We provide two programmatic frameworks for integrating philosophical research on understanding with complementary work in computer science, psychology, and neuroscience. First, philosophical theories of understanding have consequences about how agents should reason if they are to understand that can then be evaluated empirically by their concordance with findings in scientific studies of reasoning. Second, these studies use a multitude of explanations, and a philosophical theory of understanding is well suited to integrating these explanations in illuminating ways.

Inferentialists about scientific representation hold that an apparatus’s representing a target system consists in the apparatus allowing “surrogative inferences” about the target. I argue that a serious problem for inferentialism arises from the fact that many scientific theories and models contain internal inconsistencies. Inferentialism, left unamended, implies that inconsistent scientific models have unlimited representational power, since an inconsistency permits any conclusion to be inferred. I consider a number of ways that inferentialists can respond to this challenge before suggesting my own (...) solution. I develop an analogy to exploitable glitches in a game. Even though inconsistent representational apparatuses may in some sense allow for contradictions to be generated within them, doing so violates the intended function of the apparatus’s parts and hence violates representational “gameplay.”. (shrink)

If ontic dependence is the basis of explanation, there cannot be mathematical explanations. Accounting for the explanatory dependency between mathematical properties and empirical phenomena poses i...

We critically engage two traditional views of scientific data and outline a novel philosophical view that we call the pragmatic-representational view of data. On the PR view, data are representations that are the product of a process of inquiry, and they should be evaluated in terms of their adequacy or fitness for particular purposes. Some important implications of the PR view for data assessment, related to misrepresentation, context-sensitivity, and complementary use, are highlighted. The PR view provides insight into the common (...) but little-discussed practices of iteratively reusing and repurposing data, which result in many datasets’ having a phylogeny—an origin and complex evolutionary history—that is relevant to their evaluation and future use. We relate these insights to the open-data and data-rescue movements, and highlight several future avenues of research that build on the PR view of data. (shrink)

Ecological-enactive approaches to cognition aim to explain cognition in terms of the dynamic coupling between agent and environment. Accordingly, cognition of one’s immediate environment (which is sometimes labeled “basic” cognition) depends on enaction and the picking up of affordances. However, ecological-enactive views supposedly fail to account for what is sometimes called “higher” cognition, i.e., cognition about potentially absent targets, which therefore can only be explained by postulating representational content. This challenge levelled against ecological-enactive approaches highlights a putative explanatory gap between (...) basic and higher cognition. In this paper, we examine scientific cognition—a paradigmatic case of higher cognition—and argue that it shares fundamental features with basic cognition, for enaction and affordance selection are central to the scientific enterprise. Our argument focuses on modeling, and on how models promote scientific understanding. We base our argument on a non-representational account of scientific understanding and on the material engagement theory, for models are hereby conceived as material objects designed for scientific engagements. Having done so, we conclude that the explanatory gap is significantly less threatening to the ecological-enactive approach than it might appear. (shrink)

Machine learning methods have recently created high expectations in the climate modelling context in view of addressing climate change, but they are often considered as non-physics-based ‘black boxes’ that may not provide any understanding. However, in many ways, understanding seems indispensable to appropriately evaluate climate models and to build confidence in climate projections. Relying on two case studies, we compare how machine learning and standard statistical techniques affect our ability to understand the climate system. For that purpose, we put five (...) evaluative criteria of understanding to work: intelligibility, representational accuracy, empirical accuracy, coherence with background knowledge, and assessment of the domain of validity. We argue that the two families of methods are part of the same continuum where these various criteria of understanding come in degrees, and that therefore machine learning methods do not necessarily constitute a radical departure from standard statistical tools, as far as understanding is concerned. (shrink)

Simple idealized models seem to provide more understanding than opaque, complex, and hyper-realistic models. However, an increasing number of scientists are going in the opposite direction by utilizing opaque machine learning models to make predictions and draw inferences, suggesting that scientists are opting for models that have less potential for understanding. Are scientists trading understanding for some other epistemic or pragmatic good when they choose a machine learning model? Or are the assumptions behind why minimal models provide understanding misguided? In (...) this paper, using the case of deep neural networks, I argue that it is not the complexity or black box nature of a model that limits how much understanding the model provides. Instead, it is a lack of scientific and empirical evidence supporting the link that connects a model to the target phenomenon that primarily prohibits understanding. (shrink)

