Florian J. Boge

Philosophical debates on scientific understanding often contrast explanatory and objectual understanding. We here propose a novel way of slicing the debate, by distinguishing _scientific_ and _scientists’_ understanding. We argue that the former must be communicable in a way that requires the availability of an explanation, while the latter may often be thoroughly objectual. We offer support for the importance of this distinction by appeal to historical case-studies and argue that it dissolves key disputes in the literature, reconciles objectualist and explanationist accounts, and offers a new reason to resist the claim that scientific understanding can exist without explanation. Our account thus explains why many authors in the debate on scientific understanding have focused on explanation, while at the same time acknowledging the epistemic value and importance of objectual understanding.

Mature scientific hypotheses are confirmed by large amounts of independent evidence. How could anyone be an anti-realist under these conditions? A classic response appeals to confirmational holism and underdetermination, but it is unclear whether traditional arguments succeed. I offer a new line of argument: If holism is interpreted as saying that the confirmation of every part of a hypothesis depends on the confirmation of the whole hypothesis, we must formulate conditions under which the confirmation received by the whole can be transferred to its parts. However, underdetermination suggests that relevant conditions are typically not met. If this is true, the confirmation received by the whole remains bounded by the priors for the parts, and we lack compelling reasons to believe substantive hypotheses based on evidence beyond the degree to which the posits involved in them are antecedently believed. A rejoinder comes from selective realism: If some posit is preserved throughout theory change, it is confirmed beyond the degree to which the containing hypothesis is. However, the variant of holism considered here exactly implies that we cannot confirm such posits in isolation. As I will show, the realist is thus forced into a dilemma: Either she succumbs to the holistic challenge, or she must embrace meta-empirical facts, such as the posit’s recurrence, as confirmatory.

A number of authors (Morgan in: Morgan and Morrison (Eds.) Models as mediators, Cambridge University Press, Cambridge, 1999; Boumans in Philos Sci 72(5):850, 2005; Morison in Philos Stud 143(1):33–57, 2009; Massimi and Bhimji in Stud Hist Philos Modern Phys 51:71–81, 2015; Parker in Brit J Philos Sci 68(1): 273–304, 2017) have argued that models can be quite literally thought of as measuring instruments. I here challenge this view by reconstructing three arguments from the literature and rebutting them. Further, I argue that models should be seen as cognitive rather than measuring instruments, and that the distinction is important for understanding scientific change: Both yield two distinct sources of insight that mutually depend on each other, and should not be equated. In particular, we may perform the exact same actions in the laboratory but conceive of them entirely differently by virtue of the models we endorse at different points in time.

Artificial intelligence (AI) has become a topic of major interest to philosophers of science. Among the issues commonly discussed is AI’s _opacity_. To remedy opacity, scientists have provided methods commonly subsumed under the label ‘eXplaibable Artificial Intelligence’ (XAI) that aim to make AI and its outputs ‘interpretable’ and ‘explainable’. However, there is little interaction between developments in XAI and philosophical debates on scientific explanation. We here improve on this situation and argue for a descriptive and a normative thesis: (i) When suitably embedded into scientific research processes, XAI methods’ outputs can facilitate genuine scientific understanding. (ii) In order for XAI outputs to fulfill this function, they should be made _testable_. We will support our theses by building on recent and long-standing ideas from philosophy of science, by comparing them to a recent framework from the XAI community, and by showcasing their applicability to case studies from the life sciences.

