Abstract
In recent years, there have been many attempts to combine XAI with the field of symbolic AI in order to generate explanations for neural networks that are more interpretable and better align with human reasoning, with one prominent candidate for this synergy being the sub-field of computational argumentation. One method is to represent neural networks with quantitative bipolar argumentation frameworks (QBAFs) equipped with a particular semantics. The resulting QBAF can then be viewed as an explanation for the associated neural network. In this paper, we explore a novel LRP-based semantics under a new QBAF variant, namely neural QBAFs (nQBAFs). Since an nQBAF of a neural network is typically large, the nQBAF must be simplified before being used as an explanation. Our empirical evaluation indicates that the manner of this simplification is all important for the quality of the resulting explanation.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
The definition of path is adopted from [1], where there exists a path via E (set of edges) from \(n_a\) to \(n_b\) (from a node to another) iff \(\exists n_1,...,n_t\) with \(n_1 = n_a\) and \(n_t = n_b\) such that \((n_1, n_2),...,(n_{t-1}, n_t) \in E\).
- 2.
Note that, with an abuse of notation, \(\theta (n, i)\) stands for \(\theta ((n, i))\), for simplicity. Unless explicitly stated, this notation is used throughout the rest of the paper.
- 3.
In this paper, we will choose \(D=\mathbb {R}\).
- 4.
Note that \(\mathcal {P}(A)\) is the power set of a set A.
References
Albini, E., Lertvittayakumjorn, P., Rago, A., Toni, F.: Deep argumentative explanations (2021). https://arxiv.org/abs/2012.05766
Baroni, P., Rago, A., Toni, F.: How many properties do we need for gradual argumentation?. In: AAAI (2018)
Baroni, P., Rago, A., Toni, F.: From fine-grained properties to broad principles for gradual argumentation: a principled spectrum. Int. J. Approx. Reason. 105, 252–286 (2019). https://doi.org/10.1016/j.ijar.2018.11.019
Dejl, A., et al.: Argflow: a toolkit for deep argumentative explanations for neural networks. In: International Foundation for Autonomous Agents and Multiagent Systems, Richland, SC, pp. 1761–1763 (2021)
Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: a large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp. 248–255 (2009). https://doi.org/10.1109/CVPR.2009.5206848
Dung, P.M.: On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic programming and n-person games. Artif. Intell. 77(2), 321–357 (1995). https://doi.org/10.1016/0004-3702(94)00041-X
Google, L.: Neuron groups - building blocks of interpretability (2018). https://bit.ly/3a483Xc
Kim, B., et al.: Interpretability beyond feature attribution: auantitative testing with concept activation vectors (TCAV). In: Dy, J., Krause, A. (eds.) Proceedings of the 35th International Conference on Machine Learning. Proceedings of Machine Learning Research, 10–15 July 2018, vol. 80, pp. 2668–2677. PMLR (2018). https://proceedings.mlr.press/v80/kim18d.html
Lertvittayakumjorn, P., Specia, L., Toni, F.: FIND: human-in-the-loop debugging deep text classifiers. In: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 332–348. Association for Computational Linguistics, November 2020. https://doi.org/10.18653/v1/2020.emnlp-main.24
Li, J., Zhang, C., Zhou, J.T., Fu, H., Xia, S., Hu, Q.: Deep-lift: deep label-specific feature learning for image annotation. IEEE Trans. Cybern. 1–10 (2021). https://doi.org/10.1109/TCYB.2021.3049630
Montavon, G., Binder, A., Lapuschkin, S., Samek, W., Müller, K.-R.: Layer-wise relevance propagation: an overview. 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. 193–209. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-28954-6_10
