REVIEW 4 major objections 5 minor 17 references
Discriminating Spatial and Temporal Relevance in Deep Taylor Decompositions for Explainable Activity Recognition
T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read Deep Taylor decomposition can be made to expose motion relevance in video by subtracting per-frame spatial explanations from the full-clip explanation.
desk verdict 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. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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.
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (4)
- [Section 3.2] 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 5] 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 5] 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 5] 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.
minor comments (5)
- [Section 4.4] 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.
- [Figure 2] 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 5] 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.
- [Related Work] 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 2] 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.
Circularity Check
No circular derivation: temporal relevance is a direct subtraction of externally computed deep Taylor maps; the conced- ed approximation is a validity concern, not circularity.
full rationale
The paper's central construction is a direct arithmetic operation on two independently computed deep Taylor relevance maps: the full-video map and a per-frame 'freeze-frame' map. No parameter is fitted and no quantity is predicted from a value that was used to fit it. The residual R_video − R_spatial is introduced in Section 3.2 as 'what is left' after removing spatial relevance, and the paper explicitly frames the method as a 'naïve representation' rather than a derived theorem. The only reference to the authors' own work is the torchexplain library (Section 4.1), which is an implementation detail and not used as evidence for the conceptual claim; the underlying deep Taylor decomposition is cited to external work (Montavon et al.; innvestigate). Section 5 contains a genuine limitation—frame-only inputs yield different predicted classes and the spatial and temporal maps do not sum to the original relevance—but this is a correctness and validity threat (the residual may be an artifact of class mismatch and model nonlinearity), not a circular derivation. The paper does not define spatial relevance in terms of temporal relevance, nor temporal relevance in terms of the output claim; the residual is a constructed object whose semantic interpretation is an untested assumption. Under the rules requiring a quoted reduction or fitted-value prediction to establish circularity, no such reduction is present.
Assumptions & free parameters
free parameters (1)
- frame repetition count n =
not specified; 16 frames used in examples
assumptions (3)
- domain assumption A 3D CNN's spatial reasoning on a frame repeated n times approximates the spatial component of its reasoning on the original moving clip.
- ad hoc to paper Relevance is additively separable: the original deep Taylor map equals the sum of a spatial map and a temporal map, so subtracting one isolates the other.
- domain assumption The class predicted for spatial-only inputs need not match the video class for the residual to be meaningful.
invented entities (1)
-
Temporal (motion) relevance map
Cite this review
Pith. "Pith review of Discriminating Spatial and Temporal Relevance in Deep Taylor Decompositions for Explainable Activity Recognition." pith.science (2026). https://pith.science/paper/X2KH2TX4
@misc{pith2026190801536,
author = {Pith},
title = {Pith review of: Discriminating Spatial and Temporal Relevance in Deep Taylor Decompositions for Explainable Activity Recognition},
year = {2026},
howpublished = {\url{https://pith.science/paper/X2KH2TX4}},
note = {Machine review of arXiv:1908.01536}
}
read the original abstract
Current techniques for explainable AI have been applied with some success to image processing. The recent rise of research in video processing has called for similar work n deconstructing and explaining spatio-temporal models. While many techniques are designed for 2D convolutional models, others are inherently applicable to any input domain. One such body of work, deep Taylor decomposition, propagates relevance from the model output distributively onto its input and thus is not restricted to image processing models. However, by exploiting a simple technique that removes motion information, we show that it is not the case that this technique is effective as-is for representing relevance in non-image tasks. We instead propose a discriminative method that produces a na\"ive representation of both the spatial and temporal relevance of a frame as two separate objects. This new discriminative relevance model exposes relevance in the frame attributed to motion, that was previously ambiguous in the original explanation. We observe the effectiveness of this technique on a range of samples from the UCF-101 action recognition dataset, two of which are demonstrated in this paper.
