CRITS: Convolutional Rectifier for Interpretable Time Series Classification

Alejandro Kuratomi; Diego Garc\'ia P\'erez; Guilherme Dinis Chaliane Junior; Tony Lindgren; Zed Lee

REVIEW 4 major objections 5 minor 33 references

CRITS: Convolutional Rectifier for Interpretable Time Series Classification

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read CRITS is a convolutional time-series classifier whose local explanations are the exact per-sample weights the model uses to make its decision, extracted without gradients, random perturbations, or upsampling.

desk verdict A genuinely exact local explainer for a simple conv time-series model, but the printed equations don't yet support the exactness claim and the classification numbers are selected on the test split; worth a serious referee if those are fixed. read the letter →

arxiv 2506.12042 v1 pith:2UBSTHMG submitted 2025-05-24 cs.LG cs.AIstat.ML

Alejandro Kuratomi , Zed Lee , Guilherme Dinis Chaliane Junior , Tony Lindgren , Diego García Pérez This is my paper · ORCID

classification cs.LGcs.AIstat.ML

keywords interpretabilitytimeseriesclassificationrectifiernetworksconvolutionalneuralsaliencymapslocalexplanationsReLUactivationsintrinsic

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

CRITS is a convolutional classifier for time series that is interpretable by construction: because the final layers are a rectifier (ReLU-only) network, the weights each input selects can be read off exactly, and the convolutional and max-pooling layers are traced back to the input by deconvolution and unpooling. The result is a local saliency map that is not an approximation, a gradient, or an upsampled heatmap, but the actual per-sample weights the model used for the decision. Tested on six public time-series datasets, CRITS achieves F1 scores close to the strong ROCKET baseline, and its explanations compare favorably with post-hoc methods on alignment, input-sensitivity, and sparsity. The paper's contribution is showing that exact intrinsic explanations and competitive classification accuracy can coexist in a convolutional time-series architecture.

What carries the argument

The load-bearing mechanism is the unwrapping identity for rectifier networks: an algorithm that turns a ReLU-only network into an equivalent linear model for a given input by recording each neuron's activation pattern. For a sample $x_i$, the rectifier's output is exactly $z_i = w_i^K \text{maxpool}(F_i) + b_i^K$. CRITS extends this by storing the max-pooling argmax for each filter and using a deconvolution/unpooling step to project the kernel weights at those positions into input space, producing per-filter masks $M_k^i$. The final input weights are the mask-weighted sum $w_i = \sum_k M_k^i \cdot w_k^i$, which makes Eq. 4 an equality rather than an approximation.

What would settle it

Train a CRITS model and, for a set of test instances, compare the actual pre-sigmoid output of the network against the reconstruction $z_i = w_i x_i + b_i^K$ computed from the extracted weights; any mismatch, or any instance with tied max-pooling argmax, would falsify the exactness claim.

Watch

Extended reading notes

Core claim

The central claim is that a convolutional time-series classifier can be built so that, for every input instance $x_i$, the pre-sigmoid output obeys the exact linear identity $z_i = w_i x_i + b_i^K$, where $w_i$ is the per-sample input weight vector recovered by a two-step unwrap: first converting the rectifier network into its equivalent linear form $w_i^K$, then projecting each max-pooling winner's kernel weights back to the input space as a mask $M_k^i$ and summing over filters. The resulting relevance map $r_i = w_i \odot x_i$ has the same resolution as the input, so no upsampling is needed, and it is obtained without gradients, noise, or random perturbations. The authors support this with experiments on three univariate and three multivariate datasets, reporting classification performance comparable to ROCKET and explanation-quality metrics where CRITS's intrinsic explanations generally match or exceed the post-hoc baselines.

Load-bearing premise

The exact reconstruction in Eq. 4 depends on the convolutional layer contributing no unaccounted bias and on each filter having a unique max-pooling argmax; the paper does not specify whether the conv layer has a bias or an activation function, so if conv biases exist the recovered weights are only the local slope, not the full decision function.