Olah, C., Mordvintsev, A., Schubert, L.: Feature visualization. Distill 2(11), e7 (2017). https://doi.org/10.23915/distill.00007
Olah, C., et al.: The building blocks of interpretability. Distill 3(03), e10 (2018). https://doi.org/10.23915/distill.00010
Potyka, N.: Interpreting neural networks as quantitative argumentation frameworks. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 7, pp. 6463–6470, May 2021. https://ojs.aaai.org/index.php/AAAI/article/view/16801
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 (ICCV), October 2017
Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. In: International Conference on Learning Representations (2015)
Smilkov, D., Thorat, N., Kim, B., Viégas, F., Wattenberg, M.: Smoothgrad: removing noise by adding noise (2017)
Synergee Fitness Worldwide, I.: (2019). https://amzn.to/3Db2xOQ
Wataree: Police van Thailand (2019). https://bit.ly/3Fi1oqx
websubstance: Baby tummy time (nd). https://bit.ly/3D8FZya
Acknowledgements
The first author was funded in part by Imperial College London under UROP (Undergraduate Research Opportunities Programme). The last author was partially funded by the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No. 101020934). Finally, Rago and Toni were partially funded by J.P. Morgan and by the Royal Academy of Engineering under the Research Chairs and Senior Research Fellowships scheme. Any views or opinions expressed herein are solely those of the authors listed.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Appendix: Lemmas for Dialectical Properties Proofs
Appendix: Lemmas for Dialectical Properties Proofs
Lemma 1
Any attacking argument has a negative strength.
\( \forall a \in A [\exists x \in \mathcal {P}(A)[a \in Att(x)] \rightarrow \sigma (a) < 0] \)
Proof
Take arbitrary \( a \in A \). Assume there exists some \( x \in \mathcal {P}(A) \) such that \( a \in Att(x) \). Since \( a \in Att(x) \), \( (a, x) \in Att \) so \( c_-(a, x) \) is true, meaning \( R_{\rho (a) \leftarrow \rho (x)} < 0 \). As \( \sigma (a) = R_{\rho (a) \leftarrow \rho (x)} \) by Definition 7, then \( \sigma (a) < 0 \). \(\square \)
Lemma 2
Any supporting argument has a positive strength.
\( \forall a \in A [\exists x \in \mathcal {P}(A)[a \in Supp(x)] \rightarrow \sigma (a) > 0] \)
Proof
Take arbitrary \( a \in A \). Assume there exists some \( x \in \mathcal {P}(A) \) such that \( a \in Supp(x) \). Since \( a \in Supp(x) \), \( (a, x) \in Supp \) so \( c_+(a, x) \) is true, meaning \( R_{\rho (a) \leftarrow \rho (x)} > 0 \). As \( \sigma (a) = R_{\rho (a) \leftarrow \rho (x)} \) by Definition 7, then \( \sigma (a) > 0 \). \(\square \)
Lemma 3
Any argument that neither supports nor attacks any group and does not represent an output node has zero strength.
\( \forall a \in A [\forall x \in \mathcal {P}(A)[(a, x) \notin Supp \wedge (a, x) \notin Att] \wedge \rho (a) \notin V_{d+1} \rightarrow \sigma (a) = 0] \)
Proof
This proposition follows immediately from Definition 7. \(\square \)
Rights and permissions
Copyright information
© 2022 The Author(s), under exclusive license to Springer Nature Switzerland AG
About this paper
Cite this paper
Sukpanichnant, P., Rago, A., Lertvittayakumjorn, P., Toni, F. (2022). Neural QBAFs: Explaining Neural Networks Under LRP-Based Argumentation Frameworks. In: Bandini, S., Gasparini, F., Mascardi, V., Palmonari, M., Vizzari, G. (eds) AIxIA 2021 – Advances in Artificial Intelligence. AIxIA 2021. Lecture Notes in Computer Science(), vol 13196. Springer, Cham. https://doi.org/10.1007/978-3-031-08421-8_30
Download citation
DOI: https://doi.org/10.1007/978-3-031-08421-8_30
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-031-08420-1
Online ISBN: 978-3-031-08421-8
eBook Packages: Computer ScienceComputer Science (R0)