Despite widespread evidence that fictional models play an explanatory role in science, resistance remains to the idea that fictions can explain. A central source of this resistance is a particular view about what explanations are, namely, the ontic conception of explanation. According to the ontic conception, explanations just are the concrete entities in the world. I argue this conception is ultimately incoherent and that even a weaker version of the ontic conception fails. Fictional models can succeed in offering genuine explanations (...) by correctly capturing relevant patterns of counterfactual dependence and licensing correct inferences. Using the example of Newtonian force explanations of the tides, I show how, even in science, fiction can be a vehicle for truth. (shrink)

How can we use models to understand real phenomena if models misrepresent the very phenomena we seek to understand? Some accounts suggest that models may afford understanding by providing causal knowledge about phenomena via how-possibly explanations. However, general equilibrium models, for example, pose a challenge to this solution since their contribution appears to be purely mathematical results. Despite this, practitioners widely acknowledge that it improves our understanding of the world. I argue that the Arrow–Debreu model provides a mathematical how-possibly explanation (...) which establishes claims of mathematical dependence. The account developed reveals how mathematical knowledge can inform claims about the world, allow ‘what-if-things-had-been-different’ inferences, and thus improve our understanding. (shrink)

In recent years, several authors have debated about the justifiability of so-called scientific imperialism. To date, however, widespread disagreements remain regarding both the identification and the normative evaluation of scientific imperialism. In this paper, I aim to remedy this situation by making some conceptual distinctions concerning scientific imperialism and by providing a detailed assessment of the most prominent objections to it. I shall argue that these objections provide a valuable basis for opposing some instances of scientific imperialism, but do not (...) yield cogent reasons to think that scientific imperialism in general is objectionable or unjustified. I then highlight three wide-ranging implications of this result for the ongoing philosophical debate about the justifiability of scientific imperialism. (shrink)

Can purely predictive models be useful in investigating causal systems? I argue ‘yes’. Moreover, in many cases not only are they useful, they are essential. The alternative is to stick to models or mechanisms drawn from well-understood theory. But a necessary condition for explanation is empirical success, and in many cases in social and field sciences such success can only be achieved by purely predictive models, not by ones drawn from theory. Alas, the attempt to use theory to achieve explanation (...) or insight without empirical success therefore fails, leaving us with the worst of both worlds – neither prediction nor explanation. Best go with empirical success by any means necessary. I support these methodological claims via case studies of two impressive feats of predictive modelling: opinion polling of political elections, and weather forecasting. (shrink)

In the current debate on the relation between how-possibly explanations (HPEs) and understanding, two seemingly irreconcilable positions have emerged, which either deny or assert HPEs’ contribution to understanding. We argue, in contrast, that there is substantial room for reconciliation between these positions. First, we show that a shared assumption is unfounded: HPEs can be interpreted as being correct explanations. Second, we argue that what we call the standard account is actually compatible with the claim that HPEs may improve understanding. Our (...) analysis not only indicates that there is room for reconciliation, but also specifies the potential remaining disagreements. (shrink)

In the natural and social sciences, it is common to use toy models—extremely simple and highly idealized representations—to understand complex phenomena. Some of the simple surrogate models used to understand opaque machine learning (ML) models, such as rule lists and sparse decision trees, bear some resemblance to scientific toy models. They allow non-experts to understand how an opaque ML model works globally via a much simpler model that highlights the most relevant features of the input space and their effect on (...) the output. The obvious difference is that the common target of a toy and a full-scale model in the sciences is some phenomenon in the world, while the target of a surrogate model is another model. This essential difference makes toy surrogate models (TSMs) a new object of study for theories of understanding, one that is not easily accommodated under current analyses. This paper provides an account of what it means to understand an opaque ML model globally with the aid of such simple models. (shrink)

Ecological-enactive approaches to cognition aim to explain cognition in terms of the dynamic coupling between agent and environment. Accordingly, cognition of one’s immediate environment depends on enaction and the picking up of affordances. However, ecological-enactive views supposedly fail to account for what is sometimes called “higher” cognition, i.e., cognition about potentially absent targets, which therefore can only be explained by postulating representational content. This challenge levelled against ecological-enactive approaches highlights a putative explanatory gap between basic and higher cognition. In this (...) paper, we examine scientific cognition—a paradigmatic case of higher cognition—and argue that it shares fundamental features with basic cognition, for enaction and affordance selection are central to the scientific enterprise. Our argument focuses on modeling, and on how models promote scientific understanding. We base our argument on a non-representational account of scientific understanding and on the material engagement theory, for models are hereby conceived as material objects designed for scientific engagements. Having done so, we conclude that the explanatory gap is significantly less threatening to the ecological-enactive approach than it might appear. (shrink)