Editorial extensions

If this is right

Local explanations come free with the architecture: no post-hoc explainer, no gradient computation, no random perturbation, and no upsampling, so the saliency maps are sharp at native input resolution.
The same architecture handles univariate and multivariate time series, with the max-pooling argmax recording which kernel position fired, so each explanation is sparse and tied to the actual input region used for the decision.
Classification performance is comparable to ROCKET on the tested datasets, with F1 scores of 0.99–1.00 on GunPoint, Strawberry, and Blink, and within 0.02–0.04 on SharePriceIncrease and Heartbeat.
The alignment, sensitivity, and sparsity measures indicate that intrinsic explanations behave well against perturbation-based and gradient-based post-hoc methods, with CRITS's input-sensitivity increasing appropriately as input noise grows.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the exactness identity survives convolutional biases or non-unique max-pooling argmaxes, the same unwrapping idea could extend to other layer types, such as average pooling, strided convolutions, or residual connections, with a generalized bookkeeping of the linear path.
The exact per-sample weights make CRITS a natural source of ground-truth explanations for benchmarking post-hoc explainers, or for generating counterfactuals by perturbing high-weight regions with guaranteed control over the logit.
CRITS's sensitivity behavior suggests the pre-sigmoid decision is locally linear across the whole input, so the same machinery could be used to estimate per-sample Lipschitz constants or to detect adversarial directions.
A testable extension would be to replace global max-pooling with k-max pooling or to add a second convolutional layer while keeping ReLU-only activations; the reconstruction would need a revised mask, but the core identity may still hold.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague, here is my read on Kuratomi et al.'s CRITS. The genuinely new piece is the combination: Sudjianto's unwrapper for fully-connected ReLU networks, attached to a single conv layer plus max-pooling, with Zeiler-Fergus-style stored-max unpooling to project the RN weights back to the input. The result is a classifier whose local explanation is exactly the per-instance weight vector of the model -- no gradients, no perturbations, no upsampling. For time series, that is a useful niche: a simple model that gives sharp, exact saliency maps. The paper is honest that the components are known; the contribution is the integration and the evaluation on six public datasets. That is real, though moderate in scope.

What it does well: the explanation evaluation is reasonable -- alignment under four perturbation types, input-sensitivity under increasing noise, and sparsity. The comparison against DeepSHAP, GradientSHAP, and SmoothGRAD is fair, and the qualitative examples in Appendix B show CRITS maps are more focused. Code is available, and the classification results are close to ROCKET, a strong baseline.

Now the soft spots, proportional to how soft they are. The math is not as clean as it should be. Eq. 1 is malformed as printed -- x_i appears inside the product defining w_i^K -- so the key identity is hard to check. More importantly, Eq. 4 claims z_i = w_i x_i + b_i^K, but this holds only if every additive constant in the forward path is folded into b_i^K. The paper never says whether the conv layer has a bias term or a ReLU activation. In most frameworks, Conv1D defaults to bias=True, and if so, the max-pooling output includes the bias, which then appears in the logit as a sum over the unwrapped weights. If the conv layer has ReLU, inactive channels contribute zero slope and the mask construction needs a rectification rule. The paper doesn't address either case. The stress-test note is right: the exactness guarantee, which is the whole point, is not supported by the equations as printed. This is fixable -- set conv bias to zero, or include it in the recursion, and state the activation -- but it needs to be done.

Second, the hyperparameter search in Appendix A selects on the test set: an 80/20 split, best config per dataset, F1 reported on that same test split, with no variance and no repeated runs. That inflates the classification numbers and makes the ROCKET comparison unfair. A proper validation split or cross-validation is needed. Max-pooling ties are also not discussed; if two positions share the max, the stored index isn't unique. Minor in practice, but part of the exactness story.

Overall, the central idea holds up. The paper is clear and the authors are honest about limitations. It deserves a serious referee, but the referee should insist on the fixes above. I would cite it if I worked on interpretable time series models, and I would bring it to a reading group. My verdict: conditional accept at a workshop, major revision at a journal.

