{"id":"49e20ba3-faa0-4318-aaae-c17fc275128b","arxiv_id":"1908.01536","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"For 3D CNNs on video, subtracting per-frame deep Taylor explanations from the full-clip explanation produces a residual map intended to isolate motion relevance.","lead":"This paper separates a video explanation into spatial and temporal parts by explaining the whole clip with deep Taylor decomposition, then subtracting the explanation of each frozen frame. The goal is to show which moving regions drive a 3D action recognition model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Temporal relevance residual rests on an untested additivity/class-matching assumption; Section 5 concedes the decomposition is only an approximation, so the central claim is not established.","rationale":"The paper proposes a simple and intuitive subtraction of deep Taylor maps to separate spatial and temporal relevance in 3D CNN explanations. The central claim, however, depends on a strong additivity assumption that is neither proven nor quantitatively validated. The reader's weakest assumption identified exactly this: the subtraction requires R_video = R_spatial + R_temporal, and the repeated-frame spatial relevance must equal the spatial relevance in the original video. The paper's own Section 5 concedes the relevance values do not sum to the original and that frame-only inputs predict different classes, which undermines the assumption. My stress-test focuses on the class-mismatch mechanism: the spatial map is computed from an input for which the explained class may not be the model's decision, making the deep Taylor propagation from a near-zero or negative output neuron potentially arbitrary. This is not an internal inconsistency—the authors honestly label the method an approximation—but it means the residual cannot be claimed to expose true motion relevance without additional evidence. The proposed concrete test (a synthetic moving-object video with ground-truth motion masks) would directly settle whether the residual localizes motion. Since the reader's conditional verdict already demands such quantitative validation, my read does not change the verdict: conditional acceptance is appropriate, pending this check.","tokens_in":6891,"tokens_out":4942,"duration_ms":52512,"concrete_test":"Construct a synthetic 16-frame video with a known moving object (e.g., a solid rectangle translating over a fixed background). Run the same fine-tuned C3D model and compute the original deep Taylor map, the per-frame spatial maps (repeat each frame n times), and the temporal residual. Threshold positive relevance at the 90th percentile and measure the IoU (or pointing-game accuracy) between the residual's support and the ground-truth moving-object mask for each frame. Also recompute the spatial maps using the frame-only model's predicted class and compare the residual; if the residual's localization score is no better than chance, or changes drastically with the class chosen, the claim that the residual is temporal relevance is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the residual R_video - R_spatial exposes motion relevance. This requires R_video to decompose additively as R_spatial + R_temporal, and requires the per-frame spatial map (computed on a repeated-frame static input) to equal the spatial component of the same frame inside the moving video. Deep Taylor decomposition is nonlinear and input-dependent, so this additivity is not guaranteed. Section 5 reports that frame-only inputs predict different classes (Wall Push-Ups, Golf Swing, Clean And Jerk for the pull-up clip). When the explained class is not the frame-only model's output, the propagated relevance from that output neuron may be small or numerically unstable, so the resulting 'spatial relevance' is not a clean attribution of the video model's decision. Even when the class matches, the static forward pass has different intermediate activations than the moving video, so equal spatial relevance cannot be assumed. The paper itself calls the decomposition an approximation and says future work must 'formalise a method for exposing the true spatial relevance.' Thus the residual may be an artifact of class mismatch and model nonlinearity, not motion relevance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses explainability of 3D CNNs for video action recognition. Using deep Taylor decomposition, the authors compute a relevance map for a video input and, for each frame, a separate relevance map obtained from a static input formed by repeating that frame n times. They interpret the per-frame map as spatial relevance and compute the residual after subtracting it from the video map as temporal (motion) relevance. The method is demonstrated on two UCF-101 samples, a pull-up clip and a tennis-serve clip, with qualitative heatmap comparisons. The authors also observe that the frame-only inputs can produce different predicted classes than the video input and explicitly describe the decomposition as an approximation.","tokens_in":7071,"tokens_out":2920,"duration_ms":33433,"significance":"If the proposed subtraction reliably isolates motion-attributed relevance, it would be a simple, low-cost extension of deep Taylor decomposition to a temporal domain and would address a genuine gap in video explainability. The paper releases its implementation, is transparent about its limitations, and connects the method to prior LRP and saliency-tube work. However, the central claim is not yet established: the paper provides no formal condition under which the residual equals temporal relevance and no