{"id":"b6c1ea82-b63a-4069-aca2-b994d5f649e2","arxiv_id":"2412.09311","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new LRP rule that scales attributions by absolute activation magnitude, plus a unified evaluation metric, is tested across three architectures and two datasets.","lead":"This paper proposes a new layer-wise relevance propagation rule, absLRP, that divides attributions by the absolute magnitude of neuron activations, and a new evaluation metric, GAE, that combines faithfulness, robustness and contrastiveness into one score. The authors report that absLRP outperforms existing attribution methods on VGG, ResNet50 and Vision Transformers on ImageNet and PascalVOC.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (12) and Algorithm 1 disagree by a factor of 2 per layer, and the absLRP rule does not conserve relevance when pre-activations are net negative; both threaten the claim that absLRP is a principled LRP variant.","rationale":"The reader correctly identified the denominator in Eq. (12) as the least-secure element of the central claim, and the lack of a conservation or axiomatic basis is one symptom of the problem I found. My analysis sharpens this into a concrete, checkable defect: Eq. (12) does not conserve relevance for neurons with negative net pre-activation, and the pseudocode in Algorithm 1 computes twice the stated rule. If either issue is confirmed, the paper's central claim is not established for the method as defined: a non-conservative rule is not a layer-wise relevance propagation rule in the usual sense, and an implementation that differs from the stated equation means the experiments may not test the proposed method. I do not call the authors dishonest; the discrepancy may be a typo in the equation or the pseudocode, but the burden is on the paper to reconcile them and to report conservation behavior. The empirical evaluation and the GAE metric may still be valuable, and the qualitative results are suggestive, so I would not reject the paper outright. A conditional acceptance requiring (a) a corrected rule or explicit handling of negative pre-activations, (b) alignment between Eq. (12), Algorithm 1, and released code, and (c) an ablation or analysis of conservation, is the appropriate verdict. This is a partial agreement with the reader because the reader's weakest assumption was about the denominator's lack of motivation, whereas I identify an internal inconsistency and a conservation failure that are more specific and more directly testable.","tokens_in":26222,"tokens_out":6758,"duration_ms":77272,"concrete_test":"Take a single VGG16 convolutional layer and 100 ImageNet samples. Compute relevance maps by a direct implementation of Eq. (12) with R^l_j = 1 for all j, and separately by the released Algorithm 1 code. If the per-layer output of Algorithm 1 is uniformly 2 times the Eq. (12) output, the two definitions conflict. Then measure the per-layer conservation ratio Q = Σ_i R^{l-1}_i / Σ_j R^l_j under Eq. (12). If layers with many negative pre-activations give Q < 0.9 (or Q far from 1), the proposed rule is non-conservative and its status as a principled LRP variant is not established.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's formal rule, Eq. (12), is R^{l-1}_i = Σ_j (x_i w_ij)_+ / (|Σ_k x_k w_kj| + ε) R^l_j. Summing the numerators over i gives (Σ_k x_k w_kj)_+, the positive part of the net pre-activation s_j. Therefore the fraction of R^l_j actually propagated through neuron j is (s_j)_+ / |s_j|, which is strictly less than 1 whenever s_j < 0, and is 0 when s_j ≤ 0 with no positive contributions. Real ReLU networks have many such neurons, so relevance is silently destroyed layer by layer; the paper criticizes RAP's arbitrary over-allocation subtraction (Sec. 2.1) but introduces an equally arbitrary under-allocation without discussion. Conservation is not an optional extra for an LRP rule: it is what lets the propagated scores be interpreted as layer-wise relevance.\n\nSeparately, Algorithm 1 does not implement Eq. (12). It computes rel = autograd.grad(ha + h, x, rel_scaling) * x, where ha is the absolute-value linear layer. For a linear layer, the gradient of h + ha at x_i is w_ij + sign(x_i)|w_ij|, so after multiplication by rel_scaling and x_i the result is 2 Σ_j (x_i w_ij)_+ R^l_j / (|s_j| + ε), i.e. twice Eq. (12). Across L layers this introduces a 2^L factor that does not cancel between residual branches of different depths. Thus the experiments may evaluate a different rule than the one defined and motivated in the text. Either the equation, the pseudocode, or the implementation must be wrong, and the central claim of a new LRP rule is not yet supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new layer-wise relevance propagation rule, absLRP, defined in Eq. (12) as propagating relevance through a layer by normalizing each positive contribution (x_i w_ij)_+ by the absolute value of the total pre-activation |\\sum_k x_k w_kj| + epsilon. The authors argue this corrects a relative-attribution issue in LRP-\\alpha1\\beta0 when