Skip to main content

Advertisement

Log in

Artificial understanding: a step toward robust AI

  • Main Paper
  • Published:
AI & SOCIETY Aims and scope Submit manuscript

Abstract

In recent years, state-of-the-art artificial intelligence systems have started to show signs of what might be seen as human level intelligence. More specifically, large language models such as OpenAI’s GPT-3, and more recently Google’s PaLM and DeepMind’s GATO, are performing amazing feats involving the generation of texts. However, it is acknowledged by many researchers that contemporary language models, and more generally, learning systems, still lack important capabilities, such as understanding, reasoning and the ability to employ knowledge of the world and common sense in order to reach or at least advance toward general intelligence. Some believe that scaling will eventually bring about these capabilities; others think that a different architecture is needed. In this paper, we focus on the latter, with the purpose of integrating a theoretical–philosophical conception of understanding as knowledge of dependence relations, with the high-level requirements and engineering design of a robust AI system, which integrates machine learning and symbolic components.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Institutional subscriptions

Similar content being viewed by others

Data availability

Data sharing is not applicable to this article as no datasets were generated or analyzed during the current study.

Notes

  1. See https://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.html, accessed on May, 2022.

  2. See Dean (2021) and Chowdhery (2022).

  3. As per its creators, GPT-3 has been trained on over 175 billion parameters and 45 TB of text gathered from all over the web. See for example https://www.springboard.com/blog/data-science/machine-learning-gpt-3-open-ai/, accessed on 27th of April, 2022. PaLM is a 540 billion parameter language model trained on “780 billion tokens of high-quality text” (Chowdhery 2022: 3).

  4. See https://twitter.com/geoffreyhinton/status/1270814602931187715, accessed on 1st of May, 2022.

  5. For these views, see Grimm (2011, 2014) and Thorisson and Kremelberg (2017).

  6. See Grimm (2014) and Kim (2010).

  7. We shall use cognitive models and world models interchangeably; both refer to a cognitive ability to construct inner models of the world or the immediate surroundings and examine and manipulate them.

  8. For further contemporary support for this view, see Grimm (2014), fn. #1.

  9. The idea of understanding as knowledge of dependence relations (i.e., the idea of expanding the notion of causation to dependence) is supported by several prominent philosophers: Woodward (2003: 6) claims that “any explanation that proceeds by showing how an outcome depends… on other variables counts as causal”; Greco (2010: 9) likewise argues that “understanding involves ‘grasping,’ ‘appreciating,’ or knowing causal relations taken in the broad sense: i.e., the sort of relations that ground explanation.”; and Kim (2010 [1994]: 183) argues that “dependence relations of various kinds serve as objective correlates of explanations.”.

  10. See Grimm (2011). Riggs (2003: 20) emphasizes the importance of the relations among parts and between the parts and the whole, when trying to understand a subject matter. Other philosophers that support the idea that understanding is largely about grasping or coming to know the relations between entities, include Zagzebski (2001: 242, 2009: 142), who asserts that understanding “involves grasping relations of parts to other parts and perhaps the relations of part to wholes”; As Grimm (2014: §VI) puts it, “the common thought here seems to be that the primary objects of understanding are the relationships (or structures) that hold among the various elements of reality.”.

  11. This description of Grimm has some interesting similarities to Pearl and Halpern (2001) Structural-Model approach. See especially §3.1.

  12. Of course, this “grasping” is fallible and the understating can be a matter of degree. See Grimm (2014: fn #21 and §V-VI, respectively).

  13. Object O is any object of understanding and it can include theories in physics, a certain proof in mathematics or logic, a person (as in, “I understand my friend”), a story or an event, an action, or a phrase in a language, to give some examples.

  14. For a structured review of GPT-3’s scope and limitations, see for example Floridi and Chiriatti (2020).

  15. See the comprehensive technical blog of Gwern Branwen—https://www.gwern.net/GPT-3, Accessed 12-May-2022.

  16. See for example Kaplan et al. (2020) and Sam Altman’s, Open AI’s CEO, blog post, which celebrates “Moore’s Law for Everything.” (https://moores.samaltman.com/, Accessed on the 13-May-2022).

  17. See Rae et al. (2022) and Thoppilan et al. (2022) for support for this claim.

  18. See Chowdhery et al. (2022) and Reed (2022).

  19. See Reed (2022).

  20. See also related work of Bengio et al. (2020).

  21. For example, in the universally used ASCII code, the binary numbers 01000001 and 01000010 stand for (are symbols for) the letters A and B, respectively.

