{"id":"61fa9d56-6bc0-478c-b152-c8dbc96dda0e","arxiv_id":"2501.11309","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Finer-CAM computes class activation maps from the difference between a target class and a similar reference class, suppressing shared features and highlighting discriminative details.","lead":"Finer-CAM is a simple upgrade to class activation maps that highlights the small visual details separating similar classes, such as two nearly identical bird species, by having the model explain the difference between a target class and a similar class instead of explaining the target class alone. This could make AI explanations more useful for fine-grained tasks in biology and other fields.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on the untested premise that gradients of y_c − γ y_d point to visually distinguishing regions; the sole quantitative support (relative confidence drop) is aligned with the method's objective, while deletion AUC shows no meaningful advantage.","rationale":"The reader's primary weakest assumption is that y_c − y_d is a faithful proxy for visually distinguishing features, and this is exactly the load-bearing concern I find. The paper provides only qualitative examples and a metric (relative confidence drop) that is inherently aligned with the method's objective. My proposed test uses CUB-200's attribute and part annotations to directly assess whether the logit-difference gradient points to the actual distinguishing parts, which would settle the proxy question without relying on the circular metric. I also note the paper's honest failure-case discussion in Sec. C.1, which explicitly limits the method to settings where logit similarity tracks visual similarity and the classifier is correct, reinforcing that the central claim needs a scope condition. The multi-modal extension and the clean linear-subtraction analysis in Eq. (7) are valuable and implemented with modest, credible engineering, but they do not resolve the empirical gap. Since this is an evidence shortfall rather than a demonstrated contradiction, the appropriate disposition is the already-assigned CONDITIONAL verdict, so no change to the reader's verdict is needed.","tokens_in":16175,"tokens_out":4742,"duration_ms":52883,"concrete_test":"On CUB-200, use the provided attribute and part annotations to define ground-truth discriminative regions: for each target class c and its most similar class d, select the parts associated with the attributes whose labels differ most between c and d (following the paper's own Sec. A.2 procedure). Then compute the fraction of Finer-CAM's top-5% activation energy that falls within these part masks, averaged over test images, and compare it to Grad-CAM and Layer-CAM. If Finer-CAM does not significantly exceed baselines on this direct discriminative-part agreement while still achieving higher RD, then RD is measuring suppression of the reference logit rather than localization of true distinguishing visual features.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The method's core step, Eq. (5), replaces the explanation target y_c with y_c − γ y_d, assuming that gradients of this logit difference localize features unique to class c relative to class d. No evidence is provided that this logit-space proxy corresponds to visual discriminativeness; the paper's own failure cases (Fig. 10) state degeneration when logit similarity does not reflect visual similarity or when the classifier is wrong. The quantitative case rests on the relative confidence drop, Eq. (10): RD = (p_c − p_c*) − (p_d − p_d*). Masking the top pixels highlighted by Finer-CAM should, by construction of Eq. (5), lower y_c more than y_d, so RD is directly aligned with the objective being optimized. The authors tune γ and aggregation on this metric (Fig. 6b), making the reported RD gains expected rather than independent confirmation. Standard deletion AUC, which does not have this built-in alignment, shows negligible differences in Tables 1, 4, and 5. The localization gains that do appear use only bounding-box pointing games, which cannot confirm that the activated regions are the specific distinguishing details. Thus the central claim that Finer-CAM reveals genuinely finer discriminative details is not yet supported by evidence that separates the method's mechanism from the metric's objective alignment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Finer-CAM, a modification of CAM-based explanation methods that replaces the explanation target, the target class logit y_c, with the logit difference y_c - γ y_d, where d is a visually similar class. The authors argue that this suppresses features shared with similar classes and highlights discriminative details. They derive the method for gradient-based and score-based CAM, extend it to CLIP-based multi-modal models, and evaluate on fine-grained datasets using deletion AUC, a new relative confidence drop metric, and pointing-game localization.","tokens_in":16387,"tokens_out":3797,"duration_ms":34630,"significance":"If the central claim is correct, the method is valuable: it is a simple, parameter-light modification compatible with existing CAM methods and could