{"id":"d6ac1ab9-a535-4aad-99e4-ff06b6c3b080","arxiv_id":"1908.08997","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Aggregating gradient-based pixel attributions into superpixels approximates LIME's region rankings at a fraction of the computational cost, including for action recognition networks on video.","lead":"This paper proposes a fast way to explain CNN decisions by converting pixel-level gradient scores into superpixel importance weights, approximating LIME with a single forward-backward pass. It could make interpretability tools practical for real-time and video applications where LIME's many sampled perturbations are too slow.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Post-hoc selection of the best gradient method per network leaves the central approximation claim untested as a fixed single-pass algorithm.","rationale":"The reader's weakest assumption concerns whether LIME and the removal metrics are valid ground truth for interpretability. That is a reasonable concern, but the more immediately load-bearing issue for the paper's own stated claim is that the approximation to LIME is demonstrated only by choosing the best of nine gradient methods separately for each network. The abstract promises a technique that approximates LIME, but the experiments do not identify a single technique; they identify a family and then select the best member per architecture after seeing the comparison. This makes the headline claim under-specified: if a user wants a single-pass explainer, they are not told which variant to use, and the speed measurement in §4.3 is attached to Guided Vanilla while the accuracy numbers sometimes come from other variants. A fixed-method evaluation, ideally with a held-out architecture, would resolve whether the approximation is robust or an artifact of selection. This is a methodological weakness that supports the existing CONDITIONAL verdict rather than overturning it.","tokens_in":7807,"tokens_out":7322,"duration_ms":69579,"concrete_test":"Fix a single pipeline (e.g., Guided Vanilla pixel scores with sum-of-absolute-values aggregation, since §4.3 uses Guided Vanilla for timing) and evaluate it on all three networks, reporting top-k agreement and removal metrics with error bars. Then, without further model selection, run the same fixed pipeline on a hold-out architecture (e.g., DenseNet or Inception-v3) and compare against LIME-50 and LIME-75. If the fixed method falls below LIME-50 agreement on any network, the headline claim requires a method-selection caveat. Alternatively, split the validation set: select the best method on one half and evaluate only that method on the second half; if the selected method's agreement drops below the reported values, the reported approximation is inflated by post-hoc selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that gradient-weighted superpixels approximate LIME in a single pass—is supported by evaluating nine pixel-scoring methods and, for each network, reporting the best performer. In Table 2, the method that lands in the '50 to 75 samples' band differs by network: Act⊙Grad-CAM for VGG16 (k=5), Guided Grad-CAM for ResNet50, and Guided Vanilla for C3D. The timing comparison in §4.3 uses Guided Vanilla throughout, so the method whose speed is advertised is not the method that produced the best approximation numbers on VGG16 and ResNet50. More importantly, the best method is selected after seeing LIME agreement, so the reported approximation is an upper bound over a family of methods, not a property of a single reproducible algorithm. For a practitioner with a new network, the paper provides no rule for choosing among the nine variants without first running LIME, which defeats the stated purpose. The claim 'our proposed technique is able to approximate LIME between 50 and 75 samples' is therefore ambiguous and not falsifiable as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Gradient Weighted Superpixels: superpixels are scored by summing the absolute values of gradient-based pixel attribution maps within each segment, yielding a single-pass alternative to LIME's perturbation-based ranking. The authors evaluate nine pixel-scoring methods (vanilla gradients, guided backpropagation, input-times-gradient, Grad-CAM variants, and activation maps) on VGG16 and ResNet50 with ImageNet, and on C3D with Kinetics-400, using two quantitative protocols: superpixel removal until misclassification and top-k agreement with LIME using 1000 samples. A qualitative extension to 3D spatio-temporal superpixels for action recognition is also presented. The central claim is that the proposed technique approximates LIME with 50 to 75 samples while requiring only a single forward and backward pass, and that