Abstract
The aim or goal of science has long been discussed by both philosophers of science and scientists themselves. In The Scientific Image (van Fraassen 1980), the aim of science is famously employed to characterize scientific realism and a version of anti-realism, called constructive empiricism. Since the publication of The Scientific Image, however, various changes have occurred in scientific practice. The increasing use of machine learning technology, especially deep learning (DL), is probably one of the major changes in the last decade. This paper aims to explore the implications of DL-aided research for the aim of science debate. I argue that, while the emerging DL-aided research is unlikely to change the state of classic opposition between constructive empiricism and scientific realism, it could offer interesting cases regarding the opposition between those who espouse truth as the aim of science and those oriented to understanding (of the kind that sacrifices truth).
This is a preview of subscription content, log in via an institution to check access.
Similar content being viewed by others
Explore related subjects
Discover the latest articles, news and stories from top researchers in related subjects.Notes
Hooker and Hooker (2018) discuss the possible limitations of DL-aided (or, more generally, machine-learning-aided) research and suspect that the necessity of finite rounding off in computational methods might make DL models unsuitable for studying those systems that are sensitive to slight differences in the state or for performing tasks that involve small constants (e.g., theory unification where constants such as 1/c or h play a key role) (Hooker and Hooker 2018, 179–180). While this shortcoming may provide a putative reason that scientists should stick to conventional approaches in these kinds of research, this is more like a prescriptive issue than the interpretative one introduced above. This paper focuses on the latter, interpretative issue and aims to discuss the implications of DL-aided scientific research for the meta-hypotheses regarding the aim of science.
It should be noted that Resnik’s (1993) primary concern is not the aim of science discourse in relation to the scientific realism debate in particular. Rowbottom (2014), on the other hand, does discuss various possible interpretations of van Fraassen’s claims and metaphors on ‘the aim of science’ or ‘success’. However, he does not explicitly discuss the rational accountability of research practices. He seems to touch on the issue in a footnote where he mentions Otávio Bueno’s suggestion that van Fraassen’s thesis “might be best understood as concerning what’s constitutive of science” (Rowbottom 2014, 1219, n.14; original emphasis). I take it that “what’s constitutive of science” here means ‘what kind of commitment (to the aim or evaluation of accepted theories) is involved in scientific research practice’. Rowbottom responds that he “need not deny this” (Rowbottom 2014, 1219, n.14) and that such an interpretation could also be accommodated in the four theses on science that he proposes. Then, I assume that the types of analyses conducted in this paper would also be compatible with his framework. The interpretation of the aim of science as the constitutive goal of science is also defended by Bird (2022).
I also suspect that it is only because the whole research context is absent in Resnik’s picture (e.g., the problem at hand, relevant background knowledge and its epistemic status, available methods, worldview, etc.) that the purpose of the research appears irrelevant to scientific practice. Once the context is considered, the issue of whether the aim of the activity and the epistemic status of accepted theories have no bearing on scientific practice is not trivial. Indeed, this is why van Fraassen (1980) devotes a whole chapter to illustrating how various research practices are compatible with constructive empiricism. The accountability of scientific practice thus deserves consideration.
However, in Sect. 6, I discuss whether these views can be further differentiated in light of DL-aided research.
However, I am not committed to the categorization of particular positions to one of these camps.
See, e.g., Google’s course material for developers (https://developers.google.com/machine-learning/crash-course/classification/roc-and-auc, accessed Oct 29th, 2022).
According to van Fraassen’s informal definition, “a theory is empirically adequate exactly if what it says about the observable things and events in this world is true—exactly if it ‘saves’ the phenomena” (van Fraassen 1980, 12).
