{"id":"c66744e2-4d95-47e9-a7f2-429f799251d6","arxiv_id":"2501.06831","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper learns sparse masks over a DCNN's top-layer filters that preserve the inferred class (contrastive) or flip it to an alter class (counterfactual), evaluated on CUB bird classification.","lead":"An explainability method identifies which internal filters of an image classifier keep a prediction unchanged and which filters, if boosted, would switch the prediction to another class. The method probes the network's internal features instead of altering input pixels, and could help debug misclassifications and build trust in high-risk applications.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 'minimum' filter sets are outputs of a learned per-class network, not per-image optima; L1 sparsity does not certify minimality, so the central minimality/transparency claim is unproven.","rationale":"The reader's weakest_assumption correctly identifies that L1 sparsity does not prove minimality. I sharpen this: the masks are not per-image optima at all, but outputs of a learned network trained on an average loss, so the 'minimum' phrase in the paper's objectives and Algorithm 1/2 descriptions is not justified even in principle. This is a load-bearing concern because the paper's transparency and misclassification analyses are built on interpreting the selected filters as the critical decision-making features. If the masks are merely sparse approximations that may vary in size and correctness across images, then claims like 'using just these filters, the model still classifies the input to its original inferred class' (Section 4.2.1) are only anecdotal and not validated by any systematic per-image check. The proposed concrete test would settle the concern directly by comparing predicted masks to true optima. The reader's verdict of CONDITIONAL already reflects this uncertainty, so the verdict should remain UNCHANGED. I do not see an internal inconsistency that would warrant REJECT; the method is a plausible sparse-attribution tool, but its headline minimality and transparency claims require the proposed verification.","tokens_in":21712,"tokens_out":5557,"duration_ms":67727,"concrete_test":"For a fixed class c and a fixed set of test images correctly classified as c, compute the true minimal MC mask per image by exhaustive search over the 512 top-layer filters (or by orthogonal matching pursuit / integer programming) using the frozen linear classification head h(g) = softmax(W^T g + b). Compare (i) the cardinality of the CFE-predicted mask to the true minimum, (ii) whether the CFE-predicted mask actually preserves the class on every image, and (iii) whether any proper subset of the CFE-predicted mask also preserves the class. Repeat for MI masks by solving min ||delta||_1 subject to argmax h(g+delta) = c' and comparing the CFE-predicted addition to the optimal delta. If the CFE masks are not minimal or are not even sufficient on a non-negligible fraction of images, the 'minimum' claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the MC and MI masks be the minimum sets of top-layer filters that preserve or flip the model's decision for a given input. But the masks are not obtained by solving the per-image optimization in Eqs. (3) and (9); they are the outputs of a trained single-layer network, CFE(g) = ReLU(dn(g)) or ReLUt(sigmoid(dn(g))), evaluated at inference. Training minimizes an average loss over a class's training images, so the learned mapping may produce masks that are sparse on average but are neither minimal nor even sufficient for a particular test image. Table 3 confirms this: for the MC model at the reported lambda=2, the testing accuracy is 99.1%, meaning about 1% of test images are not preserved by the predicted mask. The paper never compares the predicted masks to the true per-image minimum, never reports the distribution of mask sizes, and never checks whether a strictly smaller subset of the predicted filters would still suffice. L1 regularization is a convex relaxation that selects one sparse solution; it provides no certificate of minimality. Because the paper's objectives (Section 1) are phrased as 'minimum set' and the transparency argument rests on those filters being exactly the necessary and sufficient ones, the core claim is not established by the evidence as presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a post-hoc explainability method for DCNN image classifiers, applied to VGG-16 on CUB-2011. Given an input image, a per-class 'counterfactual explanation' (CFE) network predicts two sets of masks over the top convolutional layer's post-GAP filter activations: a binary 'minimum correct' (MC) mask that, when multiplied with the activations, is claimed to be the minimum set of filters sufficient to preserve the model's original decision, and a nonnegative 'minimum incorrect' (MI) additive perturbation that, when added to the activations, is claimed to flip the decision to a specified alter class. The masks are trained by minimizing cross-entropy, an L1 sparsity penalty, and (for MC) a negative logits loss. The paper evaluates the masks through qualitative visualizations, a misclassification case study, a user study, a class-recall ablation, and a comparison with GradCAM and SCOUT.","tokens_in":21921,"tokens_out":5019,"duration_ms":52664,"significance":"If the claims were fully established, the method would be a useful addition to the explainability toolbox: it is model-intrusive, produces compact filter-level explanations, and the misclassification analysis in Section 4.2.2 is a compelling demonstration that filter-level reasoning can diagnose an erroneous decision. The paper also ships a user study and an attempt at quantitative comparison, which is more than many XAI papers provide. However, the central 'minimum set' claim is currently a fitted