{"id":"3dab4b30-e83a-4f6c-a1ed-c575d0b576f5","arxiv_id":"2501.06841","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"FCVE uses a decoder to turn modified convolutional filters into visual counterfactual explanations for MNIST and Fashion-MNIST classifiers.","lead":"An XAI method called FCVE generates counterfactual images by altering internal filters of a trained CNN and decoding the modified activations back into an image. The authors claim the result is faithful to the model's reasoning, but they do not measure faithfulness directly and their images are farther from the original than some baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The decoder is trained on intact activations (Eq. 3) but invoked on CFE-modified activations (Eq. 8), so the generated counterfactuals may be decoder artifacts; no validity or inversion-fidelity check supports the faithfulness claim.","rationale":"The reader's weakest assumption identifies exactly the gap I consider most load-bearing: the decoder is trained on one distribution and used on another. The paper provides no evidence that the decoder generalizes to modified activations, and without such evidence the generated images cannot be claimed to reflect the model's internal reasoning. My proposed check would settle this by directly measuring both whether the generated counterfactual is actually classified as the target class and whether the decoder faithfully reconstructs the modified feature vector. The qualitative figures alone are not sufficient, especially since Table 1 shows proximity scores worse than all baselines, contradicting the 'minimum changes' claim, and the reported FID values (0.50 on MNIST, 2.02 on FMNIST) are strikingly low and reported without error bars or sufficient experimental detail. These issues reinforce the reader's rejection, so I do not recommend changing the verdict.","tokens_in":13662,"tokens_out":2903,"duration_ms":31409,"concrete_test":"On a held-out set from MNIST and FMNIST, for each query x and target class c_hat, compute the modified feature vector a* = C_conv(x, F_MC_i, F_MI_i), generate x' = D(a*), and then measure: (1) counterfactual validity, the fraction of generated x' that the classifier assigns to c_hat; (2) decoder inversion error on modified features, ||C_conv(x') - a*||_2, compared with the same error on intact training features. If validity is not near 100% or if modified-feature inversion error is substantially larger than intact-feature error, the decoder does not faithfully invert altered activations and the visual explanations are not faithful to the model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that FCVE images faithfully reflect the classifier's internal decision-making. The load-bearing step is the decoder's out-of-distribution generalization. In Eq. (3), the decoder D is trained only on intact feature vectors C_conv(x_i), minimizing MAE to reconstruct x_i. In Eq. (8), D is then applied to altered features C(x, F_MC_i, F_MI_i) produced by the CFE filters. Those filters are specifically optimized to change the classifier's decision, so the modified feature vectors move away from the training distribution. Nothing in Section 3.2, Algorithm 1, or the loss constrains D on such modified features. If D has not learned to invert modified feature vectors, the output images are plausible-looking decoder artifacts rather than visualizations of the classifier's reasoning. The paper's referral to Tariq et al. (2022) validates the identified filters, not the rendered images; it does not establish that decoding a modified vector shows what the filter change means. Without a counterfactual validity check or an inversion-fidelity comparison, the abstract's faithfulness claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FCVE, a post-hoc visual explanation method that generates counterfactual images by modifying filter activations in a pretrained CNN's last convolutional layer rather than altering input pixels. The method combines a previously developed counterfactual/contrastive filter identification model (Tariq et al. 2022) with a decoder trained to reconstruct input images from intact feature vectors; during explanation, the decoder renders images from filter-modified activations as visual counterfactual explanations. The authors claim these explanations are both plausible and faithful to the classifier's internal decision process, and report qualitative comparisons on MNIST and Fashion-MNIST along with quantitative proximity and FID scores against several baselines.","tokens_in":13752,"tokens_out":3823,"duration_ms":37565,"significance":"If the central claims were established, the paper would make a useful contribution by offering a visualization of internal filter-level changes rather than pixel perturbations, which could make counterfactual explanations more interpretable. The qualitative examples on MNIST and FMNIST are suggestive, and the idea of decoding modified feature activations is a reasonable direction. However, the paper does not actually evaluate faithfulness, its quantitative proximity results contradict the 'minimum changes' claim, and the FID computation appears incorrect. As it stands, the evidence does not support the paper's main assertions, so the contribution remains a promising but