neurons differ in absolute activation magnitude, and they apply the rule to VGG, ResNet50, and ViT-Base using a contrastive output initialization. They also introduce a new evaluation metric, Global Attribution Evaluation (GAE), which combines a local-consistency score (robustness plus faithfulness via gradient-guided masking) with a contrastiveness score based on mosaics, and they report large GAE improvements for absLRP over many baselines on ImageNet and PascalVOC. The paper additionally reports ROAD, Lipschitz, and Focus scores, and performs ablations and qualitative comparisons for Transformers.","tokens_in":26641,"tokens_out":6167,"duration_ms":58798,"significance":"If the proposed rule is taken at face value, absLRP is an attractively simple and architecture-agnostic LRP variant, and the paper deserves credit for tackling Vision Transformer explainability at pixel level and for attempting a multi-component evaluation metric. The authors provide code, compare a large set of attribution methods, and include ablation studies. However, the central quantitative claim rests mainly on GAE, a metric introduced in the same paper, and the independent standard metrics give only mixed support, especially for ViT. The manuscript contains no conservation or axiomatic analysis of Eq. (12), and there is an apparent mismatch between the formal rule and the provided pseudocode. These issues currently leave the headline contribution under-supported, although the core idea is plausible and the requested fixes appear addressable.","major_comments":[{"comment":"Algorithm 1 does not implement Eq. (12). For a linear or convolutional layer, the gradient of h + ha with respect to x_i is w_ij + sign(x_i)|w_ij|, and after multiplication by x_i and rel_scaling this yields 2 (x_i w_ij)_+ R_j^l / (|s_j| + epsilon), not (x_i w_ij)_+ R_j^l / (|s_j| + epsilon). Each layer processed by Algorithm 1 therefore introduces an extra factor of 2, and across a deep network the implemented rule differs from the defined rule by a factor exponential in depth. Because all experiments, ablations, and qualitative figures are produced with the implementation, the paper must reconcile Eq. (12), Algorithm 1, and the released code; as written, the claim that absLRP is the rule in Eq. (12) is not supported.","section":"Section 3.1, Eq. (12) and Algorithm 1"},{"comment":"Eq. (12) does not conserve relevance. Summing R_i^{l-1} over i gives sum_j [ (sum_i (x_i w_ij)_+) / (|sum_k x_k w_kj| + epsilon) ] R_j^l. The numerator equals |sum_k x_k w_kj| only when all contributions x_i w_ij have the same sign; in general sum_i (x_i w_ij)_+ differs from |sum_k x_k w_kj|, so relevance is systematically amplified or suppressed layer by layer. The paper motivates LRP through the redistribution property in Eq. (1) but never states whether absLRP is intended to be conservative, and the toy example in Figure 1 involves only positive contributions and cannot reveal this behavior. The authors should prove conservation under stated conditions, add a normalization that restores it, or explicitly characterize and justify the non-conservation.","section":"Section 3.1, Eq. (12)"},{"comment":"The claim of consistent superiority is not supported by the external metrics. On ViT-Base/ImageNet (Table 9), absLRP has lower Focus (0.826) than GradCAM and HiResCAM (0.913) and worse ROAD_MoRF (0.457 vs 0.317); on ViT-Base/PascalVOC (Table 13) the same pattern holds (Focus 0.689 vs 0.800/0.785; ROAD_MoRF 0.195 vs 0.131). On VGG and ResNet absLRP is competitive, but the \"significant margin\" claimed in the abstract appears only on GAE, a metric defined by the same authors. Since GAE is also aligned with the method's design goals (sparse, contrastive, pixel-level maps), the paper should provide a sensitivity analysis of GAE to its parameters and normalization choices, and should either temper the abstract or justify why GAE should override the standard metrics where they disagree.","section":"Section 4.1, Tables 7-9 and 11-13"},{"comment":"The GAE metric is incompletely specified. Section 3.2 states that \"we select the top k percent of input features\" but neither the value of k nor the feature-block granularity (pixel-wise versus patch-wise) is given anywhere in the paper or appendices; only T=10 is reported in Section 4.1. Without k and the masking implementation, the experiments are not reproducible, and the metric's sensitivity to these choices is unknown. Given that GAE carries the paper's main quantitative conclusion, this omission blocks verification of the primary result.","section":"Section 3.2, Local consistency definition"}],"minor_comments":[{"comment":"The definition of R^{l-1}_{i in P,N} is hard to parse, and the sentence \"where Psi^{l-1}_i is the mean value of non-zero-neurons in the layer l-1\" does not explain how a per-neuron subtraction term is computed from a mean; please clarify the notation.","section":"Section 2.1, Eq. (5)"},{"comment":"The GradCAM++ Focus score is printed as \"747\" rather than \"0.747\", which appears to be a typographical error.","section":"Table 8"},{"comment":"References [46] and [47] are the same paper (Smilkov et al. 2017); please remove the duplicate.","section":"References"},{"comment":"The paragraph following Table 9 contains the doubled phrase \"Additionally, Additionally\".","section":"Section 4.1"},{"comment":"The caption uses the acronym \"RAMP\" while the text and the rest of the paper use \"absLRP\"; the shorter acronym is never defined.","section":"Figure 1 caption"},{"comment":"In the pseudocode, the variable x is used both as the function input and inside autograd.grad(ha + h, x, rel_scaling), while the parameter is named input; this makes the intended dependence of rel on the layer's input ambiguous.