Figures
Reference graph
Works this paper leans on
-
[1]
On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation
[Bach et al., 2015] Sebastian Bach, Alexander Binder, Gr´egoire Montavon, Frederick Klauschen, Klaus-Robert M¨uller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one, 10(7):e0130140,
work page 2015
-
[7]
3D convolutional neural networks for human action recognition
[Ji et al., 2013] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3D convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence, 35(1):221–231,
work page 2013
-
[9]
Explaining nonlinear classification deci- sions with deep taylor decomposition
[Montavon et al., 2017] Gr´egoire Montavon, Sebastian La- puschkin, Alexander Binder, Wojciech Samek, and Klaus- Robert M ¨uller. Explaining nonlinear classification deci- sions with deep taylor decomposition. Pattern Recogni- tion, 65:211–222,
2017
-
[10]
Grad-cam: Visual explana- tions from deep networks via gradient-based localization
[Selvaraju et al., 2017] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explana- tions from deep networks via gradient-based localization. In Proceedings of the IEEE International Conference on Computer Vision, pages 618–626,
work page 2017
-
[11]
Deep inside convolutional networks: Visualising image classification models and saliency maps
[Simonyan et al., 2013] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps. arXiv preprint arXiv:1312.6034,
arXiv 2013
-
[12]
[Soomro et al., 2012] Khurram Soomro, Amir Roshan Za- mir, and Mubarak Shah. UCF101: A dataset of 101 hu- man actions classes from videos in the wild.arXiv preprint arXiv:1212.0402,
arXiv 2012
-
[14]
Saliency tubes: Visual ex- planations for spatio-temporal convolutions,
[Stergiou et al., 2019] Alexandros Stergiou, Georgios Ka- pidis, Grigorios Kalliatakis, Christos Chrysoulas, Remco Veltkamp, and Ronald Poppe. Saliency tubes: Visual ex- planations for spatio-temporal convolutions,
work page 2019
-
[16]
Top-down neural attention by excitation backprop
[Zhang et al., 2018] Jianming Zhang, Sarah Adel Bargal, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down neural attention by excitation backprop. International Journal of Computer Vision , 126(10):1084–1102,
work page 2018
Show all 17 references
-
[17]
Learning deep features for discriminative localization
[Zhou et al., 2016] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceed- ings of the IEEE conference on computer vision and pat- tern recognition, pages 2921–2929, 2016
2016
-
[2009]
Can spatiotemporal 3D CNNs retrace the history of 2D CNNs and ImageNet? In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June
[Hara et al., 2018] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3D CNNs retrace the history of 2D CNNs and ImageNet? In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June
2018
-
[2010]
Quo vadis, action recognition? a new model and the kinetics dataset
[Carreira and Zisserman, 2017] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 6299–6308,
2017
-
[2012]
Interpretable human action recogni- tion in compressed domain
[Srinivasan et al., 2017] Vignesh Srinivasan, Sebastian La- puschkin, Cornelius Hellge, Klaus-Robert M ¨uller, and Wojciech Samek. Interpretable human action recogni- tion in compressed domain. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (I...
2017
-
[2013]
The kinetics human action video dataset
[Kay et al., 2017] Will Kay, Joao Carreira, Karen Si- monyan, Brian Zhang, Chloe Hillier, Sudheendra Vijaya- narasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950,
2017 arXiv
-
[2015]
How to explain individual classifi- cation decisions
[Baehrens et al., 2010] David Baehrens, Timon Schroeter, Stefan Harmeling, Motoaki Kawanabe, Katja Hansen, and Klaus-Robert M ˜Aˇzller. How to explain individual classifi- cation decisions. Journal of Machine Learning Research , 11(Jun):1803–1831,
2010
-
[2017]
Visualizing higher-layer features of a deep network
[Erhan et al., 2009] Dumitru Erhan, Yoshua Bengio, Aaron Courville, and Pascal Vincent. Visualizing higher-layer features of a deep network. a,
2009
-
[2018]
Evaluating feature impor- tance estimates
[Hooker et al., 2018] Sara Hooker, Dumitru Erhan, PieterJan Kindermans, and Been Kim. Evaluating feature impor- tance estimates. arXiv preprint arXiv:1806.10758,
2018 arXiv
-
[2019]
Learning spa- tiotemporal features with 3D convolutional networks
[Tran et al., 2015] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spa- tiotemporal features with 3D convolutional networks. In Proceedings of the IEEE international conference on com- puter vision, pages 4489–4497,
2015
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.