unvalidated idea rather than a demonstrated method.","major_comments":[{"comment":"The proximity scores reported for FCVE (0.098 on MNIST, 0.198 on FMNIST) are the highest among all methods in Table 1; for comparison, CEM achieves 0.016 and 0.013 on the same datasets. The paper defines proximity as an L1 distance where lower values indicate counterfactuals closer to the original input. These numbers directly contradict the central claim, stated in the Abstract and Section 5, that the method makes 'minimum changes necessary' to the input. The authors need to reconcile this discrepancy or substantively revise the minimality claim, since the 'minimum change' property is a core selling point of the paper.","section":"Section 4.3.1, Table 1"},{"comment":"Equation (10) is not the Fréchet Inception Distance. The standard FID is ||mu - mu'||^2 + Tr(Σ + Σ' - 2(ΣΣ')^{1/2}) where mu and Σ are the mean and covariance of feature distributions; the paper instead writes Tr(x + x' - 2*sqrt(x.x')) without defining x and x' as covariance matrices, and the surrounding text describes the metric as 'feature-wise subtraction of the query images and their respective counterfactuals,' which is a per-pair distance, not a distributional metric. Furthermore, the reported FID values (0.50 for MNIST, 2.02 for FMNIST) are strikingly low and given without error bars or any description of how many samples were used. The quantitative claim of outperforming baselines is therefore not supported.","section":"Section 4.3.2, Eq. (10)"},{"comment":"Faithfulness, the property that appears in the title and the abstract, is not directly evaluated for the proposed visual explanations. The text states that the authors 'refer the reader to Tariq et al. [2022]' for evidence that the identified filters are faithful. That prior work validates the filters themselves, not the decoded images produced by the decoder used here. The present paper's claim is that the generated counterfactual images 'reflect the internal decision-making process of the model,' but no experiment checks whether a generated image x' actually flips the classifier's prediction or whether the decoded visual changes correspond to the internal filter modifications. Without such a direct evaluation, the faithfulness claim is inherited from a prior work rather than demonstrated for the proposed method.","section":"Section 4, first paragraph"},{"comment":"The decoder D is trained on intact feature vectors C_conv(x_i) but is applied at test time to modified feature vectors C(x, F_MC_i, F_MI_i) produced by the CFE filter changes. The paper provides no evidence that D generalizes to these out-of-distribution activations; if it does not, the generated counterfactuals could be artifacts of the decoder rather than visualizations of the classifier's reasoning. The paper should include an inversion-fidelity check on modified activations (for example, measuring reconstruction error on feature vectors perturbed to the same degree as those used in Eq. (8)) or verify that the decoded image is classified into the intended target class. Since the entire visual explanation rests on this step, this issue is load-bearing.","section":"Section 3.2, Eq. (3) and Eq. (8)"}],"minor_comments":[{"comment":"The text says 'The decoder is asynchronous as the encoder and decoder have variable depths'; the intended word is likely 'asymmetric.'","section":"Section 3.2"},{"comment":"The sentence 'Proximity property explanations the counterfactuals, meaning Faithful to the original instance' is grammatically garbled and should be rewritten.","section":"Section 4.3.1"},{"comment":"The notation dsict in Eq. (9) is undefined; presumably this is an L1 distance and the notation should be clarified.","section":"Section 4.3.1, Eq. (9)"},{"comment":"The paper discusses several diffusion-based methods (e.g., Augustin et al. 2022, Jeanneret et al. 2022) in the related work but does not include them in the experimental comparison; adding such recent baselines would strengthen the evaluation.","section":"General"},{"comment":"The reference list has inconsistent formatting; for example, some entries include DOIs or volume/page numbers while others are incomplete, and the Tariq et al. 2022 entry appears to be self-cited with a DOI but no page range.","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper's principal claims are contradicted by its own quantitative results and the faithfulness evaluation is deferred to a self-cited prior work. The FID values are implausibly low and the formula appears incorrect, suggesting the reported numbers may not measure FID at all. The decoder out-of-distribution problem is a fundamental methodological gap that would require substantial new experiments and possibly a redesign of the approach to address. I do not see how these issues could be fixed within the scope of a revision; hence a reject rather than a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper combines the authors' earlier filter-level counterfactual identification with a decoder to render those filter changes as images. The combination is new, and the qualitative results are genuinely nice—7 to 9, pullover to coat, and so on. But the central faithfulness claim is not tested in this paper, and the quantitative evidence that is reported contradicts the “minimum changes” promise.