property of a learned surrogate, not a verified property of the explained model, and the quantitative comparison is underspecified. The significance of the contribution is therefore conditional on reframing and additional validation.","major_comments":[{"comment":"The central claim that the method identifies the 'minimum set' of filters is not established. The masks are outputs of a per-class dense network trained on the class-averaged loss, not solutions of the per-image optimization written in Eqs. (3) and (9). L1 regularization selects one sparse solution and provides no certificate that no smaller or alternative mask exists. Table 3 itself shows that at the reported lambda=2 the testing accuracy is 99.1%, meaning that for about 1% of test images the predicted MC mask does not even preserve the decision, which contradicts the 'necessary' interpretation. Please either reframe the claims as 'sparse sufficient masks' or provide per-image minimality evidence, for example by comparing predicted masks against greedy or exhaustive search on a sample of images, reporting the distribution of mask sizes, and testing whether removing any individual predicted filter breaks the prediction.","section":"Section 1 and Section 3.1-3.2, Eqs. (3) and (9)"},{"comment":"The quantitative comparison with GradCAM and SCOUT is under-specified for the proposed filter-based method. The recall and precision metrics are defined over an 'explanation region', but the paper never states how a binary MC filter mask or an MI perturbation is converted into a spatial explanation region on the input image. Without this mapping, the numbers in Table 5 are not reproducible and the comparison is not meaningful. Please specify the spatial aggregation rule, describe how part annotations are matched to the region, and ideally release code for the evaluation.","section":"Section 4.5, Table 5"},{"comment":"The paper treats scalar modification of post-GAP filter activations as equivalent to enabling or disabling visual concepts in the input. This assumption is not validated: a GAP value is an average over spatial locations, and zeroing or adding to that scalar does not remove or insert the corresponding visual feature in the image or in the feature maps. The manual image edits in Section 4.2.1 (Fig. 4) and Section 4.2.2 (Fig. 7) are a different intervention and do not test the scalar-level operation. As a result, the claims that the masks correspond to 'features and concepts' and that the model is made 'transparent' rest on an unverified causal linkage. Please add experiments that compare scalar-level interventions with spatial feature-map interventions or input-level perturbations on the same images.","section":"Section 3.1, Eq. (2), and Section 3.2, Eq. (8)"},{"comment":"The sparsity-loss analysis does not support the 'minimum' conclusion. The table reports average filter counts and accuracy values as a function of lambda, but it never checks whether a strictly smaller subset of the predicted filters would still suffice for a given image, nor whether a different sparse subset would be equally effective. Thus the trade-off shown in Table 3 is between the surrogate's training objective and its output size, not a demonstration that the CFE model finds the minimal sufficient set. Please include a minimality check or weaken the conclusion accordingly.","section":"Section 4.4.2, Table 3"}],"minor_comments":[{"comment":"The cross-entropy loss is written as if ci were a probability vector, but ci is defined earlier as a class index. Please use one-hot notation or a sparse softmax cross-entropy formulation to avoid ambiguity.","section":"Section 3.1, Eq. (4)"},{"comment":"The 'Rand. disabled' column appears to report a single random draw of filters; please state how many random subsets were used and report the variance, since the conclusion that random disabling has little effect depends on that.","section":"Table 2"},{"comment":"The 'Filters' columns report averages without standard deviations or percentiles; given that the masks vary per image, a distribution would be more informative than a point average.","section":"Section 4.4.2, Tables 3 and 4"},{"comment":"The reported VGG-16 test accuracy of 69.5% on CUB-2011 is not contextualized against standard baselines; adding a reference to prior reported accuracies on this benchmark would help the reader judge the explained model.","section":"Section 4.1"},{"comment":"The paper claims 'consistently better' recall/precision than SCOUT, but Table 5 reports only a few classes and no significance testing; please state how many images and class pairs were used and whether the differences are statistically reliable.","section":"Section 4.5"}],"recommendation":"major_revision","confidential_remarks":"The paper would be materially stronger if the authors either performed a genuine minimality check (greedy search over filter subsets on a sample) or explicitly reframed the contribution as 'sparse sufficient mask prediction' rather than 'minimum set identification.' The comparison protocol also needs to be made concrete before the numbers in Table 5 can be trusted. I see no reason to doubt the qualitative demonstrations, but the current manuscript does not yet substantiate its headline claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper does something genuinely new: it trains a small network on top of the post-GAP activations of a pre-trained CNN to output two kinds of filter masks—a binary contrastive mask (MC) that preserves the inferred class when multiplied, and a non-negative counterfactual mask (MI) that, when added, flips the decision to a chosen alter class. That combination—learned sparse mask plus learned additive perturbation, both over top-layer filters—isn't in the cited literature, and the misclassification case study (the Myrtle warbler / red-winged blackbird example) is a nice demonstration of how such filter-level reasoning can expose background or spurious features.