Referee Report

4 major / 5 minor

Summary. The paper proposes CRITS, a time series classifier consisting of a convolutional layer, a global max-pooling layer, and a fully-connected rectifier network, designed to provide intrinsic local explanations. The explanation for a given input is obtained by unwrapping the rectifier network into per-sample linear weights and then projecting the selected convolutional kernel weights back to the input space through stored max-pooling indices. The authors claim that the resulting relevance values are the exact per-sample weights used by the model, with no gradients, random perturbations, or upsampling. The method is evaluated on six time series datasets against ROCKET for classification performance and against DeepSHAP, GradientSHAP, and SmoothGRAD for explanation alignment, input-sensitivity, and sparsity.

Significance. If the exactness claim can be made rigorous, CRITS is a genuinely useful contribution: it extends the rectifier-network unwrapper of Sudjianto et al. to convolutional time series architectures, removing the need for post-hoc gradient or perturbation-based explainers, and it provides sharp, input-aligned saliency maps. The code is publicly released, the experimental comparison uses standard public datasets, and the three explanation-quality metrics are reasonable proxies for fidelity, stability, and simplicity. However, the central theoretical claim is currently not fully supported as written, and the empirical evaluation has methodological gaps that weaken the reported performance comparison. The contribution is therefore promising but needs nontrivial revision before it can be accepted.

major comments (4)

[§3.3, Eq. (1)] Equation (1) is malformed: the right-hand side includes w0 x_i inside the product defining w_i^K, making the left-hand side dimensionally inconsistent with a vector of K max-pooling weights. As printed, this equation cannot define the local Jacobian used in Eq. (4). Please provide the correct unwrapping formula from Sudjianto et al., with all symbols defined, and ensure that w_i^K has the required interpretation.
[§3.3, Eqs. (2)–(4)] The exact reconstruction z_i = w_i x_i + b_i^K depends on every additive constant and nonlinearity in the forward path being accounted for. The manuscript never states whether the convolutional layer has a bias term or an activation function. If the convolutional layer has biases c_k (the common Conv1D default), then the k-th max-pooling output is max_t conv_k(x)_t + c_k, and the logit contains an additional term sum_k w_i,k^K c_k that is not present in Eq. (2). If the convolutional layer also applies a ReLU activation, the local slope is zero on inactive channels and the mask construction requires an explicit rectification rule. Please specify the exact architecture used, including bias and activation settings, and either revise Eqs. (2)–(4) to include all contributions or state explicitly which components are disabled.
[§3.3, max-pooling and Fig. 2] The paper does not discuss ties in global max-pooling. If two or more positions in a feature map attain the same maximum value, the stored argmax is not unique, and the mask M_i^k is therefore not uniquely defined. This makes the extracted weights w_i and the resulting explanation ambiguous, directly affecting the claimed exactness. Please state the tie-breaking rule and explain how the exactness claim is maintained under ties.
[Table 2 and Appendix A] The classification results are reported as the best F1-score on the test set after a random search over 500 configurations, with no separate validation procedure and no variance estimate. Selecting the final model on the test split invalidates the comparison with ROCKET as a fair assessment of classification performance. Please report means and standard deviations over multiple train/test splits, or use a nested validation scheme, and indicate whether the ROCKET baseline was evaluated on the same held-out splits.

minor comments (5)

[§3.3, Eq. (2)] The bias notation is inconsistent: Eq. (2) uses b_n and b_L, while Eq. (4) uses b_i^K, and the relationship between these quantities is not defined. Please unify the notation.
[§3.3, Eq. (4)] The product w_i x_i is ambiguous because w_i and x_i are both defined as T × m objects. Please clarify whether Eq. (4) denotes a dot product summing over all elements, an element-wise product followed by summation, or something else.
[Fig. 1] Figure 1 contains a label 'ReLU kernel', but the text does not clearly state whether the convolutional layer itself applies a ReLU activation. Please define the activation of the convolutional layer explicitly in Section 3.2.
[§4.2, Fig. 4] The text states that the initial input-sensitivity of CRITS is 'at least 0.0000001'; please use scientific notation (10^-7) and report the actual values from the experiments.
[§3.4, Sparsity] The sparsity threshold of 0.01 is arbitrary, and no sensitivity analysis is provided. Please justify the threshold or show that the qualitative results are robust to its choice.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: CRITS' exact-relevance claim is an analytic property of its piecewise-linear architecture, not a fitted constant relabeled as a prediction.