\n\nThe new part is a decoder applied to modified filter activations. As far as the cited literature goes, nobody has visualized filter-level counterfactuals this way. Leaning on the prior filter-identification work for filter-level faithfulness is legitimate self-citation, not a defect.\n\nThe soft spots are load-bearing. The decoder is trained only on intact feature vectors (Eq. 3) and then invoked on modified features (Eq. 8). Nothing constrains the decoder to behave on those out-of-distribution inputs, and there is no check that the reconstructed images are actually classified as the target class. The paper explicitly defers faithfulness to Tariq et al. 2022, which validates the filters, not the rendered images. So the visuals may be decoder artifacts rather than a window into the classifier.\n\nThe quantitative section compounds the problem. Proximity scores put FCVE last among the five baselines, directly undermining the “minimum changes” claim. The FID values (0.50 and 2.02) are implausibly low, no error bars or run-level variance are reported, and Eq. 10 is not the FID formula—it's a garbled mix of means, covariances, and a spurious square root. No code, no decoder architecture detail, no counterfactual validity metric. These are fixable, but they are not present.\n\nWho is this for? A reader working on counterfactual visual explanations might use this as a motivation to do the evaluation properly. As a validated method, it's not there yet.\n\nI'd send it to peer review, because the idea is worth referee time and the problems are specific and addressable. But I'd expect a serious referee to require substantial revisions before anything in the paper is relied on.","headline":"New decoder-based filter visualization with nice qualitative results, but faithfulness is untested and the reported metrics contradict the paper's own claims.","tokens_in":14397,"tokens_out":4255,"would_cite":false,"duration_ms":41559,"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 claims that counterfactual visual explanations can be generated by changing a CNN's internal filter activations and decoding them, so the resulting image shows what the model's own learned features would need to change to flip…","keywords":["counterfactual explanations","visual explanations","explainable AI","faithfulness","plausibility","CNN filters","feature visualization","image classifiers"],"falsifier":"Take the counterfactual image $x'$ produced by FCVE for a source image $x$ and a target class, run the original classifier on $x'$, and record its predicted label. If this label is not the target class, or if the prediction fails to stay in the source class when only $F_{MC}$ is applied, the explanation is not faithful to the model; this classifier-on-counterfactual accuracy is not reported in the paper.","tokens_in":13341,"feed_emoji":"🖼️","tokens_out":10714,"duration_ms":99837,"temperature":0.7,"pith_summary":"This paper sets out to make counterfactual explanations of image classifiers both plausible and faithful by moving the explanation out of pixel space. Instead of searching for pixels to perturb, the proposed FCVE model locates the classifier's own high-level filters: which last-layer activations must stay on to keep a prediction, and which must be boosted to flip it toward a target class. A decoder is first trained to reconstruct the input from the unmodified activation vector, and is then applied to the vector after those filter changes, producing an image that shows the model's internal features changing with its decision. If this works as claimed, a user could ask what would make this image classified differently and receive a visual answer generated from inside the network, not from a pixel-level perturbation search.","feed_headline":"Swapping a network's internal filters shows what changes its verdict","feed_subtitle":"Minimum changes to high-level filters render 'what-if' images that mirror the model's own reasoning.","key_machinery":"The load-bearing objects are the two counterfactual-filter masks and the decoder that visualizes them. $F_{MC}$, called the minimum-correct mask, is an approximately binary vector marking which of the $n$ filters in the top convolutional layer must remain active for the classifier to keep its original prediction. $F_{MI}$, the minimum-incorrect mask, is a vector of positive magnitudes that specifies how much each filter activation should be boosted or altered to move the prediction toward a target class. The decoder, trained once with the classifier frozen, learns the inverse map from the last convolutional feature vector back to the input image; applying the masks to that feature vector and decoding produces the counterfactual image. The masks supply what to change, and the decoder supplies how that change looks.","core_discovery":"On its own terms, the central claim is that a CNN decision can be faithfully visualized without touching pixel values. The method takes a previously trained counterfactual-filter model that outputs two masks over the $n$ filters of the last convolutional layer: a minimum-correct mask $F_{MC} \\in [0,1]^n$, which keeps the original class active, and a minimum-incorrect mask $F_{MI} \\in \\mathbb{R}_+^n$, which gives the magnitudes by which