quantitative validation. The two qualitative examples, while suggestive, do not by themselves demonstrate that the residual is not an artifact of class mismatch or model nonlinearity.","major_comments":[{"comment":"The separation method rests on an additivity assumption that is not stated or justified. The paper implicitly assumes that the video-level deep Taylor map decomposes as R_video = R_spatial + R_temporal and that the spatial map computed from a repeated-frame static input equals the spatial component of the same frame inside the moving video. Deep Taylor decomposition is nonlinear and input-dependent, so neither identity follows from the LRP rules in Section 4.3. The manuscript needs either a formal derivation of sufficient conditions for this additivity or a direct empirical test on inputs where the temporal relevance is known by construction.","section":"Section 3.2"},{"comment":"The reported class mismatch directly undermines the spatial maps. For the pull-up clip, the frame-only inputs are predicted as Wall Push-Ups, Golf Swing, or Clean And Jerk rather than Pull-Ups. The spatial relevance is then backpropagated from the output neuron of Pull-Ups under a forward pass where that class is not the model's decision, so the resulting map can be small, unstable, or dominated by the model's lack of evidence for that class. This is not a clean spatial attribution of the video model's Pull-Ups decision, and the residual therefore cannot be interpreted as temporal relevance without additional analysis.","section":"Section 5"},{"comment":"The paper's own evidence indicates the decomposition is not exact: the authors state that the sum of spatial and temporal relevance is not equal to the original relevance and that spatial relevance for a non-dominant class can exceed that for the dominant class. These observations contradict the additive model required by the method. The authors call the result an approximation, but the conclusion goes on to say that formalising true spatial relevance is future work. Given this, the central claim that the residual 'exposes relevance in the frame attributed to motion' is not currently supported.","section":"Section 5"},{"comment":"The empirical evaluation is limited to two qualitative examples, with no quantitative measure of whether the residual corresponds to motion. A reproducible validation would help: for example, synthetic videos with known pixel-level motion, or a comparison against an independent motion-localisation method such as saliency tubes, with a numeric overlap metric. The statement that 'similar results were observed for other test samples' is not accompanied by details and cannot be checked.","section":"Section 5"}],"minor_comments":[{"comment":"The repetition count n is described as 'a typical sample size-number', which is vague. The paper should state the exact value used (apparently 16 for the shown samples) and ideally report sensitivity to this choice, since a different n changes both the network's internal activations and the resulting spatial map.","section":"Section 4.4"},{"comment":"The caption is hard to parse because the left and right panels are said to be from two different 16-frame samples but the sample names are not given in the caption. Please label the rows and panels explicitly as 'pull-ups' and 'tennis serve' and make clear which frames are displayed.","section":"Figure 2"},{"comment":"The phrase 'downweighting relevance' in Section 3.2 is inaccurate, since the operation actually subtracts one map from another. Please use consistent terminology ('subtract') throughout.","section":"Section 5"},{"comment":"The description of saliency tubes [Stergiou et al., 2019] would benefit from a sentence explaining how its output and evaluation differ from the proposed residual map, since it is the closest existing method for motion localisation in 3D CNN explanations.","section":"Related Work"},{"comment":"Minor typographical issues include 'white-box' being inconsistently hyphenated and the missing comma after 'state-of-the-art performance' in the first paragraph of Section 2.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the trick is real and new as far as I can tell; the interpretation is not yet supported. The construction — repeat a single frame n times, backprop deep Taylor through a 3D CNN on that static clip, subtract that map from the full-video map — is a natural extension of contrastive LRP and is not in the cited literature. They also ship the torchexplain code, which is a concrete contribution for anyone who wants to reproduce or adapt the idea. The paper is honest: Section 5 explicitly says the decomposition is an approximation, the spatial and temporal relevance do not sum to the original, and the frame-only inputs predict different classes. That candor is real and useful. The two qualitative examples are at least plausible — the residual highlights the man’s body and the tennis ball rather than background edges — which suggests the method does something visually interesting. The central claim that the residual is temporal relevance, however, rests on an additivity assumption that is never tested. Deep Taylor is nonlinear and input-dependent; there is no reason to expect R_video = R_spatial + R_temporal, nor that a frame repeated n times produces the same spatial relevance as the same frame inside a moving clip. The class-mismatch problem the authors themselves report makes this worse: explaining class A for a static frame that the model classifies as B can produce relevance that is a numerical artifact, not a clean spatial attribution. And there is no quantitative