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The main blockers are the Eq. (12)/Algorithm 1 factor-of-2 mismatch and the unaddressed non-conservation of the proposed rule. The evaluation is heavily weighted toward a self-proposed metric, and the external metrics do not consistently support the abstract's claim of clear advantage. I do not recommend rejection because the idea is simple, timely for ViT explainability, and potentially useful after the rule is clarified and the claims are recalibrated. I would ask the authors to reconcile the formal rule with the released code, provide a conservation or normalization analysis, report all GAE parameters, and rerun or rephrase the headline conclusions accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real novelty here is small but present: Eq. 12 defines a new LRP denominator using the absolute preactivation, and GAE is a new single-score evaluation combining gradient masking, attribution similarity, and mosaic contrastiveness. The experimental sweep is broad—VGG, ResNet50, ViT on ImageNet and PascalVOC, with many baselines and a code link. That is worth taking seriously.\n\nWhat the paper does well: adapting the rule to residual connections, batch/layer norm, and self-attention is nontrivial, and the authors include external metrics (ROAD, Lipschitz, Focus) alongside their own. The citation pattern is fine, with no obvious gaps.\n\nNow the soft spots. The central claim of clear superiority does not hold up on the external metrics. On ViT, absLRP gets lower Focus than GradCAM and HiResCAM (0.826 vs. 0.913), and its ImageNet ViT ROAD MoRF is worse than GradCAM's (0.457 vs. 0.317). The big wins show up mainly on the proposed GAE metric, which is a self-assessment. The authors acknowledge some of this, but the abstract and conclusion overstate it.\n\nMore serious: Eq. 12 is not conservative. Summing the numerators gives the positive part of the net preactivation, so when a neuron's preactivation is negative, a fraction of relevance is silently destroyed. The paper criticizes RAP's arbitrary over-allocation subtraction, but absLRP's under-allocation is equally arbitrary and undiscussed. For an LRP rule, conservation is what makes propagated scores interpretable.\n\nAnd there is a concrete mismatch: Algorithm 1 does not implement Eq. 12. The autograd of h+ha doubles the positive contributions per layer, so the experiments evaluate a different rule—scaled by 2^L, with branch-depth-dependent weighting in residual nets. Either the equation or the pseudocode is wrong, and the authors need to say which and whether results change.\n\nMinor: GAE's hyperparameters T, top-k, and epsilon are free, and the metric is not validated against independent human or perturbation benchmarks. The qualitative figures are suggestive but not evidence.\n\nBottom line: this is a solid engineering paper with a shaky central claim. A serious referee should engage with it, but the authors need to fix the implementation/equation mismatch, address conservation or justify non-conservation, and tone down the superiority claim to \"competitive\" on external metrics. I would not cite it in its current form.","headline":"A plausible but unproven LRP variant with a self-proposed metric; the algorithm/equation mismatch and non-conservation need fixing before the claims hold.","tokens_in":27163,"tokens_out":2590,"would_cite":false,"duration_ms":28880,"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":"This paper claims that changing one denominator in layer-wise relevance propagation—normalizing by the absolute total pre-activation of each neuron rather than by the sum of its positive contributions—fixes misattributions between neurons…","keywords":["explainable artificial intelligence","layer-wise relevance propagation","vision transformer","attribution-based evaluation","contrastive explanations","absolute magnitude normalization","pixel-level attribution","image classification"],"falsifier":"Build two small networks with identical input–output behavior, one obtained from the other by scaling a hidden neuron's incoming weights and compensating in the next layer; if absLRP assigns different attribution maps to the two (as Figure 1's toy example suggests it would), then the method tracks parameter magnitude rather than functional importance, and the claimed 'correction' becomes a