In this respect, realists typically hold that the empirical adequacy of a theory is best explained by its approximate truth (at least in relevant aspects); thus, the oppositional scheme that contrasts commitment to the empirical adequacy of a theory and its truth may not make much sense to these realists. However, this is one of the points of controversy, and anti-realists have submitted alternative explanations for success (e.g., Kukla 1996; Bueno 1999; Stanford 2001; Lyons 2003). Realists who prefer the localist approach also eschew sweeping inference from empirical success to approximate truth. As I take a neutral position in this paper, and as our focus is on the axiological issue, I will put this point aside for now.
A variant of this method, called stochastic gradient descent, is more commonly used in deep learning studies. There are other variants, too, but their basic purpose is the same, i.e., to adjust the weights to minimize the loss function.
For example, a convolutional neural network (CNN) is widely used in image recognition tasks. It has additional types of layers, called convolutional layers and pooling layers, which process the information transmitted from the preceding layer in a different way.
More precise characterizations of the blackboxness or opacity of DL models or other computational methods are suggested in the literature, typically as the deficiency of various types of transparency (See, e.g., Creel 2020; Duede forthcoming; Humphreys 2009; Lipton 2018; Leslie 2019; Zerilli 2022). To follow the taxonomy that Zerilli (2022) develops based on Lipton (2018) and Leslie (2019), the blackboxness that I have in mind in this paper is the lack of semantic explanation, i.e., “an interpretation of the functions of the individual parts of the algorithmic system in the generation of its output” (Leslie 2019, 41–42).
AUC (area under the (ROC) curve) is one of the commonly used evaluation measures for classification models. It takes values of [0, 1]. Intuitively, AUC = 1 means that the model classified the test samples perfectly, while AUC = 0 means the opposite (i.e., classified in a completely wrong manner), and AUC = 0.5 means that the performance was no better than that achieved by chance.
Of course, Giere criticizes such an account as ‘ad hoc’ and ‘vacuous’. My point here is only to show how, whether persuasive to realists or not, the constructive empiricists’ strategy is applicable when theoretical considerations precede empirical testing.
I suspect that this point holds even in the case of highly idealized models; in these cases, the intended explanandum is highly restricted (e.g., to the qualitative aspect of the system) or highly abstract (i.e., without intended to be applied to particular systems).
For more detail, see, e.g., https://www.kaggle.com/alexisbcook/data-leakage (Accessed July 25, 2021).
Morreno-Torres et al. (2012) provides more formal definitions of the various types of dataset shifts.
More generally, Zednik and Boelsen (2020) discuss the exploratory role of ML and identify three ways in which various interpretation methods could contribute to hypothesis formation, i.e., hypotheses about previously unknown regularities, causal relations and algorithms behind the behavior of certain computational systems (such as brains).
As I have employed a bottom-up approach, I do not pretend that this analysis is generally applicable. Nonetheless, I believe that these observations would hold in other DL-aided studies as well, insofar as they share the same features as those discussed here. Thus, they would serve at least as a first approximation, subject to examination or revision in future case studies, taking into account the idiosyncrasies of each discipline.
There is a more recent version of instrumentalism, called cognitive instrumentalism (Rowbottom 2011), which advises us to take only some types of theoretical discourse literally (e.g., those associated with observable properties or those using analogies with observable systems). Again, like a-ontological instrumentalism, this view concerns the interpretation of theoretical discourse rather than the aim of science. In addition, as mentioned in Sect. 2, the author is skeptical about the usefulness of discussing the aim of science. However, speculating on what cognitive instrumentalists would say about the aim of science, they are likely to hold a view similar to that of the conceptual instrumentalists, if they hold that the theoretical discourses that they take literally are in some way more desirable than other theoretical discourses.
Eamon Duede (forthcoming) argues that interpretability is required mainly in the context of justification (i.e., when the outputs of a DL model are claimed to be new scientific knowledge), and that even uninterpreted DL models can contribute to scientific research in the context of discovery.