  22. See Weidinger (2021).

  23. See min. 46:30, https://www.youtube.com/watch?v=EeqwFjqFvJA, Accessed 22-May, 2022.

  24. Representative studies of this approach include Mao (2019), Raedt et al. (2020), Oltramari (2020), Chitnis et al. (2021), and a relevant interesting research summary from IBM: https://research.ibm.com/blog/ai-neurosymbolic-common-sense, Accessed 25-May-2022.

  25. See https://www.youtube.com/watch?v=GibjI5FoZsE, Accessed 24-May-2022.

  26. See for example Sychev (2021: 731–2) for a similar approach: “we can develop a human-like intelligence as a system where neural networks generate new ideas and strategies given the context and random noise … symbolic reasoning assesses their applicability and the level of risk using available knowledge before trying them in the environment, then the ideas that passed logical verification are implemented under conscious control.”.

  27. As stated in Dean (2021), “Pathways could enable multimodal models that encompass vision, auditory, and language understanding simultaneously.”

  28. For example, when conversing with a human, the system can analyze the auditory input and extract the semantic value as well as the sentiment, and integrate this analysis with a visual analysis of the posture and gestures, for example, to gain a more complete understanding of the pragmatics of the conversation.

  29. The Façade design pattern is a software structural design pattern whose responsibility is to hide the complexities of a system and provide an interface using which any user of the system can access it.

  30. These types of operations, commonly implemented to support persistent storage applications, are usually referred to by the acronym CRUD: Create, Read, Update, Delete.

  31. Henceforth, we will use cognitive and world models interchangeably.

  32. See for example Scholkopf et al. (2021) and Bengio et al. (2020).

  33. See CYC technology overview, https://www.cyc.com/wp-content/uploads/2019/09/Cyc-Technology-Overview.pdf, Accessed 29-May-2022.

  34. See for example the research summary from IBM: https://research.ibm.com/blog/ai-neurosymbolic-common-sense, Accessed 25-May-2022.

  35. See Vogel (1998).

  36. See CYC technology overview (fn. #28), Sect. 5.6.

  37. An additional prominent approach is championed by Yoshua Bengio, who is considered one of the founding fathers of our current-days deep learning systems. To reiterate what was already mentioned at the beginning of Section IV, both the approaches mentioned here agree on the capabilities that current systems lack. However, each approach suggests a different path to implementation. Bengio believes that we should keep the current framework of deep learning, but “build in some of the functional advantages of classical AI rule-based symbolic manipulation in neural nets, but in implicit way.” (~ 46:30, from the AI Debate between Gary Marcus and Yoshua Bengio, https://www.youtube.com/watch?v=EeqwFjqFvJA, Accessed 04-Dec-2022). For Bengio’s opinion, see also his keynote lecture “Deep Learning Cognition” in AI in 2020 and Beyond.

References

  • Allen C, Wallach W (2011) Moral machines: contradiction in terms, or abdication of human responsibility? In: Lin P, Abney K, Bekey G (eds) Robot ethics: the ethical and social implications of robotics. MIT Press, Cambridge, pp 55–68

    Google Scholar 

  • Bengio Y et al. (2020) “A meta-transfer objective for learning to disentangle causal mechanisms.” ArXiv abs/1901.10912

  • Chitnis R et al. (2021). “Learning neuro-symbolic relational transition models for bilevel planning.” ArXiv abs/2105.14074

  • Chowdhery et al. (2022). “PaLM: Scaling Language Modeling with Pathways”. https://doi.org/10.48550/arXiv.2204.02311

  • Dean J (2021) “Introducing pathways: a next-generation AI architecture”. Google’s Blog, https://blog.google/technology/ai/introducing-pathways-next-generation-ai-architecture/, Accessed 17 May 2022

  • Firt E (2020) The missing G. Ai&society 35:995–1007

    Google Scholar 

  • Floridi L, Chiriatti M (2020) GPT-3: its nature, scope, limits, and consequences. Mind Mach 30:681–694. https://doi.org/10.1007/s11023-020-09548-1

    Article  Google Scholar 

  • Greco J (2010) Achieving knowledge. Cambridge University Press, New York

    Book  Google Scholar 

  • Grimm S (2011) Understanding. In: Berneker S, Pritchard D (eds) The Routledge companion to epistemology. Routledge, New York

    Google Scholar 

  • Grimm SR (2014) Understanding as knowledge of causes. In: Fairweather A (ed) Virtue epistemology naturalized. Springer International Publishing, pp 329–345

    Chapter  Google Scholar 

  • Kaplan J et al. (2020) “Scaling laws for neural language models”. arXiv 2001.08361

