveal the evidence the network actually used. Nothing in the paper independently tests this premise. The evaluation is the authors' visual inspection of a small number of CIFAR-10 and ImageNet examples, with no localization metrics, no ground-truth object annotations, no deletion/insertion measurements, no model with a known decision rule, and no user study. The paper itself concedes the missing validation in Section IV-B: 'in the future, we plan to conduct more comprehensive evaluation studies that include doing user's surveys.' A mask can satisfy the objective by preserving the predicted class probability while highlighting a small, spurious, or merely sufficient set of pixels, so the observed qualitative examples do not establish that the mask corresponds to the model's actual evidence. A secondary reproducibility problem compounds this: Section III defines Lsparse as the sum of |W_ij + τ| with τ = 20, but Algorithm 1, line 9, sets Lsparse = |W|_1, which penalizes W toward 0 rather than toward −τ; the two formulations will produce different masks, and the text provides no reconciliation. The method may work, but the central claim is empirically unverified as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NeuroMask, a model-agnostic method for explaining image classification decisions by learning a multiplicative mask over the input image. Two frozen copies of a pretrained classifier receive the original image and the masked image; the mask weights are optimized with RMSProp to minimize a weighted sum of a prediction-preservation term (cross-entropy between the two model outputs), a sparsity term, and a smoothness term. The authors evaluate NeuroMask on CIFAR-10 and ImageNet with pretrained convolutional networks and compare the resulting masks qualitatively against saliency maps, SmoothGrad, Grad-CAM, Guided Backprop, LIME, and LRP-epsilon. Based on visual inspection, they conclude that NeuroMask produces accurate and interpretable explanations that localize the parts of the image most relevant to the model's decision.","tokens_in":8316,"tokens_out":4001,"duration_ms":41141,"significance":"If the central claim were fully supported, NeuroMask would be a useful addition to the interpretability toolbox: it is model-agnostic, requires no retraining or architecture modification, and produces a directly visualizable mask. The paper also compares against six established explanation methods, which is a reasonable starting point. The conceptual formulation, combining output preservation with sparsity and smoothness, is clear and the visual examples in Figures 2 and 3 are suggestive. However, the current evidence is almost entirely qualitative. The abstract and Section IV-B claim that NeuroMask 'successfully localizes' relevant image parts, but no quantitative localization, faithfulness, or human-evaluation evidence is provided, and the paper itself concedes the missing validation at the end of Section IV-B. In addition, the sparsity term is defined inconsistently between Section III and Algorithm 1, which directly affects what objective is actually optimized. No code, hyperparameters, or numerical results are supplied, limiting reproducibility. The central idea is defensible, but the empirical claim needs substantially more support.","major_comments":[{"comment":"The central claim that NeuroMask 'successfully localizes the parts of the input image which are most relevant to the DNN decision' rests entirely on visual inspection of a small number of examples. There are no quantitative localization metrics (e.g., pointing game or IoU with ground-truth object masks), no deletion/insertion or output-preservation statistics, no comparison against a model with a known decision rule, and no user study. Because the objective itself encourages the masked input to preserve the classifier's output, output preservation alone cannot distinguish a mask that reflects the model's actual evidence from a mask that merely finds some small sufficient pixel set. The paper explicitly acknowledges this gap in Section IV-B ('in the future, we plan to conduct more comprehensive evaluation studies that include doing user's surveys'). I request quantitative evaluation on a benchmark with ground-truth annotations, plus at least one faithfulness measure, before the localization claim can be accepted.","section":"Abstract and Section IV-B"},{"comment":"There is a direct inconsistency in the definition of the sparsity objective. In Section III, the paper defines Lsparse(W) = sum_{i,j} |W_{i,j} + tau| with tau = 20, chosen so that sigma(-tau) is approximately zero, which biases mask weights toward -tau and hence mask values toward 