Recently, Zerilli (2022) argues that if we take what Daniel Dennet calls an intentional stance toward ML systems (i.e., if we regard them as rational agents with certain beliefs and aims), loss of faithfulness to the system’s exact behavior does not necessarily hinder the interpretation from providing justification for the system’s decision. However, he mainly discusses the deployment of ML systems in societal contexts, such as an estimation of an offender’s recidivism risk. It remains to be discussed whether the same point applies to ML-aided scientific research, especially in the context of justification. Sullivan (2020) also argues that DL models can provide understanding insofar as there is “empirical evidence supporting the link connecting the model to the target phenomenon” (Sullivan 2020, 18). The question would be, then, how faithful an interpretation needs to be to the actual behavior of the system to establish such a link.
References
AI HLEG. 2019. Ethics guidelines for trustworthy AI. European Commission.
Alzubaidi, Abeer, J., Tepperb, and Ahmad Lotfi. 2020. A novel deep mining model for effective knowledge discovery from omics data. Artificial Intelligence in Medicine 104: 101821. https://doi.org/10.1016/j.artmed.2020.101821.
Asay, Jamin. 2019. Going local: a defense of methodological localism about scientific realism. Synthese 196: 587–609. https://doi.org/10.1007/s11229-016-1072-6.
Bird, Alexander. 2007. What is scientific progress? Noûs, 41: 64–89.
Bird, Alexander. 2010. The epistemology of science—a bird’s-eye view. Synthese 175: 5–16.
Bird, Alexander. 2022. Knowing science. Oxford University Press.
Bueno, Otávio. 1999. What is structural empiricism? Scientific change in an empiricist setting. Erkenntnis 50: 59–85. https://doi.org/10.1023/A:1005434915055.
Creel, Kathleen. 2020. Transparency in complex computational systems. Philosophy of Science 87 (4): 568–589. https://doi.org/10.1086/709729.
De Regt, Henk. 2015. Scientific understanding: truth or dare? Synthese 192: 3781–3797. https://doi.org/10.1007/s11229-014-0538-7.
De Regt, Henk. 2017. Understanding scientific understanding. Oxford University Press.
De Regt, Henk. 2020. Understanding, values, and the aims of science. Philosophy of Science 87: 921–932.
Duede, Eamon. (forthcoming). Deep Learning Opacity in Scientific Discovery. Philosophy of Science. https://doi.org/10.1017/psa.2023.8.
Egg, Matthias. 2016. Expanding our grasp: causal knowledge and the problem of unconceived alternatives. British Journal for Philosophy of Science 67: 115–141.
Elgin, Catherine. 2004. True enough. Philosophical Issues 14: 113–131.
Elgin, Catherine. 2017. True enough. The MIT Press.
Erasmus, Adrian, T.D.P. Brunet, and E. Fisher. 2021. What is interpretability? Philosophy and Technology 34: 833–862. https://doi.org/10.1007/s13347-020-00435-2.
Fine, Arthur. 2001. The scientific image twenty years later. Philosophical Studies 106: 107–122.
Giere, Ronald. 1988. Explaining science: a cognitive approach. University of Chicago Press.
Goodman, Brice, and Seth Flaxman. 2017. European Union regulations on algorithmic decision-making and a ‘right to explanation’. AI Magazine 38(3): 50–57. https://doi.org/10.1609/aimag.v38i3.2741.
Harker, David. 2013. How to split a theory: defending selective realism and convergence without proximity. British Journal for Philosophy of Science 64: 79–106.
Hooker, Giles and C. Hooker. 2018. Machine learning and the future of realism. Spontaneous Generations 9 (1): 174–182.
Humphreys, Paul. 2009. The philosophical novelty of computer simulation methods. Synthese 169: 615–626. https://doi.org/10.1007/s11229-008-9435-2.
Kaufman, Shachar, Saharon Rosset, Claudia Perlich, and Ori Stitelman. 2011. Leakage in Data Mining: Formulation, Detection, and Avoidance. KDD’11: Proceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining. August 2011: 556–563. https://doi.org/10.1145/2020408.2020496.