filter activations must be altered to move the prediction to a chosen target class. A decoder $D$ is trained with mean-absolute-error loss to invert the intact feature vector $C_{\\mathrm{conv}}(x)$, and the counterfactual is defined as $x' = D(C(x, F_{MC}, F_{MI}))$, where $C(x, F_{MC}, F_{MI})$ denotes the classifier's feature vector after the masks are applied. The paper argues that because $x'$ is produced by changing internal activations rather than by editing pixels, the visual explanation reflects the model's own decision process, and because the filter changes are minimal, the counterfactual stays close to the original image.","pith_inferences":["An implicit test the paper leaves open is decision-level validity: feeding each generated $x'$ back into the classifier and measuring how often the label actually becomes the target class would directly separate faithful counterfactuals from visually plausible artifacts.","Interpolating the $F_{MI}$ boost from zero to its full value would trace a visual trajectory from the source class to the target class, offering a concept-level analogue of latent-space interpolation that the paper does not explore.","The decoder's reliability could be audited by constructing modified activations for images whose class is known, for example zeroing the predicted filters of a digit and decoding, and checking that the reconstructions degrade only in the class-relevant regions.","The very low FID scores reported relative to pixel-based baselines may partly reflect that the decoder outputs are reconstructions from the training distribution; a comparison against classification accuracy on the generated images would disentangle reconstruction quality from counterfactual validity."],"forward_implications":["Once the decoder is trained, counterfactual images are produced in a single forward pass, with no per-example optimization loop over pixels.","The explanation is anchored in the network's own filter vocabulary, so it can show which learned concepts, not which pixels, the model treats as decisive.","Because only the minimal correct and incorrect filter masks are edited, the method aims to satisfy the counterfactual proximity property by construction.","The same filter-plus-decoder recipe is tied to any convolutional classifier with an accessible last-layer feature vector, independent of the image dataset."],"supporting_citations":[{"why":"Supplies the counterfactual-filter CFE model whose $F_{MC}$ and $F_{MI}$ masks the proposed method visualizes; FCVE is built directly on it.","marker":"Tariq et al. [2022]"},{"why":"Provides the MNIST benchmark used to train the decoder and evaluate counterfactuals.","marker":"LeCun [1998]"},{"why":"Provides Fashion-MNIST, the second dataset used for evaluation.","marker":"Xiao et al. [2017]"},{"why":"ExpGAN is a pixel-based counterfactual baseline compared in the visual and quantitative results.","marker":"Samangouei et al. [2018]"},{"why":"CEM is the contrastive-explanation baseline compared on proximity and FID.","marker":"Dhurandhar et al. [2018]"},{"why":"CVE is the standard counterfactual visual explanation baseline that the paper compares against.","marker":"Goyal et al. [2019]"},{"why":"C3LT is the cycle-consistency latent-transformation baseline compared in the experiments.","marker":"Khorram and Fuxin [2022]"},{"why":"Supports the premise that top-layer filters capture high-level concepts, justifying filter-level rather than pixel-level edits.","marker":"Bau et al. [2020]"}],"fun_headline_variants":["Filter swaps produce true what-if images","Counterfactual visuals from CNN's own filters","No pixel edits: internal concepts explain verdicts","Faithful counterfactual explanations via minimal filter changes","See CNN reasoning: minimal filter tweaks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire visual explanation rests on the decoder being able to render feature vectors it never saw during training: it is trained only on intact activations, but is then used on activations in which filters have been zeroed or boosted, and the paper does not establish that these modified vectors stay inside the decoder's reliable reconstruction region.","fun_headline_variants_meta":{"raw":{"variants":["Filter swaps produce true what-if images","Counterfactual visuals from CNN's own filters","No pixel edits: internal concepts explain verdicts","Faithful counterfactual explanations via minimal filter changes","See CNN reasoning: minimal filter tweaks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1406,"prompt_tokens":915,"completion_tokens":491,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":422}},"tokens_in":531,"tokens_out":491,"duration_ms":5193,"temperature":1.0,"reasoning_tokens":422,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:49:43.091861+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the counterfactual image $x'$ produced by FCVE for a source image $x$ and a target class, run the original classifier on $x'$, and record its predicted label. If this label is not the target class, or if the prediction fails to stay in the source class when only $F_{MC}$ is applied, the explanation is not faithful to the model; this classifier-on-counterfactual accuracy is not reported in the paper.","supporting_citations":[],"review_version":1}