0. In contrast, Algorithm 1 line 9 sets Lsparse = |W|_1, which biases mask weights toward 0. These are different objectives and will produce different masks and different gradients. The paper provides no reconciliation of the two formulations. The authors must state which objective is actually minimized and ensure that Algorithm 1, the text, and any released code agree.","section":"Section III, Sparseness cost, and Algorithm 1, line 9"},{"comment":"The behavior of NeuroMask depends critically on the weighting coefficients lambda_p, lambda_sp, lambda_sm, the sparsity offset tau, the RMSProp learning rate alpha and decay beta, and the number of optimization steps T, but none of these values are reported anywhere in the manuscript. Without this information the experiments are not reproducible, and it is unclear how sensitive the qualitative results are to the balance between prediction preservation, sparsity, and smoothness. I request that all hyperparameters be reported and that at least a small ablation study over the main loss weights be included.","section":"Section III, Ltotal and Algorithm 1"}],"minor_comments":[{"comment":"The heading 'Smoothness cost Lsparse' and the following sentence 'the definition of Lsparse is chosen to be the L1 norm of the 2nd derivative' should refer to Lsmooth, not Lsparse.","section":"Section III, Smoothness cost"},{"comment":"The paper cites Inception-v3 as reference [22], but [22] is Montavon et al.'s deep Taylor decomposition paper; the Inception-v3 architecture should be cited as [21] (Szegedy et al.).","section":"Section IV-B"},{"comment":"The algorithm input is written as x in R^{H x W}, whereas the text defines x in R^{H x W x 3}; the notation should be aligned.","section":"Algorithm 1"},{"comment":"The introduction says Section III describes the 'assumptions' of NeuroMask, but Section III contains no explicit list of assumptions; please adjust the wording or add an assumptions subsection.","section":"Section I"},{"comment":"Figure 4 is referred to as 'Figure IV-C' in the text; the cross-reference should be to Figure 4.","section":"Section IV-C"},{"comment":"There are repeated typos and grammatical errors, including 'state-of-art', 'intepretability', 'the the model', and 'systemically occlude'; these should be corrected in a journal version.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as an early-stage workshop-style paper rather than a complete journal submission. The two main barriers are (1) the absence of any quantitative validation of the central localization/interpretability claim, and (2) the unresolved inconsistency in the sparsity objective between the text and Algorithm 1. Both are fixable with additional experiments and careful revision, but they are load-bearing for the current claims. I would also encourage the authors to check the reference numbering before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's the quick take on arXiv:1908.04389. It's a small paper on explaining image classifiers by learning a mask that suppresses uninformative pixels while preserving the model's output. The core idea is not new—the paper itself cites Fong & Vedaldi (2017) as \"most similar\"—and the differences are incremental: cross-entropy prediction loss plus a Laplacian smoothness term. There is no quantitative comparison against that baseline or any other.\n\nWhat the paper does well: the algorithm is clearly described, the two-clone architecture is simple and sensible, and the qualitative figures (which I can't see, but the text describes) suggest the masks can highlight object-level regions. The authors are honest about the similarity to [15] and about the absence of a user study.\n\nWhere it falls short, proportionally: The central claim that NeuroMask \"successfully localizes\" relevant parts and is \"accurate and interpretable\" is supported only by visual inspection of a handful of examples. No localization metrics, no ground-truth comparison, no deletion/insertion tests, no error bars, no hyperparameter values. The paper's own future-work section admits the missing evaluation. This means the claim is, as far as the written paper goes, unverified. There is also a real internal inconsistency: Section III defines the sparsity loss as sum |W_ij + tau| with tau=20 (pushing W toward -tau), while Algorithm 1 line 9 uses |W|_1 (pushing toward 0). These produce different optima, and the text never reconciles them. That's a concrete bug in the write-up.\n\nThe circularity concern is real but not fatal by itself: any mask that preserves the output will by construction produce a \"faithful\" explanation under that objective; the question is whether the mask also corresponds to what a human or ground truth would call relevant. The paper doesn't test that.