Khalifa, Kareem. 2020. Understanding, truth, and epistemic goals. Philosophy of Science 87 (5): 944–956. https://doi.org/10.1086/710545.
Kukla, Andre. 1996. Antirealist explanations of the success of Science. Philosophy of Science 63: 298–305. http://www.jstor.org/stable/188539.
Laudan, Larry. 1984. Science and values. University of California Press.
LeCun, Yann, Y. Bengio, and G. Hinton. 2015. Deep learning. Nature 521: 436–444. https://doi.org/10.1038/nature14539.
Leslie, David. 2019. Understanding artificial intelligence ethics and safety: A guide for the responsible design and implementation of AI systems in the public sector. The Alan Turing Institute. https://doi.org/10.5281/zenodo.3240529.
Lipton, Zachary C. 2018. The mythos of model interpretability. acm Queue 16 (3): 31–57. https://doi.org/10.1145/3236386.3241340.
Litjens, Geert, Thijs Kooi, Babak Ehteshami Bejnordi, et al. 2017. A survey on deep learning in medical image analysis. Medical Image Analysis 42: 60–88. https://doi.org/10.1016/j.media.2017.07.005.
Lyons, Timothy. 2003. Explaining the success of a scientific theory. Philosophy of Science 70: 891–901.
Lyons, Timothy. 2005. Toward a purely axiological realism. Erkenntnis 63: 167–204. https://doi.org/10.1007/s10670-005-3225-8.
Lyons, Timothy, and Peter Vickers. eds. 2021. Contemporary scientific realism: the challenge from the history of Science. Oxford University Press.
Magnus, P. D., and Craig Callender. 2004. Realist ennui and the base rate fallacy. Philosophy of Science 71: 320–338.
Mater, Adam C., and Michelle L. Coote. 2019. Deep learning in chemistry. Journal of Chemical Information and Modeling 59: 2545–2559. https://doi.org/10.1021/acs.jcim.9b00266.
Morreno-Torres, Jose. G., T. Raeder, R. Alaiz-Rodríguez, et al. 2012. A unifying view on dataset shift in classification. Pattern Recognition 45: 521–530. https://doi.org/10.1016/j.patcog.2011.06.019.
Muraoka, Koki, Yuki Sada, Daiki Miyazaki, et al. 2019. Linking synthesis and structure descriptors from a large collection of synthetic records of zeolite materials. Nature Communications 10: 4459. https://doi.org/10.1038/s41467-019-12394-0.
Nawar, Tamer. 2021. Veritism refuted? Understanding, idealization, and the facts. Synthese 198: 4295–4313. https://doi.org/10.1007/s11229-019-02342-2.
Okatani, Takayuki. 2015. Deep learning. Tokyo: Kodansha.
Onishi, Yukinori. 2017. Defending the selective confirmation strategy. Studies in History and Philosophy of Science Part A 64: 1–10.
Páez, Andrés. 2019. The pragmatic turn in explainable artificial intelligence (XAI). Minds & Machines 29: 441–459. https://doi.org/10.1007/s11023-019-09502-w.
Peters, Dean. 2014. What elements of successful scientific theories are the correct targets for ‘selective’ scientific realism? Philosophy of Science 81: 377–397.
Potochnik, Angela. 2015. The diverse aims of science. Studies in History and Philosophy of Science 53: 71–80. https://doi.org/10.1016/j.shpsa.2015.05.008.
Potochnik, Angela. 2017. Idealization and the aims of science. The University of Chicago Press.
Quiñonero-Candela, Joaquin, M. Sugiyama, A. Schwaighofer, and N.D. Lawrence. 2009. Dataset shift in machine learning. The MIT Press.
Raví, Daniele, Charence Wong, Fani Deligianni, Melissa Berthelot, et al. 2017. Deep learning for health informatics. IEEE Journal of Biomedical and Health Informatics 21(1): 4–21.
Resnik, David. 1993. Do scientific aims justify methodological rules? Erkenntnis 38: 223–232.