design choice that a reader can test in minutes.","tokens_in":26034,"feed_emoji":"🔍","tokens_out":12539,"duration_ms":115349,"temperature":0.7,"pith_summary":"This paper claims that a single change to layer-wise relevance propagation—dividing by the absolute value of a neuron's total incoming signal instead of by the sum of its positive contributions—fixes a real error in how credit is shared between neurons of different activation magnitudes. On that basis it introduces absLRP and shows it produces sparse, noise-free, class-specific attribution maps on three architectures, including the Vision Transformer. It also introduces Global Attribution Evaluation (GAE), a single score that folds faithfulness, robustness, and localization together, and uses it to rank more than a dozen attribution methods on ImageNet and PascalVOC. In those comparisons absLRP outscores every method on VGG and ViT-Base and statistically ties contrastive LRP on ResNet50, while the authors acknowledge the ViT margin is smaller and leaves room for improvement. If the claim is correct, deep-network explanations can be obtained from a simple, architecture-agnostic rule that works at pixel resolution even when a Vision Transformer has very few patches.","feed_headline":"New LRP variant tops attribution tests on VGG and ViT","feed_subtitle":"A simple absolute-magnitude rule yields class-specific pixel-level maps; it tops prior methods on VGG and ViT.","key_machinery":"The load-bearing object is the absLRP rule in Eq. (12), which redistributes neuron $j$'s relevance to input $i$ as $(x_i w_{ij})_+ / (|\\sum_k x_k w_{kj}| + \\epsilon) R_j^l$: positive contributions in the numerator, absolute pre-activation magnitude in the denominator. It is implemented through automatic differentiation, where the ratio of incoming relevance to the absolute output of a layer is used as the gradient scaling, so the same code works for convolutions, residual branches, normalizations, and attention blocks with only layer-specific bookkeeping. The second mechanism is the Global Attribution Evaluation (GAE) metric, which multiplies a local consistency score—obtained by correlating MoRF/LeRF output-drop curves with attribution-map similarity curves, plus a faithfulness term from gradient-based impact maps—by a contrastiveness score derived from four-image mosaics.","core_discovery":"The central discovery on the paper's own terms is that existing LRP rules misallocate relative relevance when neurons in the same layer have different absolute activation magnitudes, and that replacing the denominator with the absolute value of the whole pre-activation corrects this. Concretely, LRP-$\\alpha_1\\beta_0$ uses $\\sum_k (x_k w_{kj})_+$ as the normalizer, so two toy networks that share the same input, hidden, and output values but differ in the magnitude of one hidden neuron's incoming weights receive identical attributions $[0.5, 0.5]$; absLRP, defined by $R_i^{l-1} = \\sum_j (x_i w_{ij})_+ / (|\\sum_k x_k w_{kj}| + \\epsilon) R_j^l$, changes the split to $[0.25, 0.75]$, which the paper argues better reflects the network's internal dynamics. From there the paper builds a full recipe for convolutional, residual, batch-normalized, layer-normalized, and self-attention layers, seeds the top layer with contrastive relevance (1 for the target class, $-1/N$ for the others), and reports that the resulting maps are the top-scoring ones under its GAE metric on VGG and ViT-Base, tying contrastive LRP on ResNet50.","pith_inferences":["AbsLRP does not conserve total relevance across layers, because its numerator sums only positive contributions while its denominator is the absolute total; the paper does not discuss this departure from LRP's conservation tradition.","The motivating toy example changes the network's weights without changing its input–output function; an alternative reading is that a good attribution should be invariant to such re-parameterizations, which would make absLRP's sensitivity to weight scale a testable weakness rather than a clear correction.","Because GAE's faithfulness term is built from gradient-based impact maps, the metric systematically favors methods whose maps align with input-gradients; attention-only methods such as Rollout are disadvantaged by construction, so GAE is best read as a comparison within gradient-aligned families.","The ViT results suggest a testable extension: on models with stronger normalization of internal activations the new denominator should matter less, so one could predict smaller absLRP gains on other LayerNorm-heavy transformer variants beyond ViT."],"forward_implications":["If absLRP's advantage holds, Vision Transformer explanations need no longer be upscaled from patch resolution: the method delivers sparse pixel-level maps even for a model with 49 patches.","The same rule transfers to non-image domains without modification; the authors demonstrate BERT sentiment attributions, so text and other discrete-input models inherit the recipe.","GAE offers a single-number ranking that unifies faithfulness, robustness, and