Abstract
Various types of saliency methods have been proposed for explaining black-box classification. In image applications, this means highlighting the part of the image that is most relevant for the current decision.
Unfortunately, the different methods may disagree and it can be hard to quantify how representative and faithful the explanation really is. We observe however that several of these methods can be seen as edge cases of a single, more general procedure based on finding a particular path through the classifier’s domain. This offers additional geometric interpretation to the existing methods.
We demonstrate furthermore that ablation paths can be directly used as a technique of its own right. This is able to compete with literature methods on existing benchmarks, while giving more fine-grained information and better opportunities for validation of the explanations’ faithfulness.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
For instance in the theory of distributions.
- 2.
The exact definitions of \(\varPi _\text {sat}\) and \(\varPi _\text {pinch}\) are largely arbitrary, what matters are their attractive fixpoints; see Fig. 2
- 3.
This still requires also dedicated mask regularity, as otherwise the adversarial contributions overwhelm and the result appears as mere noise.
- 4.
The pointing game is generally used under the assumption that neither the classifier nor the saliency method have any direct training knowledge about the position annotations, i.e. it is not a test of how well a trained task generalizes but of an extrinsic notion of saliency.
- 5.
By “first 1000” we mean the 1000 images with the lowest ids. Note that the VOC and COCO sets are in random order, so that this should be a reasonably representative and reproducible selection. Comparing the score of Grad-CAM to the one on the full datasets confirms this.
The astute reader may notice that on the other hand, with only the first 100 images the results are systematically worse. This is less due to these images being more difficult, than artifact of the way the TorchRay benchmark gathers the results: specifically, it counts success rate for each class separately and averages in the end, but rates classes that are not even present in the smaller subset as 0% success.
- 6.
The authors in [5] emphasize that their method avoids hyperparameters, yet their examples rely on no fewer than 5 hard-coded number constants.
- 7.
This can be interpreted as applying prior knowledge of the location of objects in the dataset. However, there are also examples of objects close to the boundary. The window post-processing prevents these from being properly localised, which is why the top pointing-game score is still lower.
- 8.
It is easy to come up with other algorithms for monotonising a (discretised) function. One could simply sort the array, but that is not optimal with respect to any of the usual function norms; or clip the derivatives to be nonnegative and then rescale the entire function, but that is not robust against noise perturbations.
- 9.
Note that the optimum is not necessarily unique.
References
Adebayo, J., Gilmer, J., Muelly, M., Goodfellow, I., Hardt, M., Kim, B.: Sanity checks for saliency maps. In: Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., Garnett, R. (eds.) Advances in Neural Information Processing Systems, vol. 31, pp. 9505–9515. Curran Associates, Inc. (2018). https://papers.nips.cc/paper/8160-sanity-checks-for-saliency-maps.pdf
Ancona, M., Ceolini, E., Öztireli, C., Gross, M.: Towards better understanding of gradient-based attribution methods for deep neural networks. CoRR (2017)
Chockler, H., Kroening, D., Sun, Y.: Explanations for occluded images. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 1234–1243 (2021)
Dabkowski, P., Gal, Y.: Real time image saliency for black box classifiers. In: Guyon, I., et al. (eds.) Advances in Neural Information Processing Systems, vol. 30. Curran Associates, Inc. (2017). https://proceedings.neurips.cc/paper/2017/file/0060ef47b12160b9198302ebdb144dcf-Paper.pdf