evaluation: no localization metric against motion annotations, no faithfulness test, no comparison against saliency tubes, which is the closest related method. Given the paper’s own caveats, the residual map is a heuristic visualization, not a validated decomposition. None of these flaws sink the paper if it is framed as an exploratory result showing a low-cost trick and exposing a limitation of off-the-shelf deep Taylor for video. But as a claim that this exposes relevance attributed to motion, it is not established. People working on explainable video understanding or interactive debugging of 3D action recognition models will get value from this. It is workshop-grade as-is, though it could become a solid paper with quantitative validation and a reframing of the residual as a heuristic. I would send it to peer review rather than desk reject — the idea is new and the code ships — but I would expect revision requiring quantitative evaluation and a clearer separation between what is demonstrated and what is conjectured.","headline":"A simple, honest frame-repeat-and-subtract trick for separating spatial from temporal relevance in deep Taylor maps, but the paper does not yet prove the residual is motion relevance.","tokens_in":7618,"tokens_out":1556,"would_cite":false,"duration_ms":16860,"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":"Deep Taylor decomposition can be made to expose motion relevance in video by subtracting per-frame spatial explanations from the full-clip explanation.","keywords":["explainable AI","deep Taylor decomposition","layer-wise relevance propagation","3D convolutional neural networks","spatio-temporal relevance","action recognition","UCF-101","motion relevance"],"falsifier":"Use a no-motion clip, a single frame repeated $n$ times, and compute $R_{\\text{video}} - R_{\\text{spatial}}$; true temporal relevance should be zero, so any nonzero residual would show the subtraction captures more than motion. A second check: across real clips, compare the residual maps to optical-flow magnitude; the residual should track moving regions and vanish where there is no flow.","tokens_in":6649,"feed_emoji":"🎬","tokens_out":9860,"duration_ms":95125,"temperature":0.7,"pith_summary":"This paper tries to establish that deep Taylor decomposition, applied directly to a 3D convolutional network's video input, does not produce an explanation that separates why a frame matters: static edges and background dominate, and the contribution of motion stays hidden. The proposed method computes a deep Taylor map for the whole clip and, separately, for each frame frozen by repeating it across the temporal dimension; subtracting the per-frame spatial maps from the clip map leaves a residual the authors interpret as temporal relevance. They demonstrate the residual on UCF-101 clips of pull-ups and a tennis serve, where it concentrates on the moving body, the head, and the ball rather than the scene's static structure. If the method works, video explanations can name not just which pixels matter, but whether they matter because of appearance or because of movement.","feed_headline":"Freeze-frame subtraction exposes motion in video AI explanations","feed_subtitle":"Subtracting per-frame explanations exposes the motion evidence hidden in video AI maps.","key_machinery":"The load-bearing mechanism is the residual decomposition $R_{\\text{temporal}} = R_{\\text{video}} - R_{\\text{spatial}}$, where $R_{\\text{video}}$ is the deep Taylor relevance of the full clip backpropagated to the input block and $R_{\\text{spatial}}$ is built one frame at a time by feeding the network a freeze frame, the same frame repeated $n$ times to match the clip's temporal extent. Deep Taylor decomposition propagates relevance backward from the chosen output neuron through layer-wise relevance propagation rules: ReLU layers pass relevance through unchanged, pooling layers distribute it by selection or averaging, convolutional layers use the $\\alpha\\beta$ rule, and the first convolutional layer uses the $z^\\beta$ rule over the pixel range $0$ to $255$. The freeze-frame padding avoids zero padding, which would inject false temporal information through a sudden all-pixels-to-black transition. The subtraction then acts as a negative mask: strongly spatial features become negative in the residual, leaving the motion-driven positive relevance visible.","core_discovery":"The paper's central claim is that relevance in a 3D CNN explanation has spatial and temporal components that a single deep Taylor pass cannot separate, and that the temporal component can be recovered by a subtraction. For a clip, one deep Taylor pass produces a combined spatio-temporal map; for each frame, another pass on the frame repeated n times produces a spatial-only map. Subtracting the second from the first yields a temporal relevance map that the authors call a discriminative relevance model, and in their examples this residual marks motion: the body and head at the turning points of a pull-up, and the ball and upper body in a tennis serve, while static structures like the metal frame, the lawn, and a watermark become negative. This is the discovery: the residual is not just noise; it is a readable motion relevance map that the original explanation leaves ambiguous.","pith_inferences":["If the temporal residual tracks optical flow or pose trajectories, it could serve as a cheap pseudo-label source for weakly supervised action localization, where only video-level class labels are available.","Because freeze-frame inputs make the model predict classes other than the clip's class, the spatial maps are computed