localization, replacing the common practice of reporting several possibly conflicting metrics.","Because the last-layer seed is contrastive (target class positive, all others negative), the maps change with the queried class, making multi-class image analysis practical.","The ResNet50 tie with contrastive LRP implies the new denominator matters most in architectures with less normalized activation scales, which is consistent with the smaller margins the authors report for ViT-Base."],"supporting_citations":[{"why":"It establishes layer-wise relevance propagation and the epsilon and alpha-beta rules that the new denominator modifies.","marker":"[10]"},{"why":"It defines Relative Attributing Propagation, the predecessor whose denominator and relevance over-allocation problems motivate the new rule.","marker":"[33]"},{"why":"It is Contrastive-LRP, the source of the target-class/other-classes relevance initialization used to seed the last layer.","marker":"[22]"},{"why":"It is Transformer Interpretability Beyond Attention Visualization, the ViT baseline whose patch-scale resolution the new method overcomes.","marker":"[14]"},{"why":"It supplies conservative propagation for transformers, whose self-attention relevance handling is referenced in the ablation study.","marker":"[2]"},{"why":"It defines the AOPC/ABPC perturbation-curve metric, one of the standard faithfulness benchmarks in the evaluation.","marker":"[41]"},{"why":"It defines ROAD, the consistent masking evaluation strategy used as a second faithfulness benchmark.","marker":"[39]"},{"why":"It defines the Focus mosaic metric, the localization approach that the contrastiveness component of GAE extends.","marker":"[6]"},{"why":"It defines the Local Lipschitz robustness estimate used as the robustness benchmark in the comparisons.","marker":"[4]"}],"fun_headline_variants":["absLRP: sharper attribution maps on VGG and ViT","Relative magnitude fix boosts LRP for Transformers","Contrastive LRP plus absolute magnitudes wins on VGG, ViT","New GAE metric ranks absLRP first on VGG and ViT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the premise that the absolute magnitude of a neuron's total incoming signal is the right yardstick for splitting credit among its inputs; the paper motivates this with a toy example but supplies no conservation, axiomatic, or independent empirical justification for that particular denominator.","fun_headline_variants_meta":{"raw":{"variants":["absLRP: sharper attribution maps on VGG and ViT","Relative magnitude fix boosts LRP for Transformers","Contrastive LRP plus absolute magnitudes wins on VGG, ViT","New GAE metric ranks absLRP first on VGG and ViT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000591,"raw_usage":{"total_tokens":2865,"prompt_tokens":1135,"completion_tokens":1730,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":751,"completion_tokens_details":{"reasoning_tokens":1655}},"tokens_in":751,"tokens_out":1730,"duration_ms":11241,"temperature":1.0,"reasoning_tokens":1655,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:05:00.304820+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build two small networks with identical input–output behavior, one obtained from the other by scaling a hidden neuron's incoming weights and compensating in the next layer; if absLRP assigns different attribution maps to the two (as Figure 1's toy example suggests it would), then the method tracks parameter magnitude rather than functional importance, and the claimed 'correction' becomes a design choice that a reader can test in minutes.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It establishes layer-wise relevance propagation and the epsilon and alpha-beta rules that the new denominator modifies."},{"cited_title":"Relative Attributing Propagation: Interpreting the Comparative Contributions of Individual Units in Deep Neural Networks","cited_arxiv_id":"1904.00605","evidence_quote":"It defines Relative Attributing Propagation, the predecessor whose denominator and relevance over-allocation problems motivate the new rule."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It is Contrastive-LRP, the source of the target-class/other-classes relevance initialization used to seed the last layer."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies conservative propagation for transformers, whose self-attention relevance handling is referenced in the ablation study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It defines the AOPC/ABPC perturbation-curve metric, one of the standard faithfulness benchmarks in the evaluation."},{"cited_title":"In Proceedings of the 38th International Conference on Machine Learning, ICML 2021 (Proceedings of Machine Learning Research, Vol","cited_arxiv_id":null,"evidence_quote":"It defines ROAD, the consistent masking evaluation strategy used as a second faithfulness benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It defines the Focus mosaic metric, the localization approach that the contrastiveness component of GAE extends."}],"review_version":1}