Highly idealized models may serve various epistemic functions, notably explanation, in virtue of representing the world. Inferentialism provides a prima facie compelling characterization of what constitutes the representation relation. In this paper, I argue that what I call factive inferentialism does not provide a satisfactory solution to the puzzle of model-based—factive—explanation. In particular, I show that making explanatory counterfactual inferences is not a sufficient guide for accurate representation, factivity, or realism. I conclude by calling for a more explicit specification of (...) model-world mismatches and properties imputation. (shrink)

One can find in the literature two sets of views concerning the relationship between understanding and explanation: that one understands only if 1) one has knowledge of causes and 2) that knowledge is provided by an explanation. Taken together, these tenets characterize what I call the narrow knowledge account of understanding. While the first tenet has recently come under severe attack, the second has been more resistant to change. I argue that we have good reasons to reject it on the (...) basis of theoretical models that provide how-possibly explanations. These models, while they do not explain in the strict sense, afford understanding. In response, I propose an alternative epistemology of understanding, broad KAU, that takes cases of theoretical modelling into account. (shrink)

Scientific understanding is a fundamental goal of science. However, there is currently no good way to measure the scientific understanding of agents, whether these be humans or Artificial Intelligence systems. Without a clear benchmark, it is challenging to evaluate and compare different levels of scientific understanding. In this paper, we propose a framework to create a benchmark for scientific understanding, utilizing tools from philosophy of science. We adopt a behavioral conception of understanding, according to which genuine understanding should be recognized (...) as an ability to perform certain tasks. We extend this notion of scientific understanding by considering a set of questions that gauge different levels of scientific understanding, covering information retrieval, the capability to arrange information to produce an explanation, and the ability to infer how things would be different under different circumstances. We suggest building a Scientific Understanding Benchmark (SUB), formed by a set of these tests, allowing for the evaluation and comparison of scientific understanding. Benchmarking plays a crucial role in establishing trust, ensuring quality control, and providing a basis for performance evaluation. By aligning machine and human scientific understanding we can improve their utility, ultimately advancing scientific understanding and helping to discover new insights within machines. (shrink)

In a recent paper, Erasmus et al. (2021) defend the idea that the ambiguity of the term “explanation” in explainable AI (XAI) can be solved by adopting any of four different extant accounts of explanation in the philosophy of science: the Deductive Nomological, Inductive Statistical, Causal Mechanical, and New Mechanist models. In this chapter, I show that the authors’ claim that these accounts can be applied to deep neural networks as they would to any natural phenomenon is mistaken. I also (...) provide a more general argument as to why the notion of explainability as it is currently used in the XAI literature bears little resemblance to the traditional concept of scientific explanation. It would be more fruitful to use the label “understandable AI” to avoid the confusion that surrounds the goal and purposes of XAI. In the second half of the chapter, I argue for a pragmatic conception of understanding that is better suited to play the central role attributed to explanation in XAI. Following De Regt (2017) and Kuorikoski and Ylikoski (2015), the conditions of satisfaction for understanding an ML system are fleshed out in terms of an agent’s success in using the system. (shrink)

Understanding is a valued trait in any epistemic practice, scientific or not. Yet, when it comes to characterizing its nature, the notion has not received the philosophical attention it deserves. We have set ourselves three tasks in this paper. First, we defend the importance of this endeavor. Second, we consider and criticize a number of proposals to this effect. Third, we defend an alternative account, focusing on abilities as the proper mark of understanding.

A popular view presents explanations in the cognitive sciences as causal or mechanistic and argues that an important feature of such explanations is that they allow us to manipulate and control the explanandum phenomena. Nonetheless, whether there can be explanations in the cognitive sciences that are neither causal nor mechanistic is still under debate. Another prominent view suggests that both causal and non-causal relations of counterfactual dependence can be explanatory, but this view is open to the criticism that it is (...) not clear how to distinguish explanatory from non-explanatory relations. In this paper, I draw from both views and suggest that, in the cognitive sciences, relations of counterfactual dependence that allow manipulation and control can be explanatory even when they are neither causal nor mechanistic. Furthermore, the ability to allow manipulation can determine whether non-causal counterfactual dependence relations are explanatory. I present a preliminary framework for manipulation relations that includes some non-causal relations and use two examples from the cognitive sciences to show how this framework distinguishes between explanatory and non-explanatory, non-causal relations. The proposed framework suggests that, in the cognitive sciences, causal and non-causal relations have the same criterion for explanatory value, namely, whether or not they allow manipulation and control. (shrink)