In Chap. 10.1007/978-3-319-95765-4_2, we located the importance of de Broglie’s research for the development of QM in his speculating about matter waves, and hence in his indirect contribution to Schrödinger’s discovery of the SE. But de Broglie’s contributions to the early development of QM of course exceeded this point. In particular, he also proposed a so called pilot wave theory, in which there would be waves and particles, and which he hoped to be a precursor to a future (fully developed) ‘theory of the double solution’. In the latter there would be additional singular solutions to the SE or the KGE, highly peaked field amplitudes with a phase coinciding with that of the regular solutions, replacing genuine, additional particles (cf. Bacciagaluppi and Valentini 2009, p. 60; Jammer 1966, pp. 292 and 357; Mehra and Rechenberg 1987, p. 1209). Pilot wave theory and the theory of the double solution were certainly both motivated by the duality of wave-like and particle-like aspects exhibited in experiments and discussed at some length in Chap. 10.1007/978-3-319-95765-4_2. In contrast to the naïve (collapse) approach discussed therein, however, de Broglie’s pilot wave theory would have particles be present at all times and only be guided or piloted by a simultaneously occurring wave-phenomenon. The theory of the double solution, he hoped, would then explain everything in terms of waves (fields) alone (cf. Dürr et al. 2012, p. 7; Mehra and Rechenberg 1987, p. 1209).

There is, it seems, a rather natural response to the conceptual problems raised by QM. This response, put frankly, is to say that ‘it’s all just epistemic!’ More precisely this would mean to deprive the quantum state of its ontological significance and to construe the theory not as a description of the actual, real situation of physical systems, but rather as a representation of the knowledge an actual or ideal observer or agent has about these. So for instance, when a quantum system passes a double slit, we cannot know exactly where it is going to turn up. But we can try to make precise predictions about its future behavior, based on our previous experience with similar situations and systems, and hence quantify, in a sense, our knowledge about its future behavior and about the occurrence of spots in various regions on the screen behind the slit. Again given past experience, a quantum mechanical state function may be the tool of choice to accomplish this task. But in the instant we see a dot appearing on the screen we can update our knowledge about the system’s actual state, since we can now be rather certain (assuming that we have not visited too many epistemology classes) that the system has occupied exactly that region on the screen at the moment of the appearance of the dot.

QM is notoriously associated with a certain ‘strangeness’ or ‘weirdness’ (e.g. Rosenblum and Kuttner 2011, p. 4; Davies 2004, p. 11) which stems, in the first place, from the divergence of the phenomena that it describes and predicts from our pre-quantum expectations. By ‘phenomenon’ we here mean, for practical reasons, something along the lines of Bogen and Woodward (1988, pp. 305–306), according to whom the phenomenon is rather what the theory predicts, which may not even be observable, whereas the data are the observables that serve as evidence for phenomena.

Deep Neural Networks (DNNs) are becoming increasingly important as scientific tools, as they excel in various scientific applications beyond what was considered possible. Yet from a certain vantage point, they are nothing but parametrized functions fθ(x)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{f}_{\varvec{\theta }}(\varvec{x})$$\end{document} of some data vector x\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varvec{x}$$\end{document}, and their ‘learning’ is nothing but an iterative, algorithmic fitting of the parameters to data. Hence, what could be special about them as a scientific tool or model? I will here suggest an integrated perspective that mediates between extremes, by arguing that what makes DNNs in science special is their ability to develop functional concept proxies (FCPs): Substructures that occasionally provide them with abilities that correspond to those facilitated by concepts in human reasoning. Furthermore, I will argue that this introduces a problem that has so far barely been recognized by practitioners and philosophers alike: That DNNs may succeed on some vast and unwieldy data sets because they develop FCPs for features that are not transparent to human researchers. The resulting breach between scientific success and human understanding I call the ‘Actually Smart Hans Problem’.

Putnam coined what is now known as the no miracles argument “[t]he positive argument for realism”. In its opposition, he put an argument that by his own standards counts as negative. But are there no positive arguments against scientific realism? I believe that there is such an argument that has figured in the back of much of the realism-debate, but, to my knowledge, has nowhere been stated and defended explicitly. This is an argument from the success of quantum physics to the unlikely appropriateness of scientific realism as a philosophical stance towards science. I will here state this argument and offer a detailed defence of its premises. The purpose of this is to both exhibit in detail how far the intuition that quantum physics threatens realism can be driven, in the light also of more recent developments, as well as to exhibit possible vulnerabilities, i.e., to show where potential detractors might attack.