it does so faster on video inputs.","tokens_in":7988,"tokens_out":4295,"duration_ms":42694,"significance":"If the central claim holds, the method is practically useful because it replaces LIME's repeated forward passes with a single-pass computation, which is especially valuable for video inputs where each perturbation is expensive. The paper covers a broad set of attribution baselines, two large-scale datasets, and three architectures including a video model, which is a strength. The main weakness is that the headline claim is not yet pinned down as a property of one fixed algorithm: the best attribution method is selected per network after seeing LIME agreement, and no error bars or significance tests are reported. The contribution is therefore promising but needs a sharper evaluation protocol before it can support the stated approximation claim.","major_comments":[{"comment":"The headline claim that the proposed technique approximates LIME with 50 to 75 samples is tested by reporting, for each network, the best performing pixel-scoring method. In Table 2 at k=5, the best method is Act⊙Grad-CAM for VGG16, Guided Grad-CAM for ResNet50, and Guided Vanilla for C3D. However, the timing experiment in Figure 6 and the text of §4.3 use guided backpropagation throughout, so the method whose speed is advertised is not the method that produced the best approximation numbers on VGG16 and ResNet50. Moreover, because the best method is selected after observing LIME agreement, the reported agreement is an upper bound over a family of nine methods rather than a property of a single reproducible algorithm. To make the central claim falsifiable, the authors should either commit to a fixed scoring method across all networks or provide a principled rule for selecting the method without accessing LIME's output, and they should report timing for the method that actually achieves the claimed agreement.","section":"§4.2, Table 2 and §4.3, Figure 6"},{"comment":"All quantitative results are reported as single averages over validation sets, without standard deviations, confidence intervals, or significance tests. Differences between the proposed methods and LIME at 50 or 75 samples are often only a few percentage points; for example, in Table 2, k=1, ResNet50 Guided Grad-CAM is 48.71% versus LIME50 at 49.27%, and in Table 1 several entries differ from LIME by less than one percentage point. Without repeated runs (or uncertainty estimates) and a paired test, the claimed equivalence between the single-pass method and LIME at 50-75 samples is not established. The authors should add error bars and statistical comparisons, particularly for the VGG16 and C3D cases where the proposed method is claimed to fall between LIME50 and LIME100.","section":"Tables 1 and 2"},{"comment":"The text states 'For each model we are able to find a weighting method that beats LIME with 100 samples.' This is contradicted by the 'Best' columns of Table 1, where a lower percentage is preferable: for VGG16 the best proposed method removes 67.28% before misclassification versus 66.15% for LIME100, for ResNet50 74.35% versus 73.59%, and for C3D 51.19% versus 47.93%. If the intended claim concerns the 'Worst' columns (removal of the lowest-ranked superpixels), that should be stated explicitly. As written, the sentence is internally inconsistent with the reported data and needs correction.","section":"§4.1, Table 1 and surrounding text"},{"comment":"The top-k experiment takes LIME with 1000 samples as the reference explanation, and the superpixel-removal experiment uses misclassification as a proxy for explanation quality. The paper does not discuss the validity of either as ground truth, despite resting its central claim ('approximate LIME') entirely on these proxies. Because disagreement with LIME could mean the proposed method is wrong, LIME is wrong, or both are partially wrong, the authors should add at least a small human-subject evaluation or an alternative faithfulness metric (e.g., occlusion sensitivity or point-game-style localization against ground-truth object regions) to justify that approximating LIME is a meaningful goal. At minimum, the limitation should be acknowledged explicitly in the experimental section.","section":"§4.2 and §5"}],"minor_comments":[{"comment":"The sentence 'Other methods of distilling the pixel values into superpixel scores were investigated, however the sum of absolute values was found to be superior' gives no details of the alternatives considered or the criterion used. An ablation table comparing aggregation rules would make this design choice reproducible.","section":"§3, paragraph after the method list"},{"comment":"There is a typo: 'the most import superpixel' should be 'the most important