The social epistemology of science has adopted agent-based computer simulations as one of its core methods for investigating the dynamics of scientific inquiry. The epistemic status of these highly idealized models is currently under active debate in which they are often associated either with predictive or the argumentative functions. These two functions roughly correspond to interpreting simulations as virtual experiments or formalized thought experiments, respectively. This paper advances the argumentative account of modeling by proposing that models serve as a means (...) to (re)conceptualize the macro-level dynamics of complex social epistemic interactions. I apply results from the epistemology of scientific modeling and the psychology of mental simulation to the ongoing debate in the social epistemology of science. Instead of considering simulation models as predictive devices, I view them as artifacts that exemplify abstract hypothetical properties of complex social epistemic processes in order to advance scientific understanding, hypothesis formation, and communication. Models need not be accurate representations to serve these purposes. They should be regarded as pragmatic cognitive tools that engender rather than replace intuitions in philosophical reasoning and argumentation. Furthermore, I aim to explain why the community tends to converge around few model templates: Since models have the potential to transform our intuitive comprehension of the subject of inquiry, successful models may literally capture the imagination of the modeling community. (shrink)

Scientific understanding is a fundamental goal of science. However, there is currently no good way to measure the scientific understanding of agents, whether these be humans or Artificial Intelligence systems. Without a clear benchmark, it is challenging to evaluate and compare different levels of scientific understanding. In this paper, we propose a framework to create a benchmark for scientific understanding, utilizing tools from philosophy of science. We adopt a behavioral conception of understanding, according to which genuine understanding should be recognized (...) as an ability to perform certain tasks. We extend this notion of scientific understanding by considering a set of questions that gauge different levels of scientific understanding, covering information retrieval, the capability to arrange information to produce an explanation, and the ability to infer how things would be different under different circumstances. We suggest building a Scientific Understanding Benchmark (SUB), formed by a set of these tests, allowing for the evaluation and comparison of scientific understanding. Benchmarking plays a crucial role in establishing trust, ensuring quality control, and providing a basis for performance evaluation. By aligning machine and human scientific understanding we can improve their utility, ultimately advancing scientific understanding and helping to discover new insights within machines. (shrink)

This paper argues for a representational semantic conception of scientific theories, which respects the bare claim of any semantic view, namely that theories can be characterised as sets of models. RSC must be sharply distinguished from structural versions that assume a further identity of ‘models’ and ‘structures’, which we reject. The practice-turn in the recent philosophical literature suggests instead that modelling must be understood in a deflationary spirit, in terms of the diverse representational practices in the sciences. These insights are (...) applied to some mathematical models, thus showing that the mathematical sciences are not in principle counterexamples to RSC. (shrink)

Because idealizations frequently advance scientific understanding, many claim that falsehoods play an epistemic role. In this paper, we argue that these positions greatly overstate idealiza...

This article compares the epistemic roles of theoretical models and model organisms in science, and specifically the role of non-human animal models in biomedicine. Much of the previous literature on this topic shares an assumption that animal models and theoretical models have a broadly similar epistemic role—that of indirect representation of a target through the study of a surrogate system. Recently, Levy and Currie have argued that model organism research and theoretical modelling differ in the justification of model-to-target inferences, such (...) that a unified account based on the widely accepted idea of modelling as indirect representation does not similarly apply to both. I defend a similar conclusion, but argue that the distinction between animal models and theoretical models does not always track a difference in the justification of model-to-target inferences. Case studies of the use of animal models in biomedicine are presented to illustrate this. However, Levy and Currie’s point can be argued for in a different way. I argue for the following distinction. Model organisms function as surrogate sources of evidence, from which results are transferred to their targets by empirical extrapolation. By contrast, theoretical modelling does not involve such an inductive step. Rather, theoretical models are used for drawing conclusions from what is already known or assumed about the target system. Codifying assumptions about the causal structure of the target in external representational media allows one to apply explicit inferential rules to reach conclusions that could not be reached with unaided cognition alone. (shrink)

Philosophers have recently paid more attention to the physical aspects of scientific models. The attention is motivated by the prospect that a model’s physical features strongly affect its use and that this suggests re-thinking modelling in terms of extended or distributed cognition. This paper investigates two ways in which physical features of scientific models affect their use and it asks whether modelling is an instance of extended cognition. I approach these topics with a historical case study, in which scientists kept (...) records not only of their findings, but also of some the mental operations that generated the findings. The case study shows how scientists can employ a physical model as an external information store, which allows alternating between mental manipulations, recording the outcome externally, and then feeding the outcome back into subsequent mental manipulations. The case study also demonstrates that a models’ physical nature allows replacing explicit reasoning with visuospatial manipulations. I argue, furthermore, that physical modelling does not need to exemplify a strong kind of extended cognition, the sort for which external features are mereological parts of cognition. It can exemplify a weaker kind, instead. (shrink)