Deep neural networks have become increasingly successful in applications from biology to cosmology to social science. Trained DNNs, moreover, correspond to models that ideally allow the prediction of new phenomena. Building in part on the literature on ‘eXplainable AI’, I here argue that these models are instrumental in a sense that makes them non-explanatory, and that their automated generation is opaque in a unique way. This combination implies the possibility of an unprecedented gap between discovery and explanation: When unsupervised models are successfully used in exploratory contexts, scientists face a whole new challenge in forming the concepts required for understanding underlying mechanisms.

Computer simulations are nowadays often directly involved in the generation of experimental results. Given this dependency of experiments on computer simulations, that of simulations on models, and that of the models on free parameters, how do researchers establish trust in their experimental results? Using high-energy physics (HEP) as a case study, I will identify three different types of robustness that I call conceptual, methodological, and parametric robustness, and show how they can sanction this trust. However, as I will also show, simulation models in HEP themselves fail to exhibit a type of robustness I call inverse parametric robustness. This combination of robustness and failures thereof is best understood by distinguishing different epistemic capacities of simulations and different senses of trust: Trusting simulations in their capacity to facilitate credible experimental results can mean accepting them as means for generating belief in these results, while this need not imply believing the models themselves in their capacity to represent an underlying reality.

Two powerful arguments have famously dominated the realism debate in philosophy of science: The No Miracles Argument (NMA) and the Pessimistic Meta-Induction (PMI). A standard response to the PMI is selective scientific realism (SSR), wherein only the working posits of a theory are considered worthy of doxastic commitment. Building on the recent debate over the NMA and the connections between the NMA and the PMI, I here consider a stronger inductive argument that poses a direct challenge for SSR: Because it is sometimes exactly the working posits which contradict each other, i.e., that which is directly responsible for empirical success, SSR cannot deliver a general explanation of scientific success.

Despite remarkable efforts, it remains notoriously difficult to equip quantum theory with a coherent ontology. Hence, Healey (2017, 12) has recently suggested that ‘‘quantum theory has no physical ontology and states no facts about physical objects or events’’, and Fuchs et al. (2014, 752) similarly hold that ‘‘quantum mechanics itself does not deal directly with the objective world’’. While intriguing, these positions either raise the question of how talk of ‘physical reality’ can even remain meaningful, or they must ultimately embrace a hidden variables-view, in tension with their original project. I here offer a neo-Kantian alternative. In particular, I will show how constitutive elements in the sense of Reichenbach (1920) and Friedman (1999, 2001) can be identified within quantum theory, through considerations of symmetries that allow the constitution of a ‘quantum reality’, without invoking any notion of a radically mind-independent reality. The resulting conception will inherit elements from pragmatist and ‘QBist’ approaches, but also differ from them in crucial respects. Furthermore, going beyond the Friedmanian program, I will show how non-fundamental and approximate symmetries can be relevant for identifying constitutive principles.

The workshop “Machine Learning: Prediction Without Explanation?” brought together philosophers of science and scholars from various fields who study and employ Machine Learning (ML) techniques, in order to discuss the changing face of science in the light of ML's constantly growing use. One major focus of the workshop was on the impact of ML on the concept and value of scientific explanation. One may speculate whether ML’s increased use in science exemplifies a paradigmatic turn towards mere pattern recognition and prediction and away from science’s traditional aim of explanation. In contrast, certain conceptions of explanation, such as statistical explanation, could turn out to fit well with the achievements of present-day ML and concede an explanatory value to these achievements after all. It is an open question how to explain ML successes themselves, and this question was in the focus of several talks in the workshop. Based on the topics raised, we will discuss the talks’ contents in more detail as organized into (i) practitioners’ perspectives, (ii) explanations from ML, (iii) explanations of ML, (iv) societal implications, and (v) global and historical perspectives.