\n\nWhat the paper does well: the architecture is simple and clearly described, the training losses (CE + L1 + logits for MC, CE + L1 for MI) are reasonable, and the authors are honest about the weakness of the synthetic ground truth used in the quantitative comparison. The qualitative visualizations with receptive fields are useful for building intuition.\n\nThe soft spots are real, though. The central claim—that these are the *minimum* sets—is not established. L1 regularization gives one sparse solution, not a certificate of minimality, and more importantly the masks are outputs of a trained dense layer, not per-image optima. Table 3 confirms this: at λ=2, about 1% of test images are not preserved by the predicted MC mask, so the mask isn't even sufficient for every input, let alone minimal. The paper never compares predicted masks to the true per-image minimum, never reports the distribution of mask sizes, and never checks whether a strictly smaller subset would suffice. Any transparency or debugging argument that rests on the \"exactly necessary and sufficient filters\" phrasing is therefore on shaky ground.\n\nThere's also a self-fulfilling component: the masks are trained with the very class labels used to evaluate them, so the \"explanations\" are somewhat circular. That's common in this literature, but it matters more here because the minimality claim is load-bearing.\n\nThe evaluation is thin: five classes, no error bars on the quantitative metrics, a user study with 30+10 participants and no significance testing, and no code or weights released. These are addressable, but they limit independent verification.\n\nWho's it for: researchers working on filter-level or concept-level explanations, especially those interested in misclassification diagnosis. A serious referee should see it, because the core idea is novel and the pipeline is clean. My recommendation: send to peer review, but the authors should soften \"minimum\" to \"small sparse set,\" add per-image sanity checks, and release code.","headline":"A genuinely novel filter-mask counterfactual explainer for DCNNs, but the 'minimum' claim is uncertified, the evaluation is thin, and the method still deserves a serious referee.","tokens_in":22510,"tokens_out":2083,"would_cite":false,"duration_ms":20666,"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":"The paper argues that a CNN image classifier's decision can be explained by two sparse sets of top-layer filters: one that alone preserves the predicted class, and one whose boosted activations switch the prediction to a chosen…","keywords":["Explainable AI","counterfactual explanation","contrastive explanation","deep convolutional neural networks","filter-level interpretability","post-hoc explainability","misclassification analysis","CUB-2011 dataset"],"falsifier":"Take a set of test images for a class, introduce into each image the exact features that the MI filters highlight (such as the red eye or bluish wing tinge), and count how often the model's decision flips to the alter class; if flips occur no more often than for equally many randomly chosen feature edits, the MI filters are not the minimal cause of the decision change the paper claims.","tokens_in":21401,"feed_emoji":"🧠","tokens_out":8137,"duration_ms":73011,"temperature":0.7,"pith_summary":"This paper proposes a way to explain a deep convolutional image classifier by looking inside it rather than by changing its input. For any image, the method learns two sparse masks over the filters of the top convolutional layer: a minimum correct (MC) mask whose filters alone keep the model's prediction at the inferred class, and a minimum incorrect (MI) mask whose filter activations, when additively boosted, flip the prediction to a chosen alter class. The masks are trained by a small explainer head on top of the frozen convolutional features, using cross-entropy with L1 sparsity penalties, plus a negative-logits term for the MC mask. The authors argue that the selected filters correspond to interpretable, class-separating concepts and that examining them reveals why a model errs, for instance when a misclassification is driven by background branches or by a missing class-defining wing spot. If this holds, explanations of DCNN decisions can be generated at the level of the features the network actually uses.","feed_headline":"A few internal filters keep a CNN's call; a few more can flip it","feed_subtitle":"Counterfactual explanations come from inside the network: sparse masks over top-layer filter activations, not pixel edits.","key_machinery":"The machinery is a counterfactual explanation (CFE) model attached to a pre-trained DCNN at the global average pooling layer. The convolutional feature extractor is frozen and shared; a single dense layer predicts a filter mask. For MC filters the mask is binarized through sigmoid and thresholded ReLU and multiplied onto the pooled activations, $F_{MC} \\circ g$, so only the selected filters contribute to the classifier head $h$. For MI filters the mask is a nonnegative additive vector produced by ReLU and added to the pooled activations, $g + F_{MI}$. The MC mask is trained with cross-entropy, an $\\ell^1$ sparsity penalty, and a negative-logits term that favors filters with high weight toward the inferred class; the MI mask is trained with cross-entropy and an $\\ell^1$ penalty on the additive values. The selected filters are then visualized by their receptive fields to show which visual features separate the inferred class from the alter class.","core_discovery":"The central discovery is that a DCNN's decision on an image can be accounted for by a small, learnable set of filters in its top convolutional layer: those that are sufficient to reproduce the decision and those whose augmented activations are