  • Kim J (2010) Explanatory knowledge and metaphysical dependence. Essays in the metaphysics of mind. Oxford University Press, New York

    Chapter  Google Scholar 

  • Mao J et al. (2019) “The neuro-symbolic concept learner: interpreting scenes words and sentences from natural supervision.” ArXiv abs/1904.12584

  • Marcus G (2020) “The next decade in AI: four steps towards robust artificial intelligence”. https://arxiv.org/abs/2002.06177

  • Marcus G (2022) “Deep learning is hitting a wall”. https://nautil.us/deep-learning-is-hitting-a-wall-14467/. Accessed on 13 May 2022

  • Marcus, G. and Davis, E. (2019). Rebooting AI: Building Artificial Intelligence We Can Trust. Vintage Books.

    Google Scholar 

  • Marcus G, Davis E (2020b) “GPT-3, Bloviator: OpenAI’s language generator has no idea what it’s talking about”. MIT technology review. https://www.technologyreview.com/2020b/08/22/1007539/gpt3-openai-language-generator-artificial-intelligence-ai-opinion/, Accessed 11-May-2022

  • Oltramari A et al. (2020) “Neuro-symbolic architectures for context understanding.” Knowledge graphs for explainable artificial intelligence. Neuro-symbolic Architectures for Context Understanding. https://doi.org/10.48550/arXiv.2003.04707

  • Pearl J, Halpern JY (2001) “Causes and explanations: a structural-model approach—part ii: explanations”. In Proceedings of the Seventeenth International Joint Conference on Artificial Intelligence (IJCAI), San Francisco, CA: Morgan Kaufmann

  • Rae JW et al. (2022) “Scaling language models: Methods, analysis and insights from training Gopher”. arXiv 2112.11446

  • Raedt De L et al. (2020) “From statistical relational to neuro-symbolic artificial intelligence.” ArXiv abs/2003.08316

  • Reed S et al. (2022) “A Generalist Agent.” ArXiv abs/2205.06175

  • Riggs W (2003) Understanding ‘virtue’ and the virtue of understanding. In: DePaul M, Zagzebski L (eds) Intellectual virtue: perspectives from ethics and epistemology. Oxford University Press, New York, pp 203–226

    Chapter  Google Scholar 

  • Scholkopf B et al (2021) Toward causal representation learning. Proc IEEE 109:612–634

    Article  Google Scholar 

  • Sutton R (2019) “The Bitter Lesson”. http://incompleteideas.net/IncIdeas/BitterLesson.html. Accessed 07 May 2022

  • Sychev O (2021) Combining neural networks and symbolic inference in a hybrid cognitive architecture. Procedia Comput Sci 190:728–734

    Article  Google Scholar 

  • Thoppilan R et al. (2022) “LaMDA: Language models for dialog applications”. arXiv 2201.08239

  • Thorisson KR, Kremelberg D (2017) Do machines understand? A short review of understanding & common sense in Artificial Intelligence. (AGI 2017 conference). http://alumni.media.mit.edu/~kris/ftp/AGI17-UUW-DoMachinesUnderstand.pdf. Accessed 07 May 2022

  • Vogel J (1998) “Inference to the best explanation”. In The Routledge Encyclopedia of Philosophy. Taylor and Francis. Retrieved 4 Dec 2022, from https://www.rep.routledge.com/articles/thematic/inference-to-the-best-explanation/v-1. doi: https://doi.org/10.4324/9780415249126-P025-1

  • Weidinger, L., et al. (2021). “Ethical and social risks of harm from Language Models.” ArXiv abs/2112.04359.

  • Wilkenfeld D (2013) Understanding as representation manipulability. Synthese 190(6):997–1016

    Article  Google Scholar 

  • Woodward J (2003) Making things happen: a theory of causal explanation. Oxford University Press, New York

    Google Scholar 

  • Zagzebski L (2001) Recovering understanding. In: Steup M (ed) Knowledge, truth, and duty: essays on epistemic justification, responsibility, and virtue. Oxford University Press, New York

    Google Scholar 

  • Zagzebski L (2009) On epistemology. Wadsworth, Belmont

    Google Scholar 

Download references

Author information

Authors and Affiliations

Authors

Corresponding author

Correspondence to Erez Firt.

Ethics declarations

Conflict of interest

On behalf of all the authors, the corresponding author states that there is no conflict of interest.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Firt, E. Artificial understanding: a step toward robust AI. AI & Soc (2023). https://doi.org/10.1007/s00146-023-01631-2

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1007/s00146-023-01631-2

Keywords

Navigation