under a different decision state than the video map; a more faithful decomposition would hold the predicted class fixed while ablating motion.","The same subtractive logic could be applied to other backpropagation-based attribution methods, with the additivity of the residual as a measurable criterion for which methods support spatial/temporal separation.","Formalizing the decomposition as an optimization—find the minimal ablation of motion that preserves the class decision—could turn the heuristic subtraction into a principled explanation objective."],"forward_implications":["Video explanations made by image-style attribution methods are dominated by spatial edge relevance, so applying such methods as-is to video misleads users about what drove the model.","The freeze-frame subtraction supplies a motion relevance map for every frame at a cost that scales linearly with the number of frames.","Temporal relevance in an activity like pull-ups concentrates at the turning points of the motion, giving a per-frame localization of when movement matters.","Regions that become negative in the residual are identified as spatially relevant rather than motion-relevant, separating static context from action evidence.","The method can be attached to existing LRP-based explainers for 3D CNNs without retraining the classifier."],"supporting_citations":[{"why":"Defines deep Taylor decomposition, the relevance propagation method whose video and freeze-frame maps are subtracted.","marker":"[Montavon et al., 2017]"},{"why":"Introduces layer-wise relevance propagation rules that the explainer implements during backpropagation.","marker":"[Bach et al., 2015]"},{"why":"Supplies the UCF-101 action recognition dataset and the pull-up and tennis-serve samples used for the demonstrations.","marker":"[Soomro et al., 2012]"},{"why":"Provides the C3D architecture on which the fine-tuned 3D CNN is based.","marker":"[Tran et al., 2015]"},{"why":"Earlier LRP-based video explanation work whose whole-clip relevance curve does not localize spatially relevant objects.","marker":"[Srinivasan et al., 2017]"},{"why":"The most similar method for localizing salient motion in 3D CNNs, used as the comparison point for saliency tubes.","marker":"[Stergiou et al., 2019]"},{"why":"Supports padding by frame repetition instead of zero padding, since replacing input values with zeros distorts attribution.","marker":"[Hooker et al., 2018]"},{"why":"Demonstrates contrastive relevance by subtracting one relevance signal from another, the conceptual template for the proposed subtraction.","marker":"[Zhang et al., 2018]"}],"fun_headline_variants":["Subtracting spatial maps reveals motion in video AI explanations","Motion relevance extracted by subtracting frames in video explanations","Deep Taylor decomposition split into spatial and temporal maps","Residual maps expose motion hidden in video AI relevance","Frame subtraction isolates motion in explainable video AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The decomposition stands on the assumption that relevance is approximately additive—$R_{\\text{video}} = R_{\\text{spatial}} + R_{\\text{temporal}}$—even though the paper itself notes the two components do not sum to the original and the model predicts different classes on freeze-frame inputs.","fun_headline_variants_meta":{"raw":{"variants":["Subtracting spatial maps reveals motion in video AI explanations","Motion relevance extracted by subtracting frames in video explanations","Deep Taylor decomposition split into spatial and temporal maps","Residual maps expose motion hidden in video AI relevance","Frame subtraction isolates motion in explainable video AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1325,"prompt_tokens":892,"completion_tokens":433,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":360}},"tokens_in":508,"tokens_out":433,"duration_ms":4678,"temperature":1.0,"reasoning_tokens":360,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:09:58.850094+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use a no-motion clip, a single frame repeated $n$ times, and compute $R_{\\text{video}} - R_{\\text{spatial}}$; true temporal relevance should be zero, so any nonzero residual would show the subtraction captures more than motion. A second check: across real clips, compare the residual maps to optical-flow magnitude; the residual should track moving regions and vanish where there is no flow.","supporting_citations":[{"cited_title":"On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation","cited_arxiv_id":null,"evidence_quote":"Introduces layer-wise relevance propagation rules that the explainer implements during backpropagation."},{"cited_title":"Learning spa- tiotemporal features with 3D convolutional networks","cited_arxiv_id":null,"evidence_quote":"Provides the C3D architecture on which the fine-tuned 3D CNN is based."},{"cited_title":"Interpretable human action recogni- tion in compressed domain","cited_arxiv_id":null,"evidence_quote":"Earlier LRP-based video explanation work whose whole-clip relevance curve does not localize spatially relevant objects."},{"cited_title":"Saliency tubes: Visual ex- planations for spatio-temporal convolutions,","cited_arxiv_id":null,"evidence_quote":"The most similar method for localizing salient motion in 3D CNNs, used as the comparison point for saliency tubes."},{"cited_title":"Top-down neural attention by excitation backprop","cited_arxiv_id":null,"evidence_quote":"Demonstrates contrastive relevance by subtracting one relevance signal from another, the conceptual template for the proposed subtraction."}],"review_version":1}