sufficient to overturn it. The paper calls the first set minimum correct (MC) and the second minimum incorrect (MI). It claims that these sets are sparse, that the filters are visually interpretable as high-level concepts, and that the contrast between them provides both contrastive (\"why this class and not that one\") and counterfactual (\"what would have to be different\") explanations that make the model transparent. It further claims that the same machinery exposes faulty reasoning in misclassifications, and reports that disabling the globally important MC filters for a class collapses class recall while leaving overall accuracy nearly unchanged.","pith_inferences":["The paper's \"minimum\" is a claim about what L1-regularized training returns, not a certificate of true minimality; a stronger test would compare the MC mask against exhaustive or random subsets of the same size.","The causal reading is strongest when the features highlighted by MI filters are introduced into real images; the paper demonstrates this on hand-edited examples, and a systematic study across many images and classes would show whether the identified filters are genuinely sufficient to flip decisions.","Because the method operates on post-pooling activations, its transparency likely degrades for lower layers, where filters are less semantic; the same architecture could be tested there to map the limits of concept-level explanations.","The MI filters point to features whose presence pushes the model across a decision boundary, which suggests a natural connection to adversarial perturbations: features that flip decisions could be checked for overlap with adversarial patterns."],"forward_implications":["If a small MC mask suffices to reproduce the original prediction, the remaining top-layer filters are not doing load-bearing work for that image, so explanations can identify where the model's decision actually lives.","The MI mask gives an actionable counterfactual: boosting the pooled activations of a handful of filters flips the decision to the alter class, meaning those filters encode features that separate the two classes.","Misclassification analysis becomes possible by contrasting MC filters for the wrong class with MI filters for the true class; in the paper's example, a background-branch filter drives a wrong prediction and a missing red wing spot explains why the true class was not chosen.","Disabling the globally important MC filters for a class drops class recall sharply (e.g., 93.3% to 30% for Red-winged blackbird) while overall accuracy falls by only about 2%, indicating these filters encode class-exclusive concepts.","User evaluations with Explanation Satisfaction scores suggest that both expert and non-expert users find the filter-based explanations understandable and useful compared with heatmap and attribute-based baselines."],"supporting_citations":[{"why":"Supplies the Caltech-UCSD Birds-2011 dataset used to train the VGG-16 model and to evaluate all explanations.","marker":"Wah et al. [2011]"},{"why":"Defines the VGG-16 architecture whose top-convolutional filters the method explains.","marker":"Simonyan and Zisserman [2014]"},{"why":"Provides the receptive-field visualization technique used to show which image regions activate the selected filters.","marker":"Zhou et al. [2014]"},{"why":"Supports the claim that top-layer filters correspond to interpretable high-level features and even whole objects.","marker":"Bau et al. [2020]"},{"why":"Grad-CAM is the main visual-explanation baseline compared in qualitative and user studies.","marker":"Selvaraju et al. [2017]"},{"why":"SCout is the counterfactual baseline and provides the recall/precision ground-truth evaluation protocol used for quantitative comparison.","marker":"Wang and Vasconcelos [2020]"},{"why":"Supplies the Explanation Satisfaction metric used in the user study.","marker":"Hoffman et al. [2018]"},{"why":"Provides the user-study protocol and the concept-level counterfactual approach the paper builds on and compares against.","marker":"Akula et al. [2020]"}],"fun_headline_variants":["A few filters in a CNN explain a decision; fewer tweaks flip it","Minimal filter sets: just enough to keep or flip a CNN call","Internal filter masks: counterfactual why and what-if for CNNs","The few filters that decide a CNN's call, and the few that flip it","Sparse masks inside a CNN give contrastive and counterfactual reasons"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that modifying a filter's post-pooling activation acts as a faithful intervention on the visual feature that filter detects, so the sparse masks reveal the concepts the model actually uses.","fun_headline_variants_meta":{"raw":{"variants":["A few filters in a CNN explain a decision; fewer tweaks flip it","Minimal filter sets: just enough to keep or flip a CNN call","Internal filter masks: counterfactual why and what-if for CNNs","The few filters that decide a CNN's call, and the few that flip it","Sparse masks inside a CNN give contrastive and counterfactual reasons"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001579,"raw_usage":{"total_tokens":6309,"prompt_tokens":962,"completion_tokens":5347,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":5248}},"tokens_in":578,"tokens_out":5347,"duration_ms":32518,"temperature":1.0,"reasoning_tokens":5248,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:50:42.237832+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of test images for a class, introduce into each image the exact features that the MI filters highlight (such as the red eye or bluish wing tinge), and count how often the model's decision flips to the alter class; if flips occur no more often than for equally many randomly chosen feature edits, the MI filters are not the minimal cause of the decision change the paper claims.","supporting_citations":[],"review_version":1}