full rationale

The paper's central claim—that CRITS extracts the exact per-sample weights used by the model—follows from the forward pass structure, not from a definitional equivalence or from the paper's own fitted values. The unwrapper of [25] is an external result by different authors, and it is used only for the fully-connected rectifier part; the extension to convolution/pooling is the paper's own derivation in Eqs. 1-3, in which w_i is obtained by projecting the stored max-pooling argmax locations back to the input space and weighting them by the RN weights. If the convolutional layer has no hidden bias/activation (or those terms are folded into b_i^K), Eq. 4 is the standard local linearization of a ReLU network, so the relevance map r_i = w_i ⊙ x_i is genuinely the additive decomposition of the pre-sigmoid logit. The alignment metric then measures what the architecture guarantees by construction, which is a sanity check rather than a fitted prediction. The classification comparison with ROCKET and the explanation comparison with SmoothGRAD, DeepSHAP, and GradientSHAP are external benchmarks. Concerns about unspecified conv bias/activation and the Eq. 1 typo are correctness/verification gaps, not circularity. No load-bearing self-citations exist: the two self-references ([11], [12]) are only motivation/related-work examples. Hence the derivation chain is self-contained and the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim relies on the Sudjianto unwrapper as an external mathematical tool, on unstated details about the conv layer, on uniqueness of max-pooling argmax, and on tuned hyperparameters per dataset. No new physical entities are introduced.

free parameters (4)

Kernel height h = 16, 30, 15, 8 per dataset (width equals number of variables m)
Chosen by random search in Appendix A; determines the patch size projected into the input-space masks in Eq. 3.
Number of convolutional kernels K = 144, 256, 240, 32, 48, 176 per dataset
Random-search hyperparameter; the number of masks summed in Eq. 3.
Rectifier network depth and width = Varies per dataset, e.g., (160, 120, 40, 30, 190) for GunPoint
Random-search hyperparameters; they enter the unwrapper product in Eq. 1.
Sparsity threshold = 0.01
Hand-set cutoff in Eq. 7 for defining relevant features; changing it changes reported sparsity scores.

assumptions (4)

standard math Sudjianto et al. unwrapper correctly recovers local linear coefficients of a ReLU network.
Used as the starting point in Section 3.3, Eqs. 1-2; the paper does not re-derive it.
domain assumption Global max-pooling records a unique argmax per filter and the local linear region around each test input has positive measure.
Needed so the deconvolution in Eq. 3 and Fig. 2 yields the exact gradient rather than a subgradient choice.
domain assumption The convolutional layer has no nonlinearity and either no bias or its bias is included in b_i^K.
The text describes conv, max-pool, then a rectifier network; whether ReLU or bias exists at the conv layer is not stated, yet Eq. 4 requires it.
domain assumption Datasets and labels are correct and the 80/20 split used for hyperparameter selection gives a valid estimate.
Performance claims in Table 2 depend on this; Appendix A selects the best config using the test split.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CRITS: Convolutional Rectifier for Interpretable Time Series Classification." pith.science (2026). https://pith.science/paper/2UBSTHMG

@misc{pith2026250612042,
  author       = {Pith},
  title        = {Pith review of: CRITS: Convolutional Rectifier for Interpretable Time Series Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2UBSTHMG}},
  note         = {Machine review of arXiv:2506.12042}
}

read the original abstract

Several interpretability methods for convolutional network-based classifiers exist. Most of these methods focus on extracting saliency maps for a given sample, providing a local explanation that highlights the main regions for the classification. However, some of these methods lack detailed explanations in the input space due to upscaling issues or may require random perturbations to extract the explanations. We propose Convolutional Rectifier for Interpretable Time Series Classification, or CRITS, as an interpretable model for time series classification that is designed to intrinsically extract local explanations. The proposed method uses a layer of convolutional kernels, a max-pooling layer and a fully-connected rectifier network (a network with only rectified linear unit activations). The rectified linear unit activation allows the extraction of the feature weights for the given sample, eliminating the need to calculate gradients, use random perturbations and the upscale of the saliency maps to the initial input space. We evaluate CRITS on a set of datasets, and study its classification performance and its explanation alignment, sensitivity and understandability.