Ribeiro, Marco Tulio, Sameer Singh, and Carlos Guestrin. 2016. ‘Why should I trust you?’. Explaining the Predictions of Any Classifier. KDD’16: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. August 2016: 1135–1144. https://doi.org/10.1145/2939672.2939778.
Rice, Collin. 2016. Factive scientific understanding without accurate representation. Biology and Philosophy 31: 81–102. https://doi.org/10.1007/s10539-015-9510-2.
Rowbottom, Darrell P. 2011. The instrumentalist’s new clothes. Philosophy of Science 78(5): 1200–1211.
Rowbottom, Darrell P. 2014. Aimless science. Synthese 191: 1211–1221. https://doi.org/10.1007/s11229-013-0319-8.
Saatsi, Juha. 2017. Replacing recipe realism. Synthese 194: 3233–3244. https://doi.org/10.1007/s11229-015-0962-3.
Saito, Koki. 2016. Deep learning from scratch. Tokyo: O’Reilly Japan.
Schmidt, Jonathan, Mário R. G. Marques, Silvana Botti, et al. 2019. Recent advances and applications of machine learning in solidstate materials science. npj Computational Materials 5: 83. https://doi.org/10.1038/s41524-019-0221-0.
Stanford, Kyle. 2001. An antirealist explanation of the success of science. Philosophy of Science 67(2): 266–284. http://www.jstor.org/stable/188724.
Stanford, Kyle. 2006. Exceeding our grasp; science, history, and the problem of unconceived alternatives. Oxford University Press.
Sullivan, Emily. 2020. Understanding from machine learning models. British Journal for Philosophy of Science 00 (2020), 1–28. https://doi.org/10.1093/bjps/axz035.
Van Fraassen, B. C. 1980. The scientific image. Oxford: Oxford University Press.
Van Fraassen, B. C. 2002. The empirical stance. Yale: Yale University Press.
Winsberg, Eric. 2010. Science in the age of computer simulation. Chicago: University of Chicago Press.
Xie, Tian, and Jeffrey C. Grossman. 2018. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties. Physical Review Letters 120: 145301.
Yan, Karen, and Jonathan Hricko. 2017. Brain networks, structural realism, and local approaches to the scientific realism debate. Studies in History and Philosophy of Science Part C 64: 1–10.
Zednik, Carlos, and Hannes Boelsen. 2020. The exploratory role of explainable artificial intelligence. In PSA2020: the 27th Biennial Meeting of the Philosophy of Science Association, Baltimore, MD. http://philsci-archive.pitt.edu/id/eprint/18005. Accessed 04 May 2023.
Zerilli, John. 2022. Explaining machine learning decisions. Philosophy of Science 89: 1–19.
Ziletti, Angelo, Devinder Kumar, Matthias Scheffler, and Luca M. Ghiringhelli. 2018. Insightful classification of crystal structures using deep learning. Nature Communications 9: 2775. https://doi.org/10.1038/s41467-018-05169-6.
Acknowledgements
I am grateful to the audience of the workshop ‘Diversity of the Scientific Realism Debate’ in Tokyo, 2019, and CLMPST 2019 in Prague, where a part of this manuscript was presented, for the inspiring discussions and comments. I also thank the anonymous reviewers and the editors-in-chief for this journal for helpful comments to improve the manuscript. This work is supported by Japan Society for the Promotion of Science (JSPS) KAKENHI Grant (Grant Numbers: JP18K12178, JP 18H00604).
Funding
Japan Society for the Promotion of Science (JSPS) KAKENHI Grant (Grant Numbers: JP18K12178, JP 18H00604).
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Competing interests
None.
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.
About this article
Cite this article
Onishi, Y. Deep Learning-Aided Research and the Aim-of-Science Controversy. J Gen Philos Sci (2024). https://doi.org/10.1007/s10838-023-09667-0
Accepted:
Published:
DOI: https://doi.org/10.1007/s10838-023-09667-0