improve fine-grained interpretability. The derivation in Eqs. (5)-(7) is mathematically sound, and the paper includes a useful multi-modal extension and qualitative consistency analysis. However, the empirical evidence is not yet convincing because the main quantitative metric is aligned with the method's objective and is used for hyperparameter selection.","major_comments":[{"comment":"The load-bearing premise that gradients of y_c - γ y_d localize visually discriminative features is asserted rather than demonstrated. The paper's own failure cases (Fig. 10) concede that Finer-CAM degenerates when the classifier is wrong or when logit similarity does not reflect visual similarity, which indicates the premise is not guaranteed. To support the central claim, the authors should validate the correspondence between logit differences and human-annotated distinguishing attributes (e.g., using CUB part/attribute annotations) or provide a controlled experiment.","section":"§3.2, Eq. (5)"},{"comment":"The relative confidence drop metric is defined as (p_c - p_c*) - (p_d - p_d*), which directly measures the quantity that Finer-CAM is designed to maximize: the gap between target and reference logits. Because γ and the number of aggregated references are tuned on this metric on the same datasets (Fig. 6b), the reported RD improvements are expected by construction and do not constitute independent evidence of finer localization. The authors should report an evaluation metric that is not aligned with the training objective, such as insertion/deletion AUC or agreement with fine-grained part annotations.","section":"§3.4, Eq. (10), Table 1"},{"comment":"Standard deletion AUC shows negligible or inconsistent differences between Finer-CAM and baselines (e.g., Birds-525 Grad-CAM 0.079 vs. +Finer 0.076; CUB 0.024 vs. 0.024; Cars 0.024 vs. 0.024). Since deletion AUC is the standard faithfulness metric and is not aligned with Finer-CAM's objective, the lack of improvement is a serious concern for the claim that Finer-CAM yields more faithful explanations. The paper should explain this discrepancy or provide a direct, non-aligned measure of discriminative-detail localization.","section":"Tables 1, 4, and 5"},{"comment":"The energy-based pointing game only evaluates whether activations fall within the object bounding box, which cannot distinguish between coarse object-level localization and localization of the specific fine-grained details that the method claims to identify. The authors should use part-level or attribute-level annotations (available for CUB-200) to show that Finer-CAM highlights the actual distinguishing regions rather than just the object.","section":"§4.2, Localization"}],"minor_comments":[{"comment":"The baseline term for the reference class appears to be missing; the expression should likely be [f(x◦H_k)c - f(xb)c] - γ[f(x◦H_k)d - f(xb)d] or an explicit assumption that the baseline terms cancel.","section":"§3.2, Eq. (9)"},{"comment":"The text uses 'Fine-CAM' where 'Finer-CAM' is intended.","section":"Supplementary B.3"},{"comment":"The label 'Aggreagte' is a typo for 'Aggregate'.","section":"Figure 6b"},{"comment":"The paper states that 'the top 5% of activated pixels' are masked but does not specify how the threshold is applied across images of different sizes; please clarify the masking procedure.","section":"§3.4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is clearly written and the core idea is appealing, but the evaluation gap is substantial. I would encourage the editor to request a major revision rather than reject, because the core mechanism is simple and could be validated with additional experiments using existing annotations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Finer-CAM is a simple, sensible reframing of CAM for fine-grained explanation: explain the logit difference y_c − γ y_d instead of the target logit y_c alone. The derivation (Eqs. 5–7) is correct and reduces to subtracting reference CAM weights from target CAM weights, so it plugs into any gradient- or score-based CAM with a one-line change. The qualitative results are genuinely good—wing-color differences, head/back distinctions, and the multi-modal CLIP extension ('red epaulets' minus 'bird') is a nice practical idea. The authors also show failure cases when the classifier is wrong or the reference class isn't visually similar, which is honest.\n\nThe soft spot is the evaluation. The headline metric (relative confidence drop, Eq. 10) is introduced in this paper and directly measures what Finer-CAM optimizes: masking top activated pixels should reduce target-class confidence more than similar-class confidence. The strength γ is then tuned on that metric (Fig. 6b), so the reported gains are, to a large degree, expected. Standard deletion AUC shows no meaningful improvement on any of the five datasets, and the pointing game only reports bounding-box overlap, which can't verify that the activated regions are the specific distinguishing details. There are no error bars or significance tests, and the paper does not cite or compare against contrastive explanation methods (e.g., CEM or other difference-based approaches), which is a notable omission given the method's framing.