\n\nWho's this for? A reader new to mask-based explanation methods might find it a useful entry point, and it could serve as a cautionary example of why qualitative evaluation is insufficient. But it doesn't advance the state of the art. I would not cite it in my own work, and I wouldn't send it to a serious referee in its current form. If the authors add a quantitative comparison to [15], report hyperparameters, and fix the Lsparse inconsistency, it could become a marginal workshop paper.","headline":"A clearly-written mask-learning explanation method that is essentially Fong & Vedaldi with a different loss, undermined by purely qualitative evaluation and a concrete inconsistency in the objective.","tokens_in":8894,"tokens_out":2525,"would_cite":false,"duration_ms":24973,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"NeuroMask learns a sparse, smooth mask over the input image that preserves a pre-trained classifier's prediction, and claims this mask localizes the pixels that drove the decision.","keywords":["explainable AI","mask learning","post-hoc explanations","image classification","saliency maps","perturbation-based explanations","deep neural network interpretability"],"falsifier":"A concrete test: on images with ground-truth object segmentations, if a NeuroMask mask preserves the classifier's class probability while its high-weight pixels have near-zero overlap with the object, the localization claim is falsified; the same test with known spurious background cues would reveal whether the explanation tracks bias instead of the object.","tokens_in":7824,"feed_emoji":"🎯","tokens_out":6771,"duration_ms":66742,"temperature":0.7,"pith_summary":"NeuroMask sets out to show that a post-hoc explanation of a deep image classifier can be generated by learning a single-channel mask over the input: multiply the image by the mask, feed it to the frozen network, and require the predicted class distribution to stay nearly the same while the mask is forced to be sparse and smooth. If the claim holds, the mask highlights the pixels that mattered to the decision without retraining or modifying the network, which would make the method usable for auditing already-deployed classifiers in settings where explanations are required. The paper demonstrates the idea with visual comparisons on CIFAR-10 and ImageNet, arguing that its masks localize objects and object parts more cleanly than several existing saliency, gradient, and perturbation methods.","feed_headline":"A learned mask shows which pixels drive a neural net's verdict","feed_subtitle":"Sparse, smooth masks hide irrelevant pixels without retraining the classifier, producing human-readable explanations.","key_machinery":"The central object is the relevance mask $m$, a per-pixel weight in $[0,1]$ learned through RMSProp by minimizing a weighted sum of three losses: $\\mathcal{L}_{\\text{pred}}$, the cross-entropy between the outputs of two frozen copies of the classifier, one fed the original image and one fed the masked image; $\\mathcal{L}_{\\text{sparse}}$, an $\\ell^1$ penalty on mask weights that drives most pixels toward zero; and $\\mathcal{L}_{\\text{smooth}}$, an $\\ell^1$ penalty on the mask convolved with a discrete Laplacian filter that forces highlighted regions to be contiguous. The trainable weights $W$ are mapped to $m$ through a sigmoid, and the two frozen model copies supply the constraint that a good explanation must be sufficient: hiding everything outside the mask should not change the verdict.","core_discovery":"The paper's central claim is that explanation can be framed as mask optimization. Given an input image $x$ and a frozen pre-trained classifier $f$, NeuroMask learns $m \\in [0,1]^{H\\times W}$ so that $f(x)$ and $f(m \\odot x)$ predict the same class, where the prediction cost is cross-entropy between the two softmax outputs, while $\\ell^1$ penalties push the mask toward being sparse and spatially smooth. The authors argue that the resulting mask is both accurate, in the sense that it localizes the parts of the image most relevant to the decision, and interpretable, in the sense that a human sees contiguous object parts rather than scattered pixels. They support this with qualitative comparison against Saliency Map, Smoothed Grad, Grad-CAM, Guided Backprop, LIME, and LRP-epsilon on CIFAR-10 and ImageNet examples.","pith_inferences":["A natural quantitative extension the paper does not run is to score masks by deletion and insertion, removing the highlighted pixels and measuring how fast the class probability drops; if