superpixel.'","section":"§4.2, first paragraph"},{"comment":"Figure 5 uses 'LIME5000' while the text and other figures refer to LIME with 5000 samples inconsistently, and Figure 6 shows only guided backpropagation even though the surrounding text discusses 'a number of the better performing methods.' The captions should specify exactly which methods and sample counts are displayed.","section":"Figure 5 and Figure 6 captions"},{"comment":"The superpixel generation methods (QuickShift for images, SLIC for video) are named, but their parameter settings (number of segments, compactness, and so on) are not given. Since all comparisons depend on the segmentation, these parameters should be reported.","section":"§4, experimental setup"},{"comment":"The sentence in §2, 'This work also introduced gradient ⊙ input as a visualisation method,' is attached to reference [12], which is the DeepLIFT paper; the gradient-times-input visualization is more commonly attributed to the Integrated Gradients paper [16]. The citation should be corrected or re-worded.","section":"References and related work"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the core idea has practical value. The main concern is that the central claim is expressed as a property of a single fast algorithm, but the experiments report the best of nine variants chosen per network, which makes the claim an upper bound rather than a fixed-method result. This is fixable by re-running the evaluation with a committed method or by providing a selection rule that does not use LIME. I did not identify any evidence of misconduct."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a reasonable engineering paper with an honest but overreaching claim. The idea—generate superpixels, aggregate gradient-based pixel scores by sum of absolute values, compare to LIME—is simple and likely useful. The video extension to C3D with spatiotemporal superpixels is the most original part. The paper also does something right: it reports LIME with 50, 75, 100, 500, and 1000 samples as a baseline curve, so you can see where gradient methods fall on the cost/quality trade-off.\n\nThe soft spots are real, and the stress-test note nails the main one. The headline claim is that gradient-weighted superpixels approximate LIME in a single pass. But the approximation results are obtained by picking the best of nine pixel-scoring methods separately for each network. For VGG16 the star is Act⊙Grad-CAM; for ResNet50 it's Guided Grad-CAM; for C3D it's Guided Vanilla. The timing experiment, meanwhile, uses Guided Vanilla throughout. So the speed being advertised belongs to a different method than the one that produced the best accuracy on two of three networks. And there is no rule for choosing among the nine variants on a new network without first running LIME—which defeats the stated purpose. As written, “approximates LIME between 50 and 75 samples” is an upper bound over a method family, not a property of a fixed algorithm. That is a load-bearing ambiguity, though not a fatal one: they could fix it by preselecting one method based on validation data or, better, by reporting all nine honestly and claiming the concept works with at least one gradient scorer.\n\nThe evaluation has additional issues. No error bars or significance tests on Tables 1 and 2, and several differences between methods are small, so the per-network “best” may be noise. The metrics assume LIME 1000 is ground truth and that removal-to-misclassification measures explanation quality; both are contested in the XAI literature. The paper says sum-of-absolute-values aggregation “was found to be superior” but shows no comparison. No code either.\n\nNone of this kills the paper. As a cheap approximation to LIME, especially for video where 1000 perturbations is genuinely painful, the idea has clear practical value. The C3D timing numbers (0.13s vs 1.0–1.5s for LIME 50–75) are striking. The paper deserves a serious referee. I would ask for a fixed method choice or a disambiguated claim, error bars, and ideally code or a human-subject sanity check. If I worked on efficient interpretability, I would cite it; as is, I would want to see the revision first.","headline":"Useful idea, overreaching claim: gradient-weighted superpixels can roughly match LIME at a fraction of the cost, but the headline result is weakened by post-hoc method selection and weak evaluation.","tokens_in":8506,"tokens_out":3688,"would_cite":false,"duration_ms":35157,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A proposed method approximates LIME's superpixel explanations by summing absolute gradient-based pixel scores within each superpixel, matching LIME run with 50-75 samples while needing only one forward and backward