Figures

Figures reproduced from arXiv: 2506.12042 by the authors.

**Figure 4.** The input-sensitivity of the explanations for each datasets. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗

**Figure 5.** The sparsity of the explanations for each dataset and method. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

Figures from the paper (2 more)

**Figure 6.** Figure 6: Relevance values, r i , for the univariate datasets [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]

**Figure 7.** Figure 7: Relevance values, r i , for the multivariate datasets [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 21 canonical work pages

[1]

In: IJCAI-17 workshop on explainable AI (XAI)

Biran, O., Cotton, C.: Explanation and justification in machine learning: A survey. In: IJCAI-17 workshop on explainable AI (XAI). vol. 8, pp. 8–13 (2017)

work page 2017
[2]

Burstein, J., Doran, C., Solorio, T.: Proceedings of the 2019 conference of the north american chapter of the association for computational linguistics: Human language technologies, volume 1 (long and short papers). In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Compu- tational Linguistics: Human Language Techn...

work page 2019
[3]

Simon and Schuster (2021)

Chollet, F.: Deep learning with Python. Simon and Schuster (2021)

work page 2021
[4]

Data Mining and Knowledge Discovery34(5), 1454–1495 (2020)

Dempster, A., Petitjean, F., Webb, G.I.: Rocket: exceptionally fast and accurate time series classification using random convolutional kernels. Data Mining and Knowledge Discovery34(5), 1454–1495 (2020)

2020
[5]

Dorle, A., Li, F., Song, W., Li, S.: Learning discriminative virtual sequences for timeseriesclassification.In:Proceedingsofthe29thACMInternationalConference on Information & Knowledge Management. pp. 2001–2004 (2020)

work page 2020
[6]

IEEE Transactions on Artificial Intel- ligence 1(1), 47–61 (2020)

Gupta, A., Gupta, H.P., Biswas, B., Dutta, T.: Approaches and applications of early classification of time series: A review. IEEE Transactions on Artificial Intel- ligence 1(1), 47–61 (2020)

work page 2020
[7]

In: Advances in Knowledge Discovery and Data Mining: 23rd Pacific-Asia Conference, PAKDD 2019, Macau, China, April 14-17, 2019, Proceedings, Part III 23

Hsu, E.Y., Liu, C.L., Tseng, V.S.: Multivariate time series early classification with interpretability using deep learning and attention mechanism. In: Advances in Knowledge Discovery and Data Mining: 23rd Pacific-Asia Conference, PAKDD 2019, Macau, China, April 14-17, 2019, Proceedings, Part III 23. pp. 541–553. Springer (2019)

work page 2019
[8]

Data mining and knowledge discovery 33(4), 917–963 (2019)

Ismail Fawaz, H., Forestier, G., Weber, J., Idoumghar, L., Muller, P.A.: Deep learn- ing for time series classification: a review. Data mining and knowledge discovery 33(4), 917–963 (2019)

2019

Show all 33 references

[9]

In: Proceedings of the 25th ACM SIGKDD International conference on knowledge discovery & Data Mining

Jia, Y., Bailey, J., Ramamohanarao, K., Leckie, C., Houle, M.E.: Improving the quality of explanations with local embedding perturbations. In: Proceedings of the 25th ACM SIGKDD International conference on knowledge discovery & Data Mining. pp. 875–884 (2019)

2019
[10]

Ieee Access8, 121033–121049 (2020)

Kiangala, K.S., Wang, Z.: An effective predictive maintenance framework for con- veyor motors using dual time-series imaging and convolutional neural network in an industry 4.0 environment. Ieee Access8, 121033–121049 (2020)