the mask is truly decision-relevant, removal should be devastating.","Because the objective only enforces output preservation, the method could in principle highlight a spurious background cue instead of the object; comparing masks on counterfactual image pairs, same object with different backgrounds, would reveal whether NeuroMask tracks the object or the dataset bias.","The smoothness penalty's scale is a free parameter; varying it and checking masks against ground-truth object segmentations would provide the first objective test of the accuracy claim, which the paper currently supports only by visual inspection."],"forward_implications":["A mask that preserves the predicted class while being sparse and smooth is a candidate minimal evidence set: pixels outside it are, by construction, dispensable for the decision.","Because the network weights are never updated, NeuroMask can be applied to any already-deployed image classifier that supports gradient backpropagation through the input.","The same three-term objective could be ported to other outputs and modalities, which the paper names as future directions: image captioning, text, and sound.","The visual comparisons imply NeuroMask's masks sit between overly noisy pixel-level methods and overly coarse region-level methods: cleaner than saliency and guided backprop, finer than Grad-CAM, and less dependent on superpixel boundaries than LIME."],"supporting_citations":[{"why":"Supplies the Saliency Map baseline whose noisy outputs NeuroMask contrasts with its own masks.","marker":"[4]"},{"why":"Supplies the LIME baseline whose superpixel dependence is the failure mode NeuroMask's smoothness term targets.","marker":"[5]"},{"why":"Supplies the Smoothed Grad baseline that NeuroMask is visually compared against and judged comparable to.","marker":"[6]"},{"why":"Supplies the Guided Backprop baseline whose noisy outputs are contrasted with NeuroMask's masks.","marker":"[9]"},{"why":"Supplies the LRP-epsilon baseline described as too conservative in what it highlights.","marker":"[10]"},{"why":"Supplies the Grad-CAM baseline described as too coarse-grained relative to NeuroMask.","marker":"[13]"},{"why":"The closest prior mask-perturbation approach, used to position NeuroMask's cost-function design as its novel contribution.","marker":"[15]"},{"why":"Supplies the RMSProp optimizer used to update the mask weights in the learning loop.","marker":"[18]"}],"fun_headline_variants":["Sparse masks explain neural net decisions without retraining","NeuroMask: Optimize a mask to show why a net predicted that","Explain any CNN by learning a sparse, smooth pixel mask","A learned mask hides pixels to reveal what a DNN actually saw"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a mask that preserves the classifier's output while being sparse and smooth is necessarily showing the evidence the network used; the paper asserts this link but never tests it against ground-truth object locations or human judgment, so a mask could in principle preserve the verdict while pointing at irrelevant pixels.","fun_headline_variants_meta":{"raw":{"variants":["Sparse masks explain neural net decisions without retraining","NeuroMask: Optimize a mask to show why a net predicted that","Explain any CNN by learning a sparse, smooth pixel mask","A learned mask hides pixels to reveal what a DNN actually saw"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000781,"raw_usage":{"total_tokens":3458,"prompt_tokens":962,"completion_tokens":2496,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":2425}},"tokens_in":578,"tokens_out":2496,"duration_ms":15844,"temperature":1.0,"reasoning_tokens":2425,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:10:46.443199+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: on images with ground-truth object segmentations, if a NeuroMask mask preserves the classifier's class probability while its high-weight pixels have near-zero overlap with the object, the localization claim is falsified; the same test with known spurious background cues would reveal whether the explanation tracks bias instead of the object.","supporting_citations":[{"cited_title":"On pixel-wise explanations for non-linear classifier deci- sions by layer-wise relevance propagation,","cited_arxiv_id":null,"evidence_quote":"Supplies the LRP-epsilon baseline described as too conservative in what it highlights."},{"cited_title":"Interpretable explanations of black boxes by meaningful perturbation,","cited_arxiv_id":null,"evidence_quote":"The closest prior mask-perturbation approach, used to position NeuroMask's cost-function design as its novel contribution."}],"review_version":1}