\n\nNone of this kills the paper. The idea is plausible, the math is clean, and the code is released. But the central claim—that Finer-CAM reveals genuinely finer discriminative details—needs stronger independent evidence. A revision should include standard deletion/insertion curves with variance, part-level or pixel-level ground truth where available, and a head-to-head against contrastive explanation baselines.\n\nVerdict: send it to peer review. The evaluation is weak but fixable, and the method is the kind of simple, broadly applicable tool that reviewers will want to see properly vetted.","headline":"Clean idea, weak evaluation: Finer-CAM's logit-difference framing is a nice practical trick, but the headline metric is aligned with its objective, so independent evidence is still needed.","tokens_in":16998,"tokens_out":2568,"would_cite":true,"duration_ms":24536,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Finer-CAM proposes that class activation maps should explain the difference between a target class and a visually similar class, not the target class alone.","keywords":["class activation map","fine-grained visual explanation","saliency map","logit difference","relative confidence drop","zero-shot localization","interpretability","explainable AI"],"falsifier":"Take a target class and pair it with its most similar class and with its least similar class, then compute Finer-CAM maps for both pairs. If the least-similar pairing matches or beats the most-similar pairing on relative confidence drop at 5% masking, then the comparison of similar classes is not what drives the reported gain.","tokens_in":15905,"feed_emoji":"🔍","tokens_out":5672,"duration_ms":42822,"temperature":0.7,"pith_summary":"Finer-CAM claims that class activation maps fail on fine-grained tasks not because of how they combine feature maps but because of what they are asked to explain: the target class's prediction score alone. The paper proposes instead to explain the difference between the target score and the score of a visually similar class (the logit difference), so that features shared with the lookalike are suppressed and the features that genuinely separate the classes are highlighted. If correct, this means any CAM variant can gain fine-grained precision simply by changing the explanation target, with no extra computation or training. The authors demonstrate the mechanism on gradient-based and score-based CAMs, and extend it to CLIP-style text prompts for localizing specific concepts in an image.","feed_headline":"Spot-the-difference trick sharpens class activation maps","feed_subtitle":"Explaining the gap to a similar class reveals the details that truly separate them.","key_machinery":"The machinery is the logit-difference target $y_c - \\gamma y_d$ in Eq. (5), where $y_c$ and $y_d$ are the raw class scores for the target and a visually similar class, and $\\gamma$ is the comparison strength. For gradient-based CAMs, differentiation linearity turns this into a per-channel weight subtraction $\\alpha^{c,d}_k = \\alpha^c_k - \\gamma \\alpha^d_k$ (Eq. 7) before the ReLU, which differs from subtracting saliency maps after the fact. This object carries the argument because it converts a class-conditional explanation into a contrastive one, and the adjustable $\\gamma$ provides a continuous dial between coarse object-level activation and fine discriminative detail.","core_discovery":"The central discovery is that the explanation target, not the explanation machinery, is what limits CAM in fine-grained settings. Formally, Finer-CAM replaces the target logit $y_c$ with $y_c - \\gamma y_d$, where $d$ is a visually similar class and $\\gamma$ controls comparison strength, and computes CAM weights from $\\partial(y_c - \\gamma y_d)/\\partial A^k_{ij}$. Because the gradient is linear, the resulting weight for channel $k$ is $\\alpha^{c,d}_k = \\alpha^c_k - \\gamma \\alpha^d_k$, so the map positively weights features that raise the target logit without also raising the lookalike's logit. The paper shows this yields saliency maps that, when their top 5% of pixels are masked, drop the target class's confidence relative to the similar class's confidence more than baseline CAMs do, and that the same comparison can be applied to score-based CAMs and to zero-shot text-prompt localization.","pith_inferences":["The same comparison trick could work in other explanation families, such as perturbation- or attention-based methods, wherever a scalar class score is available.","Because the reference class is a free choice, the method doubles as a way of asking a model what would make an image look like class $d$ instead of class $c$, which could support interactive debugging of confusions.","The paper's failure cases suggest the improvement depends on choosing references with genuinely similar visual features; an adversarial test with least-similar references would show how much of the gain is due to the similarity selection.","The multi-modal extension offers a concrete audit tool: where classifier-driven and attribute-driven Finer-CAM maps disagree, either the classifier is using non-attribute cues or the