pass through the network.","keywords":["interpretability","explainable AI","superpixels","gradient-based saliency","LIME approximation","action recognition","convolutional neural networks","video explanation"],"falsifier":"Take a fixed set of images, compute the gradient-weighted superpixel ranking and the LIME-1000 ranking for each image, and measure per-image rank correlation such as Kendall's tau; if the correlation is no better than LIME with 50 samples, or if it drops sharply when the number of superpixels changes, the claim that the method approximates LIME is falsified. A human study in which LIME-1000 explanations are rated as clearly more useful than the one-pass explanations would also undercut the practical claim.","tokens_in":7600,"feed_emoji":"⚡","tokens_out":4995,"duration_ms":46899,"temperature":0.7,"pith_summary":"This paper tries to show that the expensive, perturbation-based explanation method LIME can be approximated by a cheap one-pass procedure: generate superpixels, score each pixel with a gradient-based saliency method, and sum the absolute scores inside each superpixel. The paper argues that this summed-gradient weighting ranks superpixels much like LIME does, with quality comparable to running LIME with 50 to 75 perturbed samples, while requiring only a single forward and backward pass through the network. The authors test the idea on ImageNet classification with VGG16 and ResNet50, and on Kinetics-400 action recognition with C3D, comparing against LIME and random ranking using superpixel-removal and top-k agreement metrics. The payoff, if true, is an interpretability tool that keeps LIME's readable superpixel format but is fast enough for video and near-real-time use.","feed_headline":"Gradient superpixels approximate LIME in a single pass","feed_subtitle":"One-pass summed-gradient weighting matches LIME's ranking at 50-75 samples and explains video frames in 0.13 seconds.","key_machinery":"The central object is superpixel-weighted saliency: first segment the input into superpixels using QuickShift for images or SLIC for video volumes, then compute a per-pixel attribution map with one backward pass from the class of interest, and finally assign each superpixel a weight equal to the sum of absolute pixel scores within it. This replaces LIME's repeated perturbation of the input and its surrogate linear model with a single deterministic aggregation, which is the mechanism that carries the speedup.","core_discovery":"The paper's central claim is that superpixels can be weighted by aggregating per-pixel attribution scores from standard gradient-based visualization techniques, and that this aggregation approximates LIME's superpixel rankings. For each superpixel, the weight is the sum of the absolute values of the pixel scores inside it; the scores come from methods such as vanilla backpropagation, guided backpropagation, input-gradient products, Grad-CAM, Grad-CAM++, and guided combinations. Quantitatively, on top-k agreement with LIME run with 1000 samples, several of these one-pass variants perform between LIME with 50 samples and LIME with 75 samples, and in superpixel-removal tests they are consistently better than random ranking and sometimes beat LIME with 100 samples. For the C3D action recognition network, the method explains a 16-frame temporal volume in 0.13 seconds on average, compared with 1.0 and 1.5 seconds for LIME with 50 and 75 samples, so the paper positions the technique as a fast alternative that keeps LIME's interpretable superpixel output.","pith_inferences":["A natural extension, not explored in the paper, is to test whether a different aggregation rule, such as the maximum or mean of absolute pixel scores, would push agreement with LIME-1000 higher than the sum does.","The one-pass ranking could be used as a cheap pre-filter that selects candidate superpixels for a subsequent LIME run, potentially reducing the number of perturbations LIME needs while keeping its exact behavior.","The paper's evaluation on C3D, an older action recognition model, leaves open whether the method transfers cleanly to modern 3D architectures; the single-pass property suggests it should, but temporal resolution issues could reappear.","If the top-k agreement metric is a valid proxy for explanation quality, then this style of evaluation could become a general benchmark for any fast explainer claiming to approximate a slower reference method."],"forward_implications":["For image classifiers, the one-pass method can produce LIME-like explanations at near-interactive speeds, since ranking quality matches LIME at 50 to 75 samples while costing only one forward and backward pass.","The same recipe extends to 3D spatiotemporal inputs, allowing action