2020
[11]

In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases

Kuratomi, A., Lindgren, T., Papapetrou, P.: Prediction of global navigation satel- lite system positioning errors with guarantees. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 562–578. Springer (2020)

2020
[12]

In: Discovery Science: 25th International Conference, DS 2022, Montpellier, France, October 10–12, 2022, Proceedings

Kuratomi, A., Miliou, I., Lee, Z., Lindgren, T., Papapetrou, P.: Juice: Justified counterfactual explanations. In: Discovery Science: 25th International Conference, DS 2022, Montpellier, France, October 10–12, 2022, Proceedings. pp. 493–508. Springer (2022) 16 Kuratomi et al

2022
[13]

Advances in neural information processing systems30 (2017)

Lundberg, S.M., Lee, S.I.: A unified approach to interpreting model predictions. Advances in neural information processing systems30 (2017)

2017
[14]

In: Fourth International Workshop on Knowledge Discovery, Knowledge Management and Decision Support

Maletzke, A.G., Lee, H.D., Batista, G.E., Rezende, S.O., Machado, R.B., Voltolini, R.F., Maciel, J.N., Silva, F.: Time series classification using motifs and character- istics extraction: a case study on ecg databases. In: Fourth International Workshop on Knowledge Discovery, ...
[15]

Artificial intelligence267, 1–38 (2019)

Miller, T.: Explanation in artificial intelligence: Insights from the social sciences. Artificial intelligence267, 1–38 (2019)

2019
[16]

Molnar, C.: Interpretable machine learning: A guide for making black-box models explainable (2021), https://christophm.github.io/interpretable-ml-book/ limo.html

2021
[17]

why should i trust you?

Ribeiro, M.T., Singh, S., Guestrin, C.: "why should i trust you?" explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD interna- tional conference on knowledge discovery and data mining. pp. 1135–1144 (2016)

2016
[18]

arXiv preprint arXiv:2104.00950 (2021)

Rojat, T., Puget, R., Filliat, D., Del Ser, J., Gelin, R., Díaz-Rodríguez, N.: Ex- plainable artificial intelligence (xai) on timeseries data: A survey. arXiv preprint arXiv:2104.00950 (2021)

2021 arXiv
[19]

Nature Machine Intelligence1(5), 206–215 (2019)

Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence1(5), 206–215 (2019)

2019
[20]

Data Mining and Knowledge Discovery35(2), 401– 449 (2021)

Ruiz, A.P., Flynn, M., Large, J., Middlehurst, M., Bagnall, A.: The great multi- variate time series classification bake off: a review and experimental evaluation of recent algorithmic advances. Data Mining and Knowledge Discovery35(2), 401– 449 (2021)

2021
[21]

In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW)

Schlegel, U., Arnout, H., El-Assady, M., Oelke, D., Keim, D.A.: Towards a rigor- ous evaluation of xai methods on time series. In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). pp. 4197–4201. IEEE (2019)

2019
[22]

In: Proceedings of the IEEE international conference on computer vision

Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad- cam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618–626 (2017)

2017
[23]

In: International conference on machine learn- ing

Shrikumar, A., Greenside, P., Kundaje, A.: Learning important features through propagating activation differences. In: International conference on machine learn- ing. pp. 3145–3153. PMLR (2017)

2017
[24]

arXiv preprint arXiv:1706.03825 (2017)

Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: Smoothgrad: re- moving noise by adding noise. arXiv preprint arXiv:1706.03825 (2017)

2017 arXiv
[25]

arXiv preprint arXiv:2011.04041 (2020)

Sudjianto, A., Knauth, W., Singh, R., Yang, Z., Zhang, A.: Unwrapping the black box of deep relu networks: interpretability, diagnostics, and simplification. arXiv preprint arXiv:2011.04041 (2020)

2020 arXiv
[26]

IEEE Access (2022)

Theissler, A., Spinnato, F., Schlegel, U., Guidotti, R.: Explainable ai for time series classification: A review, taxonomy and research directions. IEEE Access (2022)