In this paper, we pursue three general aims: (I) We will define a notion of fundamental opacity and ask whether it can be found in High Energy Physics (HEP), given the involvement of machine learning (ML) and computer simulations (CS) therein. (II) We identify two kinds of non-fundamental, contingent opacity associated with CS and ML in HEP respectively, and ask whether, and if so how, they may be overcome. (III) We address the question of whether any kind of opacity, contingent or fundamental, is unique to ML or CS, or whether they stand in continuity to kinds of opacity associated with other scientific research.

Large scale experiments at CERN’s Large Hadron Collider rely heavily on computer simulations, a fact that has recently caught philosophers’ attention. CSs obviously require appropriate modeling, and it is a common assumption among philosophers that the relevant models can be ordered into hierarchical structures. Focusing on LHC’s ATLAS experiment, we will establish three central results here: with some distinct modifications, individual components of ATLAS’ overall simulation infrastructure can be ordered into hierarchical structures. Hence, to a good degree of approximation, hierarchical accounts remain valid at least as descriptive accounts of initial modeling steps. In order to perform the epistemic function Winsberg Model-based reasoning in scientific discovery. Kluwer Academic/plenum Publishers, New York, pp 255–269, 1999) assigns to models in simulation—generate knowledge through a sequence of skillful but non-deductive transformations—ATLAS’ simulation models have to be considered part of a network rather than a hierarchy, in turn making the associated simulation modeling messy rather than motley. Deriving knowledge-claims from this ‘mess’ requires two sources of justification: holistic validation :253–262, 2010; in Carrier M, Nordmann A Science in the context of application. Springer, Berlin, pp 115–130, 2011), and model coherence. As it turns out, the degree of model coherence sets HEP apart from other messy, simulation-intensive disciplines such as climate science, and the reasons for this are to be sought in the historical, empirical and theoretical foundations of the respective discipline.

The question of where, between theory and experiment, computer simulations (CSs) locate on the methodological map is one of the central questions in the epistemology of simulation (cf. Saam Journal for General Philosophy of Science, 48, 293–309, 2017). The two extremes on the map have them either be a kind of experiment in their own right (e.g. Barberousse et al. Synthese, 169, 557–574, 2009; Morgan 2002, 2003, Journal of Economic Methodology, 12(2), 317–329, 2005; Morrison Philosophical Studies, 143, 33–57, 2009; Morrison 2015; Massimi and Bhimji Studies in History and Philosophy of Science Part B: Studies in History and Philosophy of Modern Physics, 51, 71–81, 2015; Parker Synthese, 169, 483–496, 2009) or just an argument executed with the aid of a computer (e.g. Beisbart European Journal for Philosophy of Science, 2, 395–434, 2012; Beisbart and Norton International Studies in the Philosophy of Science, 26, 403–422, 2012). There exist multiple versions of the first kind of position, whereas the latter is rather unified. I will argue that, while many claims about the ‘experimental’ status of CSs seem unjustified, there is a variant of the first position that seems preferable. In particular I will argue that while CSs respect the logic of (deductively valid) arguments, they neither agree with their pragmatics nor their epistemology. I will then lay out in what sense CSs can fruitfully be seen as experiments, and what features set them apart from traditional experiments nonetheless. I conclude that they should be seen as surrogate experiments, i.e. experiments executed consciously on the wrong kind of system, but with an exploitable connection to the system of interest. Finally, I contrast my view with that of Beisbart (European Journal for Philosophy of Science, 8, 171–204, 2018), according to which CSs are surrogates for experiments, arguing that this introduces an arbitrary split between CSs and other kinds of simulations.

Computer simulations are involved in numerous branches of modern science, and science would not be the same without them. Yet the question of how they can explain real-world processes remains an issue of considerable debate. In this context, a range of authors have highlighted the inferences back to the world that computer simulations allow us to draw. I will first characterize the precise relation between computer and target of a simulation that allows us to draw such inferences. I then argue that in a range of scientifically interesting cases they are particular abductions and defend this claim by appeal to two case studies.