Fong, R., Patrick, M., Vedaldi, A.: Understanding deep networks via extremal perturbations and smooth masks, pp. 2950–2958 (2019)
Fong, R.C., Vedaldi, A.: Interpretable explanations of black boxes by meaningful perturbation (2017)
Kindermans, P.-J., et al.: The (un)reliability of saliency methods. In: Samek, W., Montavon, G., Vedaldi, A., Hansen, L.K., Müller, K.-R. (eds.) Explainable AI: Interpreting, Explaining and Visualizing Deep Learning. LNCS (LNAI), vol. 11700, pp. 267–280. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-28954-6_14
Koh, P.W., Liang, P.: Understanding black-box predictions via influence functions. 70, 1885–1894 (2017). https://proceedings.mlr.press/v70/koh17a.html
Petsiuk, V., Das, A., Saenko, K.: Rise: randomized input sampling for explanation of black-box models. CoRR (2018)
Rudin, C.: Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat. Mach. Intell. 1(5), 206–215 (2019). https://doi.org/10.1038/s42256-019-0048-x
Sagemüller, J.: Pytorch implementation of the ablation path saliency method (2023). https://github.com/leftaroundabout/ablation-paths-pytorch
Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad-CAM: visual explanations from deep networks via gradient-based localization (2017)
Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: visualising image classification models and saliency maps. CoRR (2013)
Sturmfels, P., Lundberg, S., Lee, S.I.: Visualizing the impact of feature attribution baselines. Distill (2020). https://doi.org/10.23915/distill.00022. https://distill.pub/2020/attribution-baselines
Sundararajan, M., Taly, A., Yan, Q.: Axiomatic attribution for deep networks. 70, 3319–3328 (2017). https://proceedings.mlr.press/v70/sundararajan17a.html
Szegedy, C., et al.: Intriguing properties of neural networks (2014)
Vedaldi, A.: Understanding deep networks via extremal perturbations and smooth masks (2019). https://github.com/facebookresearch/TorchRay
Weller, A.: Transparency: motivations and challenges. In: Samek, W., Montavon, G., Vedaldi, A., Hansen, L.K., Müller, K.-R. (eds.) Explainable AI: Interpreting, Explaining and Visualizing Deep Learning. LNCS (LNAI), vol. 11700, pp. 23–40. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-28954-6_2
Zhang, J., Bargal, S.A., Lin, Z., Brandt, J., Shen, X., Sclaroff, S.: Top-down neural attention by excitation backprop. Int. J. Comput. Vis. 126(10), 1084–1102 (2017). https://doi.org/10.1007/s11263-017-1059-x
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Appendices
Appendix A Canonical Time Reparametrisation
The function \(m :[0,1] \rightarrow \mathbb {R}\) defined by \(m(t) {:}{=}\int _{\varOmega } \varphi (t)\) is increasing and goes from zero to one (since we assume that \(\int _{\varOmega } 1 = 1\)).
Note first that if \(m(t_1) = m(t_2)\), then \(\varphi (t_1) = \varphi (t_2)\) from the monotonicity property. Indeed, supposing for instance that \(t_1 \le t_2\), and defining the element \(\theta {:}{=}\varphi (t_2) - \varphi (t_1) \) we see that on the one hand \(\int _{\varOmega } \theta = 0\), on the other hand, \(\theta \ge 0\), so \(\theta = 0\) and thus \(\varphi (t_1) = \varphi (t_2)\).
Now, define . Pick \(s \in [0,1]\).
If \(s \in \textsf{M}\) we define \(\psi (s) {:}{=}\varphi (t)\) where \(m(t) = s\) (and this does not depend on which \(t\) fulfills \(m(t) = s\) from what we said above). We remark that \(\int _{\varOmega }\psi (s) = \int _{\varOmega }\varphi (t) = m(t) = s\).
Now suppose that \(s \not \in \textsf{M}\). Define \(s_1 {:}{=}\sup (\textsf{M}\cap [0,s])\) and \(s_2 {:}{=}\inf (\textsf{M}\cap [s,1])\) (neither set are empty since \(0\in \textsf{M}\) and \(1 \in \textsf{M}\)). Since \(s_1\in \textsf{M}\) and \(s_2\in \textsf{M}\), there are \(t_1\in [0,1]\) and \(t_2\in [0,1]\) such that \(m(t_1) = s_1\) and \(m(t_2) = s_2\). Finally define \(\psi (s) {:}{=}\varphi (t_1) + (s - s_1)\frac{\varphi (t_2) - \varphi (t_1)}{s_2 - s_1} \). In this case, \(\int _{\varOmega }\psi (s) = m(t_1) + (s-s_1) \frac{m(t_2) - m(t_1)}{s_2-s_1} = s\). The path \(\psi \) constructed this way is still monotone, and it has the constant speed property, so it is an ablation path.