2022
[27]

Journal of Computational Science58, 101539 (2022)

Veerappa, M., Anneken, M., Burkart, N., Huber, M.F.: Validation of xai explana- tions for multivariate time series classification in the maritime domain. Journal of Computational Science58, 101539 (2022)

2022
[28]

arXiv preprint arXiv:2106.12569 (2021)

Widdicombe, A., Julier, S.J.: Gradient-based interpretability methods and bina- rized neural networks. arXiv preprint arXiv:2106.12569 (2021)

2021 arXiv
[29]

In: Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13

Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. pp. 818–833. Springer (2014) CRITS 17

2014
[30]

In: Proceedings of the AAAI Conference on Artificial Intelligence

Zhang, X., Gao, Y., Lin, J., Lu, C.T.: Tapnet: Multivariate time series classification with attentional prototypical network. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 6845–6852 (2020)

2020
[31]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., Torralba, A.: Learning deep features for discriminative localization. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2921–2929 (2016)

2016
[32]

In: 2021 International Joint Conference on Neural Networks (IJCNN)

Zhou, L., Ma, C., Shi, X., Zhang, D., Li, W., Wu, L.: Salience-cam: Visual explana- tions from convolutional neural networks via salience score. In: 2021 International Joint Conference on Neural Networks (IJCNN). pp. 1–8. IEEE (2021) 18 Kuratomi et al. A Hyperparameter tuning ...

2021
[329]

Atlantis Press (2013)

2013

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

In: IJCAI-17 workshop on explainable AI (XAI)

Biran, O., Cotton, C.: Explanation and justification in machine learning: A survey. In: IJCAI-17 workshop on explainable AI (XAI). vol. 8, pp. 8–13 (2017)

work page 2017

[2] [2]

Burstein, J., Doran, C., Solorio, T.: Proceedings of the 2019 conference of the north american chapter of the association for computational linguistics: Human language technologies, volume 1 (long and short papers). In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Compu- tational Linguistics: Human Language Techn...

work page 2019

[3] [3]

Simon and Schuster (2021)

Chollet, F.: Deep learning with Python. Simon and Schuster (2021)

work page 2021

[4] [4]

Data Mining and Knowledge Discovery34(5), 1454–1495 (2020)

Dempster, A., Petitjean, F., Webb, G.I.: Rocket: exceptionally fast and accurate time series classification using random convolutional kernels. Data Mining and Knowledge Discovery34(5), 1454–1495 (2020)

2020

[5] [5]

Dorle, A., Li, F., Song, W., Li, S.: Learning discriminative virtual sequences for timeseriesclassification.In:Proceedingsofthe29thACMInternationalConference on Information & Knowledge Management. pp. 2001–2004 (2020)

work page 2020

[6] [6]

IEEE Transactions on Artificial Intel- ligence 1(1), 47–61 (2020)

Gupta, A., Gupta, H.P., Biswas, B., Dutta, T.: Approaches and applications of early classification of time series: A review. IEEE Transactions on Artificial Intel- ligence 1(1), 47–61 (2020)

work page 2020

[7] [7]

In: Advances in Knowledge Discovery and Data Mining: 23rd Pacific-Asia Conference, PAKDD 2019, Macau, China, April 14-17, 2019, Proceedings, Part III 23

Hsu, E.Y., Liu, C.L., Tseng, V.S.: Multivariate time series early classification with interpretability using deep learning and attention mechanism. In: Advances in Knowledge Discovery and Data Mining: 23rd Pacific-Asia Conference, PAKDD 2019, Macau, China, April 14-17, 2019, Proceedings, Part III 23. pp. 541–553. Springer (2019)

work page 2019

[8] [8]

Data mining and knowledge discovery 33(4), 917–963 (2019)

Ismail Fawaz, H., Forestier, G., Weber, J., Idoumghar, L., Muller, P.A.: Deep learn- ing for time series classification: a review. Data mining and knowledge discovery 33(4), 917–963 (2019)

2019