recognition networks to be explained in well under a second per 16-frame clip.","Because explanations are generated by backpropagating from a chosen class, the method can highlight different classes in the same image or video, not just the predicted class.","The superpixel-removal results suggest the method is particularly good at identifying low-importance regions, which is useful for debugging when a network relies on unexpected or spurious evidence.","If LIME's ranking is taken as the reference, the method offers a practical speed-quality trade-off point for any deployment where LIME's default 1000 samples is too slow."],"supporting_citations":[{"why":"LIME is the baseline the paper aims to approximate, providing the superpixel-ranking reference and the perturbation-based method being replaced.","marker":"[10]"},{"why":"Vanilla backpropagation supplies one of the pixel-scoring techniques whose absolute values are summed inside superpixels.","marker":"[14]"},{"why":"Guided backpropagation supplies another pixel-scoring technique and is reported as the most consistent weighting method across networks.","marker":"[15]"},{"why":"Grad-CAM provides the weighted activation-map scoring technique used to weight superpixels.","marker":"[11]"},{"why":"Grad-CAM++ is one of the gradient-based pixel scoring variants compared for superpixel weighting.","marker":"[4]"},{"why":"QuickShift generates the superpixels used for image classification experiments.","marker":"[19]"},{"why":"SLIC generates the spatiotemporal segments used for the action recognition experiments.","marker":"[1]"},{"why":"C3D is the 3D convolutional network used to test the method on video inputs.","marker":"[17]"},{"why":"The released C3D weights for Kinetics-400 come from the R(2+1)D work and are used to initialize the action recognition model.","marker":"[18]"}],"fun_headline_variants":["One-pass gradient sums give LIME-like superpixel rankings","Gradient-weighted superpixels explain video frames in 0.13s","Single-pass gradient superpixels rival LIME at 50–75 samples","Gradient superpixel scores approximate LIME in a single pass","Gradient superpixels: LIME-level rankings at video speed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that LIME's superpixel rankings, computed with 1000 samples, are the correct reference for explanation quality; if LIME can be unreliable or if the top-k and removal proxies do not track what humans find interpretable, then matching LIME at 50 to 75 samples does not by itself establish that the one-pass method is truly interpretable.","fun_headline_variants_meta":{"raw":{"variants":["One-pass gradient sums give LIME-like superpixel rankings","Gradient-weighted superpixels explain video frames in 0.13s","Single-pass gradient superpixels rival LIME at 50–75 samples","Gradient superpixel scores approximate LIME in a single pass","Gradient superpixels: LIME-level rankings at video speed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001225,"raw_usage":{"total_tokens":5015,"prompt_tokens":907,"completion_tokens":4108,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":4014}},"tokens_in":523,"tokens_out":4108,"duration_ms":29196,"temperature":1.0,"reasoning_tokens":4014,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:59:16.694366+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed set of images, compute the gradient-weighted superpixel ranking and the LIME-1000 ranking for each image, and measure per-image rank correlation such as Kendall's tau; if the correlation is no better than LIME with 50 samples, or if it drops sharply when the number of superpixels changes, the claim that the method approximates LIME is falsified. A human study in which LIME-1000 explanations are rated as clearly more useful than the one-pass explanations would also undercut the practical claim.","supporting_citations":[{"cited_title":"Why should I trust you?","cited_arxiv_id":null,"evidence_quote":"LIME is the baseline the paper aims to approximate, providing the superpixel-ranking reference and the perturbation-based method being replaced."},{"cited_title":"Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra","cited_arxiv_id":null,"evidence_quote":"Grad-CAM provides the weighted activation-map scoring technique used to weight superpixels."},{"cited_title":"Quick shift and kernel methods for mode seeking","cited_arxiv_id":null,"evidence_quote":"QuickShift generates the superpixels used for image classification experiments."},{"cited_title":"A closer look at spatiotemporal convolutions for action recognition","cited_arxiv_id":null,"evidence_quote":"The released C3D weights for Kinetics-400 come from the R(2+1)D work and are used to initialize the action recognition model."}],"review_version":1}