Appendix B \(\mathcal {L}^{\infty }\)-Optimal Monotonicity Projection
The algorithm proposed in Appendix C for optimising monotone paths uses updates that can locally introduce nonmonotonicity in the candidate \(\hat{\varphi }_1\), so that it is necessary to project back onto a monotone path \(\varphi _1\). The following routineFootnote 8 performs such a projection in a way that is optimal in the sense of minimising the \(\mathcal {L}^{\infty }\)-distanceFootnote 9, i.e.,
for all \(\textbf{r}\in \varOmega \) and any other monotone path \(\vartheta \).
The algorithm works separately for each \(\textbf{r}\), i.e., we express it as operating simply on continuous functions \(p: [0,1]\rightarrow \mathbb {R}\).
The final step effectively flattens out, in a minimal way, any region in which the function was decreasing.
In practice, this algorithm is executed not on continuous functions but on a PCM-discretised representation; this changes nothing about the algorithm except that instead as real numbers, \(l,r\) and \(t\) are represented by integral indices.
Appendix C Path Optimisation Algorithm
As said in Sect. 5, our optimisation algorithm is essentially gradient descent of a path \(\varphi \): it repeatedly seeks the direction within the space of all paths that (first ignoring the monotonicity constraint) would affect the largest increase to \(P(\varphi )\) as per Sect. 4, for any of the defined score functions. Algorithm 2 shows the details of how this is done in presence of our constraints. In case of \({P_{\uparrow \downarrow }}\), the state \(\varphi \) is understood to consist of the two paths \(\varphi _\uparrow \) and \(\varphi _\downarrow \).
As discussed before, the use of a gradient requires a metric to obtain a vector from the covector-differential, which could be either the implicit \(\ell ^2\) metric on the discretised representation (pixels), or a more physical kernel/filter-based metric. In the present work, we base this on the regularisation filter.
Unlike with the monotonisation condition, the update can easily be made to preserve speed-constness by construction, by projecting for each \(t\) the gradient \(\textbf{g}\) on the sub-tangent-space of zero change to \(\int _\varOmega \varphi (t)\), by subtracting the constant function times \(\int _\varOmega \textbf{g}(t)\). Note this requires the measure of \(\varOmega \) to be normalised, or else considered at this point.
Then we apply these gradients time-wise as updates to the path, using a scalar product in the channel-space to obtain the best direction for \(\varphi \) itself (as opposed to the corresponding image composite \(x_{\varphi ,t}\)).
The learning rate \(\gamma \) can be chosen in different ways. What we found to work best is to normalise the step size in a \(\mathcal {L}^{\infty }\) sense, such that the strongest-affected pixel in the mask experiences a change of at most 0.7 per step. This is small enough to avoid excessively violating the constraint, but not so small to make the algorithm unnecessarily slow (Fig. 6).
Appendix D Baseline Choice
The baseline image is prominently present in the input for much of the ablation path, and it is therefore evident that it will have a significant impact on the saliency. In line with previous work, we opted for a blurred baseline for the examples in the main paper, but even then there is still considerable freedom in the choice of blurring filter. Figure 7 shows two examples, where the result is not fundamentally, but still notably different.
Rights and permissions
Copyright information
© 2023 The Author(s), under exclusive license to Springer Nature Switzerland AG
About this paper
Cite this paper
Sagemüller, J., Verdier, O. (2023). Ablation Path Saliency. In: Longo, L. (eds) Explainable Artificial Intelligence. xAI 2023. Communications in Computer and Information Science, vol 1901. Springer, Cham. https://doi.org/10.1007/978-3-031-44064-9_19
Download citation
DOI: https://doi.org/10.1007/978-3-031-44064-9_19
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-031-44063-2
Online ISBN: 978-3-031-44064-9
eBook Packages: Computer ScienceComputer Science (R0)