Abstract
Explanation can form the basis, in any lawfully behaving environment, of plans, summaries, justifications, analysis and predictions, and serve as a method for probing their validity. For systems with general intelligence, an equally important reason to generate explanations is for directing cumulative knowledge acquisition: Lest they be born knowing everything, a general machine intelligence must be able to handle novelty. This can only be accomplished through a systematic logical analysis of how, in the face of novelty, effective control is achieved and maintained—in other words, through the systematic explanation of experience. Explanation generation is thus a requirement for more powerful AI systems, not only for their owners (to verify proper knowledge and operation) but for the AI itself—to leverage its existing knowledge when learning something new. In either case, assigning the automatic generation of explanation to the system itself seems sensible, and quite possibly unavoidable. In this paper we argue that the quality of an agent’s explanation generation mechanism is based on how well it fulfils three goals – or purposes – of explanation production: Uncovering unknown or hidden patterns, highlighting or identifying relevant causal chains, and identifying incorrect background assumptions. We present the arguments behind this conclusion and briefly describe an implemented self-explaining system, AERA (Autocatlytic Endogenous Reflective Architecture), capable of goal-directed self-explanation: Autonomously explaining its own behavior as well as its acquired knowledge of tasks and environment.
This is a preview of subscription content, log in via an institution to check access.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
Other types of explanation than causal have been proposed. Teleological explanations are explanations focused on utility (to explain by defining the purpose or intent of the thing to be explained [2]). But nowhere nearly all things in need of explaining have intent or utility behind them.
- 2.
Providing adequate levels of transparency modern machine learning and AI systems such as reinforcement learners and deep neural networks, with adequate levels of transparency, requires considerable post-hoc effort and skill in interpreting algorithms, and most of the time it is essentially prohibitive due to cost.
- 3.
Traditionally, ‘ampliative reasoning’ refers to any process that relies on abduction and induction in any combination to achieve a particular result (cf. [16]); we include (defeasible, non-axiomatic) deduction in that list.
- 4.
- 5.
For convenience we include, as part of the ‘encoding’ of an explanation, any references to related but different phenomena intended to better match an explainee’s knowledge—that is, to explain something better to a particular explainee, due to their particular knowledge at the time of the explanation generation.
- 6.
This certainly is a factor in all explanations produced by one human for another. It may not, however, be relevant for self-explanation generation since the meaning of a low-value (or zero-value, i.e. worthless) explanation produced for oneself is undefined.
References
Bieger, J., Thórisson, K.R.: Evaluating understanding. In: IJCAI Workshop on Evaluating General-Purpose AI, Melbourne, Australia (2017)
Cohen, J.: Teleological explanation. Proc. Aristot. Soc. 51, 255–292 (1950)
Halpern, J.Y., Pearl, J.: Causes and explanations: a structural-model approach – part I: causes. Br. J. Philos. Sci. 56, 889–911 (2005)
Halpern, J.Y., Pearl, J.: Causes and explanations: a structural-model approach – Part II: Explanations. Br. J. Philos. Sci. 56, 843–847 (2005)
Hilton, D.J.: Conversational processes and causal explanation. Psychol. Bull. 107(1), 65–81 (1990)
Hilton, D.J., Slugoski, B.R.: Knowledge-based causal attribution: the abnormal conditions focus model. Psychol. Rev. 93(1), 75–88 (1986). https://doi.org/10.1037/0033-295X.93.1.75
Josephson, J., Josephson, S.: Abductive Inference: Computation, Philosophy, Technology. Computation, Philosophy, Technology, Cambridge University Press (1996)
Lapuschkin, S., Wäldchen, S., Binder, A., Montavon, G., Samek, W., Müller, K.R.: Unmasking clever hans predictors and assessing what machines really learn. Nat. Commun. 10(1) (2019). https://doi.org/10.1038/s41467-019-08987-4
Lombrozo, T.: The structure and function of explanations. Trends Cogn. Sci. 10(10), 464–470 (2006)
Miller, T.: Explanation in artificial intelligence: insights from the social sciences (2017)
Nivel, E., Thórisson, K.R.: Replicode: a constructivist programming paradigm and language. Technical report RUTR-SCS13001, Reykjavik University School of Computer Science (2013)
Nivel, E., Thórisson, K.R.: Towards a programming paradigm for control systems with high levels of existential autonomy. In: Kühnberger, K.-U., Rudolph, S., Wang, P. (eds.) AGI 2013. LNCS (LNAI), vol. 7999, pp. 78–87. Springer, Heidelberg (2013). https://doi.org/10.1007/978-3-642-39521-5_9
Nivel, E., et al.: Bounded recursive self-improvement (2013)
Palacio, S., Lucieri, A., Munir, M., Hees, J., Ahmed, S., Dengel, A.: Xai handbook: towards a unified framework for explainable AI (2021)
Pearl, J.: Causality: Models, Reasoning and Inference, 2nd edn. Cambridge University Press, New York, NY, USA (2009)
Psillos, S.: An explorer upon untrodden ground: peirce on abduction. In: Handbook of the History of Logic, vol. 10, pp. 117–151. Elsevier (2011)
Rörbeck, H.: Self-Explaining Artificial Intelligence: On the Requirements for Autonomous Explanation Generation. M.Sc. Thesis, Dept. Comp. Sci., Reykjavik University (2022)
Strevens, M.: The causal and unification approaches to explanation unified-causally. Noûs 38(1), 154–176 (2004)
Thórisson, K.R.: A new constructivist AI: from manual construction to self-constructive systems. In: Wang, P., Goertzel, B. (eds.) Theoretical Foundations of Artificial General Intelligence, vol. 4, pp. 145–171 (2012)
Thórisson, K.R.: Seed-programmed autonomous general learning. Proc. Mach. Learn. Res. 131, 32–70 (2020)
Thórisson, K.R., Kremelberg, D., Steunebrink, B.R., Nivel, E.: About understanding. In: Steunebrink, B., Wang, P., Goertzel, B. (eds.) AGI -2016. LNCS (LNAI), vol. 9782, pp. 106–117. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-41649-6_11
Thórisson, K.R., Talbot, A.: Cumulative learning with causal-relational models. In: Iklé, M., Franz, A., Rzepka, R., Goertzel, B. (eds.) AGI 2018. LNCS (LNAI), vol. 10999, pp. 227–237. Springer, Cham (2018). https://doi.org/10.1007/978-3-319-97676-1_22
Thórisson, K.R.: The explanation hypothesis in general self-supervised learning. Proc. Mach. Learn. Res. 159, 5–27 (2021)
Woodward, J.: Making things Happen: A Theory of Causal Explanation. Oxford University Press, Oxford (2005)
Acknowledgments
This work was supported in part by Cisco Systems, the Icelandic Institute for Intelligent Machines and Reykjavik University.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2023 The Author(s), under exclusive license to Springer Nature Switzerland AG
About this paper
Cite this paper
Thórisson, K.R., Rörbeck, H., Thompson, J., Latapie, H. (2023). Explicit Goal-Driven Autonomous Self-Explanation Generation. In: Hammer, P., Alirezaie, M., Strannegård, C. (eds) Artificial General Intelligence. AGI 2023. Lecture Notes in Computer Science(), vol 13921. Springer, Cham. https://doi.org/10.1007/978-3-031-33469-6_29
Download citation
DOI: https://doi.org/10.1007/978-3-031-33469-6_29
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-031-33468-9
Online ISBN: 978-3-031-33469-6
eBook Packages: Computer ScienceComputer Science (R0)