attribute annotation is incomplete."],"forward_implications":["Any gradient-based CAM (Grad-CAM, Layer-CAM, and similar) can be turned into a finer-grained explainer by substituting the logit-difference derivative, without retraining or extra forward passes.","Score-based CAMs get the same recipe via Eq. (9): weight each feature map by how much it raises the target score relative to the reference score.","In zero-shot multi-modal models, comparing a concept prompt against a base-category prompt localizes that concept more accurately than prompting the concept alone.","The relative confidence drop metric (Eq. 10) tracks the target-vs-lookalike confidence gap rather than target confidence alone, exposing differences that standard deletion curves miss.","Aggregating comparisons over the top few similar classes produces a saliency map that captures multiple discriminative traits of the target class."],"supporting_citations":[{"why":"Grad-CAM supplies the baseline weighting formula and the linearity result that the channel weight equals the classifier weight, which Finer-CAM modifies with the comparison term.","marker":"[37]"},{"why":"Score-CAM provides the score-based weighting scheme that Finer-CAM extends by subtracting the reference class's score effect.","marker":"[41]"},{"why":"Layer-CAM serves as another gradient-based CAM baseline on which Finer-CAM is applied and evaluated.","marker":"[17]"},{"why":"The original CAM formulation defines the linear-combination paradigm that Finer-CAM preserves while changing the explanation target.","marker":"[48]"},{"why":"CLIP is the multi-modal zero-shot model that Finer-CAM is extended to, using text embeddings as comparison targets.","marker":"[32]"},{"why":"RISE supplies the deletion-curve evaluation protocol that motivates the paper's proposed relative confidence drop metric.","marker":"[30]"}],"fun_headline_variants":["Subtract lookalike class to sharpen CAM for fine-grained tasks","Compare target class with similar ones to reveal finer details","Finer-CAM: spotting the difference sharpens activation maps","New CAM variant compares classes to highlight unique details","Subtract the similar class logit for finer CAM explanations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the difference between two class scores, $y_c - y_d$, faithfully points to the visual features that distinguish the classes, so that gradients of that difference land on genuinely discriminative image regions.","fun_headline_variants_meta":{"raw":{"variants":["Subtract lookalike class to sharpen CAM for fine-grained tasks","Compare target class with similar ones to reveal finer details","Finer-CAM: spotting the difference sharpens activation maps","New CAM variant compares classes to highlight unique details","Subtract the similar class logit for finer CAM explanations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000687,"raw_usage":{"total_tokens":3147,"prompt_tokens":1009,"completion_tokens":2138,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":2057}},"tokens_in":625,"tokens_out":2138,"duration_ms":17479,"temperature":1.0,"reasoning_tokens":2057,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:25:59.479160+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a target class and pair it with its most similar class and with its least similar class, then compute Finer-CAM maps for both pairs. If the least-similar pairing matches or beats the most-similar pairing on relative confidence drop at 5% masking, then the comparison of similar classes is not what drives the reported gain.","supporting_citations":[{"cited_title":"Grad-cam: Visual explanations from deep networks via gradient-based localization","cited_arxiv_id":null,"evidence_quote":"Grad-CAM supplies the baseline weighting formula and the linearity result that the channel weight equals the classifier weight, which Finer-CAM modifies with the comparison term."},{"cited_title":"Score-cam: Score-weighted visual explanations for convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Score-CAM provides the score-based weighting scheme that Finer-CAM extends by subtracting the reference class's score effect."},{"cited_title":"Layercam: Exploring hierarchical class activation maps for localization","cited_arxiv_id":null,"evidence_quote":"Layer-CAM serves as another gradient-based CAM baseline on which Finer-CAM is applied and evaluated."},{"cited_title":"bird” as comparing pairs in the zero-shot classification setting. It shows the “ground truth","cited_arxiv_id":null,"evidence_quote":"The original CAM formulation defines the linear-combination paradigm that Finer-CAM preserves while changing the explanation target."},{"cited_title":"Learn- ing transferable visual models from natural language super- vision","cited_arxiv_id":null,"evidence_quote":"CLIP is the multi-modal zero-shot model that Finer-CAM is extended to, using text embeddings as comparison targets."},{"cited_title":"Rise: Randomized input sampling for explanation of black-box models","cited_arxiv_id":null,"evidence_quote":"RISE supplies the deletion-curve evaluation protocol that motivates the paper's proposed relative confidence drop metric."}],"review_version":1}